[tutorial] edited hands-on practices (#1899)

* Add handson to ColossalAI.

* Change names of handsons and edit sequence parallel example.

* Edit wrong folder name

* resolve conflict

* delete readme

examples/tutorial/hybrid_parallel/README.md

@@ -0,0 +1,27 @@
+# Handson 1: Multi-dimensional Parallelism with Colossal-AI
+
+
+## Install Colossal-AI and other dependencies
+
+```bash
+sh install.sh
+```
+
+
+## Prepare Dataset
+
+We use CIFAR10 dataset in this example. The dataset will be downloaded to `../data` by default. 
+If you wish to use customized directory for the dataset. You can set the environment variable `DATA` via the following command.
+
+```bash
+export DATA=/path/to/data
+```
+
+
+## Run on 2*2 device mesh
+
+Current configuration setting on `config.py` is TP=2, PP=2.
+
+```bash
+colossalai run --nproc_per_node 4 train.py --config config.py
+```

examples/tutorial/hybrid_parallel/config.py

@@ -0,0 +1,36 @@
+from colossalai.amp import AMP_TYPE
+
+# hyperparameters
+# BATCH_SIZE is as per GPU
+# global batch size = BATCH_SIZE x data parallel size
+BATCH_SIZE = 256
+LEARNING_RATE = 3e-3
+WEIGHT_DECAY = 0.3
+NUM_EPOCHS = 10
+WARMUP_EPOCHS = 3
+
+# model config
+IMG_SIZE = 224
+PATCH_SIZE = 16
+HIDDEN_SIZE = 512
+DEPTH = 4
+NUM_HEADS = 4
+MLP_RATIO = 2
+NUM_CLASSES = 1000
+CHECKPOINT = False
+SEQ_LENGTH = (IMG_SIZE // PATCH_SIZE)**2 + 1    # add 1 for cls token
+
+# parallel setting
+TENSOR_PARALLEL_SIZE = 2
+TENSOR_PARALLEL_MODE = '1d'
+
+parallel = dict(
+    pipeline=2,
+    tensor=dict(mode=TENSOR_PARALLEL_MODE, size=TENSOR_PARALLEL_SIZE),
+)
+
+fp16 = dict(mode=AMP_TYPE.NAIVE)
+clip_grad_norm = 1.0
+
+# pipeline config
+NUM_MICRO_BATCHES = parallel['pipeline']

examples/tutorial/hybrid_parallel/install.sh

@@ -0,0 +1,4 @@
+pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 torchaudio==0.11.0 --extra-index-url https://download.pytorch.org/whl/cu113
+pip install colossalai==0.1.10+torch1.12cu11.3 -f https://release.colossalai.org
+pip install titans
+colossalai check -i

examples/tutorial/hybrid_parallel/train.py

@@ -0,0 +1,116 @@
+import os
+import colossalai
+import torch
+
+from tqdm import tqdm
+from colossalai.context import ParallelMode
+from colossalai.core import global_context as gpc
+from colossalai.logging import get_dist_logger
+from colossalai.nn import CrossEntropyLoss
+from colossalai.nn.lr_scheduler import CosineAnnealingWarmupLR
+from colossalai.utils import is_using_pp, get_dataloader
+from colossalai.pipeline.pipelinable import PipelinableContext
+from titans.model.vit.vit import _create_vit_model
+from titans.dataloader.cifar10 import build_cifar
+
+
+def main():
+    # initialize distributed setting
+    parser = colossalai.get_default_parser()
+    args = parser.parse_args()
+
+    # launch from torch
+    colossalai.launch_from_torch(config=args.config)
+
+    # get logger
+    logger = get_dist_logger()
+    logger.info("initialized distributed environment", ranks=[0])
+
+    if hasattr(gpc.config, 'LOG_PATH'):
+        if gpc.get_global_rank() == 0:
+            log_path = gpc.config.LOG_PATH
+            if not os.path.exists(log_path):
+                os.mkdir(log_path)
+            logger.log_to_file(log_path)
+
+    use_pipeline = is_using_pp()
+
+    # create model
+    model_kwargs = dict(img_size=gpc.config.IMG_SIZE,
+                        patch_size=gpc.config.PATCH_SIZE,
+                        hidden_size=gpc.config.HIDDEN_SIZE,
+                        depth=gpc.config.DEPTH,
+                        num_heads=gpc.config.NUM_HEADS,
+                        mlp_ratio=gpc.config.MLP_RATIO,
+                        num_classes=10,
+                        init_method='jax',
+                        checkpoint=gpc.config.CHECKPOINT)
+
+    if use_pipeline:
+        pipelinable = PipelinableContext()
+        with pipelinable:
+            model = _create_vit_model(**model_kwargs)
+        pipelinable.to_layer_list()
+        pipelinable.policy = "uniform"
+        model = pipelinable.partition(
+            1, gpc.pipeline_parallel_size, gpc.get_local_rank(ParallelMode.PIPELINE))
+    else:
+        model = _create_vit_model(**model_kwargs)
+
+    # count number of parameters
+    total_numel = 0
+    for p in model.parameters():
+        total_numel += p.numel()
+    if not gpc.is_initialized(ParallelMode.PIPELINE):
+        pipeline_stage = 0
+    else:
+        pipeline_stage = gpc.get_local_rank(ParallelMode.PIPELINE)
+    logger.info(
+        f"number of parameters: {total_numel} on pipeline stage {pipeline_stage}")
+
+    # create dataloaders
+    root = os.environ.get('DATA', '../data/cifar10')
+    train_dataloader, test_dataloader = build_cifar(
+        gpc.config.BATCH_SIZE, root, pad_if_needed=True)
+
+    # create loss function
+    criterion = CrossEntropyLoss(label_smoothing=0.1)
+
+    # create optimizer
+    optimizer = torch.optim.AdamW(model.parameters(
+    ), lr=gpc.config.LEARNING_RATE, weight_decay=gpc.config.WEIGHT_DECAY)
+
+    # create lr scheduler
+    lr_scheduler = CosineAnnealingWarmupLR(optimizer=optimizer,
+                                           total_steps=gpc.config.NUM_EPOCHS,
+                                           warmup_steps=gpc.config.WARMUP_EPOCHS)
+
+    # initialize
+    engine, train_dataloader, test_dataloader, _ = colossalai.initialize(model=model,
+                                                                         optimizer=optimizer,
+                                                                         criterion=criterion,
+                                                                         train_dataloader=train_dataloader,
+                                                                         test_dataloader=test_dataloader)
+
+    logger.info("Engine is built", ranks=[0])
+
+    data_iter = iter(train_dataloader)
+
+    for epoch in range(gpc.config.NUM_EPOCHS):
+        # training
+        engine.train()
+
+        if gpc.get_global_rank() == 0:
+            description = 'Epoch {} / {}'.format(epoch, gpc.config.NUM_EPOCHS)
+            progress = tqdm(range(len(train_dataloader)), desc=description)
+        else:
+            progress = range(len(train_dataloader))
+        for _ in progress:
+            engine.zero_grad()
+            engine.execute_schedule(data_iter, return_output_label=False)
+            engine.step()
+            lr_scheduler.step()
+
+
+if __name__ == '__main__':
+    main()

examples/tutorial/large_batch_optimizer/README.md

@@ -0,0 +1,17 @@
+# Handson 4: Comparison of Large Batch Training Optimization
+
+## Prepare Dataset
+
+We use CIFAR10 dataset in this example. The dataset will be downloaded to `../data` by default. 
+If you wish to use customized directory for the dataset. You can set the environment variable `DATA` via the following command.
+
+```bash
+export DATA=/path/to/data
+```
+
+
+## Run on 2*2 device mesh
+
+```bash
+colossalai run --nproc_per_node 4 train.py --config config.py
+```

examples/tutorial/large_batch_optimizer/config.py

@@ -0,0 +1,36 @@
+from colossalai.amp import AMP_TYPE
+
+# hyperparameters
+# BATCH_SIZE is as per GPU
+# global batch size = BATCH_SIZE x data parallel size
+BATCH_SIZE = 512
+LEARNING_RATE = 3e-3
+WEIGHT_DECAY = 0.3
+NUM_EPOCHS = 10
+WARMUP_EPOCHS = 3
+
+# model config
+IMG_SIZE = 224
+PATCH_SIZE = 16
+HIDDEN_SIZE = 512
+DEPTH = 4
+NUM_HEADS = 4
+MLP_RATIO = 2
+NUM_CLASSES = 1000
+CHECKPOINT = False
+SEQ_LENGTH = (IMG_SIZE // PATCH_SIZE)**2 + 1    # add 1 for cls token
+
+# parallel setting
+TENSOR_PARALLEL_SIZE = 2
+TENSOR_PARALLEL_MODE = '1d'
+
+parallel = dict(
+    pipeline=2,
+    tensor=dict(mode=TENSOR_PARALLEL_MODE, size=TENSOR_PARALLEL_SIZE),
+)
+
+fp16 = dict(mode=AMP_TYPE.NAIVE)
+clip_grad_norm = 1.0
+
+# pipeline config
+NUM_MICRO_BATCHES = parallel['pipeline']

examples/tutorial/large_batch_optimizer/train.py

@@ -0,0 +1,117 @@
+import os
+import colossalai
+import torch
+
+from tqdm import tqdm
+from colossalai.context import ParallelMode
+from colossalai.core import global_context as gpc
+from colossalai.logging import get_dist_logger
+from colossalai.nn import CrossEntropyLoss
+from colossalai.nn.lr_scheduler import CosineAnnealingWarmupLR
+from colossalai.nn.optimizer import Lars, Lamb
+from colossalai.utils import is_using_pp, get_dataloader
+from colossalai.pipeline.pipelinable import PipelinableContext
+from titans.model.vit.vit import _create_vit_model
+from titans.dataloader.cifar10 import build_cifar
+
+
+def main():
+    # initialize distributed setting
+    parser = colossalai.get_default_parser()
+    args = parser.parse_args()
+
+    # launch from torch
+    colossalai.launch_from_torch(config=args.config)
+
+    # get logger
+    logger = get_dist_logger()
+    logger.info("initialized distributed environment", ranks=[0])
+
+    if hasattr(gpc.config, 'LOG_PATH'):
+        if gpc.get_global_rank() == 0:
+            log_path = gpc.config.LOG_PATH
+            if not os.path.exists(log_path):
+                os.mkdir(log_path)
+            logger.log_to_file(log_path)
+
+    use_pipeline = is_using_pp()
+
+    # create model
+    model_kwargs = dict(img_size=gpc.config.IMG_SIZE,
+                        patch_size=gpc.config.PATCH_SIZE,
+                        hidden_size=gpc.config.HIDDEN_SIZE,
+                        depth=gpc.config.DEPTH,
+                        num_heads=gpc.config.NUM_HEADS,
+                        mlp_ratio=gpc.config.MLP_RATIO,
+                        num_classes=10,
+                        init_method='jax',
+                        checkpoint=gpc.config.CHECKPOINT)
+
+    if use_pipeline:
+        pipelinable = PipelinableContext()
+        with pipelinable:
+            model = _create_vit_model(**model_kwargs)
+        pipelinable.to_layer_list()
+        pipelinable.policy = "uniform"
+        model = pipelinable.partition(
+            1, gpc.pipeline_parallel_size, gpc.get_local_rank(ParallelMode.PIPELINE))
+    else:
+        model = _create_vit_model(**model_kwargs)
+
+    # count number of parameters
+    total_numel = 0
+    for p in model.parameters():
+        total_numel += p.numel()
+    if not gpc.is_initialized(ParallelMode.PIPELINE):
+        pipeline_stage = 0
+    else:
+        pipeline_stage = gpc.get_local_rank(ParallelMode.PIPELINE)
+    logger.info(
+        f"number of parameters: {total_numel} on pipeline stage {pipeline_stage}")
+
+    # create dataloaders
+    root = os.environ.get('DATA', '../data/cifar10')
+    train_dataloader, test_dataloader = build_cifar(
+        gpc.config.BATCH_SIZE, root, pad_if_needed=True)
+
+    # create loss function
+    criterion = CrossEntropyLoss(label_smoothing=0.1)
+
+    # create optimizer
+    optimizer = Lars(model.parameters(), lr=gpc.config.LEARNING_RATE,
+                     weight_decay=gpc.config.WEIGHT_DECAY)
+
+    # create lr scheduler
+    lr_scheduler = CosineAnnealingWarmupLR(optimizer=optimizer,
+                                           total_steps=gpc.config.NUM_EPOCHS,
+                                           warmup_steps=gpc.config.WARMUP_EPOCHS)
+
+    # initialize
+    engine, train_dataloader, test_dataloader, _ = colossalai.initialize(model=model,
+                                                                         optimizer=optimizer,
+                                                                         criterion=criterion,
+                                                                         train_dataloader=train_dataloader,
+                                                                         test_dataloader=test_dataloader)
+
+    logger.info("Engine is built", ranks=[0])
+
+    data_iter = iter(train_dataloader)
+
+    for epoch in range(gpc.config.NUM_EPOCHS):
+        # training
+        engine.train()
+
+        if gpc.get_global_rank() == 0:
+            description = 'Epoch {} / {}'.format(epoch, gpc.config.NUM_EPOCHS)
+            progress = tqdm(range(len(train_dataloader)), desc=description)
+        else:
+            progress = range(len(train_dataloader))
+        for _ in progress:
+            engine.zero_grad()
+            engine.execute_schedule(data_iter, return_output_label=False)
+            engine.step()
+            lr_scheduler.step()
+
+
+if __name__ == '__main__':
+    main()

examples/tutorial/opt/README.md

@@ -0,0 +1 @@
+# Handson 5: Fine-tuning and Serving for OPT from Hugging Face

examples/tutorial/opt/inference/README.md

@@ -0,0 +1,77 @@
+# Overview
+
+This is an example showing how to run OPT generation. The OPT model is implemented using ColossalAI.
+
+It supports tensor parallelism, batching and caching.
+
+# How to run
+
+Run OPT-125M:
+```shell
+python opt_fastapi.py opt-125m
+```
+
+It will launch a HTTP server on `0.0.0.0:7070` by default and you can customize host and port. You can open `localhost:7070/docs` in your browser to see the openapi docs.
+
+## Configure
+
+### Configure model
+```shell
+python opt_fastapi.py <model>
+```
+Available models: opt-125m, opt-6.7b, opt-30b, opt-175b.
+
+### Configure tensor parallelism
+```shell
+python opt_fastapi.py <model> --tp <TensorParallelismWorldSize>
+```
+The `<TensorParallelismWorldSize>` can be an integer in `[1, #GPUs]`. Default `1`.
+
+### Configure checkpoint
+```shell
+python opt_fastapi.py <model> --checkpoint <CheckpointPath>
+```
+The `<CheckpointPath>` can be a file path or a directory path. If it's a directory path, all files under the directory will be loaded.
+
+### Configure queue
+```shell
+python opt_fastapi.py <model> --queue_size <QueueSize>
+```
+The `<QueueSize>` can be an integer in `[0, MAXINT]`. If it's `0`, the request queue size is infinite. If it's a positive integer, when the request queue is full, incoming requests will be dropped (the HTTP status code of response will be 406).
+
+### Configure bathcing
+```shell
+python opt_fastapi.py <model> --max_batch_size <MaxBatchSize>
+```
+The `<MaxBatchSize>` can be an integer in `[1, MAXINT]`. The engine will make batch whose size is less or equal to this value.
+
+Note that the batch size is not always equal to `<MaxBatchSize>`, as some consecutive requests may not be batched.
+
+### Configure caching
+```shell
+python opt_fastapi.py <model> --cache_size <CacheSize> --cache_list_size <CacheListSize>
+```
+This will cache `<CacheSize>` unique requests. And for each unique request, it cache `<CacheListSize>` different results. A random result will be returned if the cache is hit.
+
+The `<CacheSize>` can be an integer in `[0, MAXINT]`. If it's `0`, cache won't be applied. The `<CacheListSize>` can be an integer in `[1, MAXINT]`.
+
+### Other configurations
+```shell
+python opt_fastapi.py -h
+```
+
+# How to benchmark
+```shell
+cd benchmark
+locust
+```
+
+Then open the web interface link which is on your console.
+
+# Pre-process pre-trained weights
+
+## OPT-66B
+See [script/processing_ckpt_66b.py](./script/processing_ckpt_66b.py).
+
+## OPT-175B
+See [script/process-opt-175b](./script/process-opt-175b/).

examples/tutorial/opt/inference/batch.py

@@ -0,0 +1,59 @@
+import torch
+from typing import List, Deque, Tuple, Hashable, Any
+from energonai import BatchManager, SubmitEntry, TaskEntry
+
+
+class BatchManagerForGeneration(BatchManager):
+    def __init__(self, max_batch_size: int = 1, pad_token_id: int = 0) -> None:
+        super().__init__()
+        self.max_batch_size = max_batch_size
+        self.pad_token_id = pad_token_id
+
+    def _left_padding(self, batch_inputs):
+        max_len = max(len(inputs['input_ids']) for inputs in batch_inputs)
+        outputs = {'input_ids': [], 'attention_mask': []}
+        for inputs in batch_inputs:
+            input_ids, attention_mask = inputs['input_ids'], inputs['attention_mask']
+            padding_len = max_len - len(input_ids)
+            input_ids = [self.pad_token_id] * padding_len + input_ids
+            attention_mask = [0] * padding_len + attention_mask
+            outputs['input_ids'].append(input_ids)
+            outputs['attention_mask'].append(attention_mask)
+        for k in outputs:
+            outputs[k] = torch.tensor(outputs[k])
+        return outputs, max_len
+
+    @staticmethod
+    def _make_batch_key(entry: SubmitEntry) -> tuple:
+        data = entry.data
+        return (data['top_k'], data['top_p'], data['temperature'])
+
+    def make_batch(self, q: Deque[SubmitEntry]) -> Tuple[TaskEntry, dict]:
+        entry = q.popleft()
+        uids = [entry.uid]
+        batch = [entry.data]
+        while len(batch) < self.max_batch_size:
+            if len(q) == 0:
+                break
+            if self._make_batch_key(entry) != self._make_batch_key(q[0]):
+                break
+            if q[0].data['max_tokens'] > entry.data['max_tokens']:
+                break
+            e = q.popleft()
+            batch.append(e.data)
+            uids.append(e.uid)
+        inputs, max_len = self._left_padding(batch)
+        trunc_lens = []
+        for data in batch:
+            trunc_lens.append(max_len + data['max_tokens'])
+        inputs['top_k'] = entry.data['top_k']
+        inputs['top_p'] = entry.data['top_p']
+        inputs['temperature'] = entry.data['temperature']
+        inputs['max_tokens'] = max_len + entry.data['max_tokens']
+        return TaskEntry(tuple(uids), inputs), {'trunc_lens': trunc_lens}
+
+    def split_batch(self, task_entry: TaskEntry, trunc_lens: List[int] = []) -> List[Tuple[Hashable, Any]]:
+        retval = []
+        for uid, output, trunc_len in zip(task_entry.uids, task_entry.batch, trunc_lens):
+            retval.append((uid, output[:trunc_len]))
+        return retval

examples/tutorial/opt/inference/benchmark/locustfile.py

@@ -0,0 +1,15 @@
+from locust import HttpUser, task
+from json import JSONDecodeError
+
+
+class GenerationUser(HttpUser):
+    @task
+    def generate(self):
+        prompt = 'Question: What is the longest river on the earth? Answer:'
+        for i in range(4, 9):
+            data = {'max_tokens': 2**i, 'prompt': prompt}
+            with self.client.post('/generation', json=data, catch_response=True) as response:
+                if response.status_code in (200, 406):
+                    response.success()
+                else:
+                    response.failure('Response wrong')

examples/tutorial/opt/inference/cache.py

@@ -0,0 +1,64 @@
+from collections import OrderedDict
+from threading import Lock
+from contextlib import contextmanager
+from typing import List, Any, Hashable, Dict
+
+
+class MissCacheError(Exception):
+    pass
+
+
+class ListCache:
+    def __init__(self, cache_size: int, list_size: int, fixed_keys: List[Hashable] = []) -> None:
+        """Cache a list of values. The fixed keys won't be removed. For other keys, LRU is applied.
+        When the value list is not full, a cache miss occurs. Otherwise, a cache hit occurs. Redundant values will be removed.
+
+        Args:
+            cache_size (int): Max size for LRU cache.
+            list_size (int): Value list size.
+            fixed_keys (List[Hashable], optional): The keys which won't be removed. Defaults to [].
+        """
+        self.cache_size = cache_size
+        self.list_size = list_size
+        self.cache: OrderedDict[Hashable, List[Any]] = OrderedDict()
+        self.fixed_cache: Dict[Hashable, List[Any]] = {}
+        for key in fixed_keys:
+            self.fixed_cache[key] = []
+        self._lock = Lock()
+
+    def get(self, key: Hashable) -> List[Any]:
+        with self.lock():
+            if key in self.fixed_cache:
+                l = self.fixed_cache[key]
+                if len(l) >= self.list_size:
+                    return l
+            elif key in self.cache:
+                self.cache.move_to_end(key)
+                l = self.cache[key]
+                if len(l) >= self.list_size:
+                    return l
+        raise MissCacheError()
+
+    def add(self, key: Hashable, value: Any) -> None:
+        with self.lock():
+            if key in self.fixed_cache:
+                l = self.fixed_cache[key]
+                if len(l) < self.list_size and value not in l:
+                    l.append(value)
+            elif key in self.cache:
+                self.cache.move_to_end(key)
+                l = self.cache[key]
+                if len(l) < self.list_size and value not in l:
+                    l.append(value)
+            else:
+                if len(self.cache) >= self.cache_size:
+                    self.cache.popitem(last=False)
+                self.cache[key] = [value]
+
+    @contextmanager
+    def lock(self):
+        try:
+            self._lock.acquire()
+            yield
+        finally:
+            self._lock.release()

examples/tutorial/opt/inference/opt_fastapi.py

@@ -0,0 +1,123 @@
+import argparse
+import logging
+import random
+from typing import Optional
+
+import uvicorn
+from energonai import QueueFullError, launch_engine
+from energonai.model import opt_6B, opt_30B, opt_125M, opt_175B
+from fastapi import FastAPI, HTTPException, Request
+from pydantic import BaseModel, Field
+from transformers import GPT2Tokenizer
+
+from batch import BatchManagerForGeneration
+from cache import ListCache, MissCacheError
+
+
+class GenerationTaskReq(BaseModel):
+    max_tokens: int = Field(gt=0, le=256, example=64)
+    prompt: str = Field(
+        min_length=1, example='Question: Where were the 2004 Olympics held?\nAnswer: Athens, Greece\n\nQuestion: What is the longest river on the earth?\nAnswer:')
+    top_k: Optional[int] = Field(default=None, gt=0, example=50)
+    top_p: Optional[float] = Field(default=None, gt=0.0, lt=1.0, example=0.5)
+    temperature: Optional[float] = Field(default=None, gt=0.0, lt=1.0, example=0.7)
+
+
+app = FastAPI()
+
+
+@app.post('/generation')
+async def generate(data: GenerationTaskReq, request: Request):
+    logger.info(f'{request.client.host}:{request.client.port} - "{request.method} {request.url.path}" - {data}')
+    key = (data.prompt, data.max_tokens)
+    try:
+        if cache is None:
+            raise MissCacheError()
+        outputs = cache.get(key)
+        output = random.choice(outputs)
+        logger.info('Cache hit')
+    except MissCacheError:
+        inputs = tokenizer(data.prompt, truncation=True, max_length=512)
+        inputs['max_tokens'] = data.max_tokens
+        inputs['top_k'] = data.top_k
+        inputs['top_p'] = data.top_p
+        inputs['temperature'] = data.temperature
+        try:
+            uid = id(data)
+            engine.submit(uid, inputs)
+            output = await engine.wait(uid)
+            output = tokenizer.decode(output, skip_special_tokens=True)
+            if cache is not None:
+                cache.add(key, output)
+        except QueueFullError as e:
+            raise HTTPException(status_code=406, detail=e.args[0])
+
+    return {'text': output}
+
+
+@app.on_event("shutdown")
+async def shutdown(*_):
+    engine.shutdown()
+    server.should_exit = True
+    server.force_exit = True
+    await server.shutdown()
+
+
+def get_model_fn(model_name: str):
+    model_map = {
+        'opt-125m': opt_125M,
+        'opt-6.7b': opt_6B,
+        'opt-30b': opt_30B,
+        'opt-175b': opt_175B
+    }
+    return model_map[model_name]
+
+
+def print_args(args: argparse.Namespace):
+    print('\n==> Args:')
+    for k, v in args.__dict__.items():
+        print(f'{k} = {v}')
+
+
+FIXED_CACHE_KEYS = [
+    ('Question: What is the name of the largest continent on earth?\nAnswer: Asia\n\nQuestion: What is at the center of the solar system?\nAnswer:', 64),
+    ('A chat between a salesman and a student.\n\nSalesman: Hi boy, are you looking for a new phone?\nStudent: Yes, my phone is not functioning well.\nSalesman: What is your budget? \nStudent: I have received my scholarship so I am fine with any phone.\nSalesman: Great, then perhaps this latest flagship phone is just right for you.', 64),
+    ("English: I am happy today.\nChinese: 我今天很开心。\n\nEnglish: I am going to play basketball.\nChinese: 我一会去打篮球。\n\nEnglish: Let's celebrate our anniversary.\nChinese:", 64)
+]
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('model', choices=['opt-125m', 'opt-6.7b', 'opt-30b', 'opt-175b'])
+    parser.add_argument('--tp', type=int, default=1)
+    parser.add_argument('--master_host', default='localhost')
+    parser.add_argument('--master_port', type=int, default=19990)
+    parser.add_argument('--rpc_port', type=int, default=19980)
+    parser.add_argument('--max_batch_size', type=int, default=8)
+    parser.add_argument('--pipe_size', type=int, default=1)
+    parser.add_argument('--queue_size', type=int, default=0)
+    parser.add_argument('--http_host', default='0.0.0.0')
+    parser.add_argument('--http_port', type=int, default=7070)
+    parser.add_argument('--checkpoint', default=None)
+    parser.add_argument('--cache_size', type=int, default=0)
+    parser.add_argument('--cache_list_size', type=int, default=1)
+    args = parser.parse_args()
+    print_args(args)
+    model_kwargs = {}
+    if args.checkpoint is not None:
+        model_kwargs['checkpoint'] = args.checkpoint
+
+    logger = logging.getLogger(__name__)
+    tokenizer = GPT2Tokenizer.from_pretrained('facebook/opt-30b')
+    if args.cache_size > 0:
+        cache = ListCache(args.cache_size, args.cache_list_size, fixed_keys=FIXED_CACHE_KEYS)
+    else:
+        cache = None
+    engine = launch_engine(args.tp, 1, args.master_host, args.master_port, args.rpc_port, get_model_fn(args.model),
+                           batch_manager=BatchManagerForGeneration(max_batch_size=args.max_batch_size,
+                                                                   pad_token_id=tokenizer.pad_token_id),
+                           pipe_size=args.pipe_size,
+                           queue_size=args.queue_size,
+                           **model_kwargs)
+    config = uvicorn.Config(app, host=args.http_host, port=args.http_port)
+    server = uvicorn.Server(config=config)
+    server.run()

examples/tutorial/opt/inference/opt_server.py

@@ -0,0 +1,122 @@
+import logging
+import argparse
+import random
+from torch import Tensor
+from pydantic import BaseModel, Field
+from typing import Optional
+from energonai.model import opt_125M, opt_30B, opt_175B, opt_6B
+from transformers import GPT2Tokenizer
+from energonai import launch_engine, QueueFullError
+from sanic import Sanic
+from sanic.request import Request
+from sanic.response import json
+from sanic_ext import validate, openapi
+from batch import BatchManagerForGeneration
+from cache import ListCache, MissCacheError
+
+
+class GenerationTaskReq(BaseModel):
+    max_tokens: int = Field(gt=0, le=256, example=64)
+    prompt: str = Field(
+        min_length=1, example='Question: Where were the 2004 Olympics held?\nAnswer: Athens, Greece\n\nQuestion: What is the longest river on the earth?\nAnswer:')
+    top_k: Optional[int] = Field(default=None, gt=0, example=50)
+    top_p: Optional[float] = Field(default=None, gt=0.0, lt=1.0, example=0.5)
+    temperature: Optional[float] = Field(default=None, gt=0.0, lt=1.0, example=0.7)
+
+
+app = Sanic('opt')
+
+
+@app.post('/generation')
+@openapi.body(GenerationTaskReq)
+@validate(json=GenerationTaskReq)
+async def generate(request: Request, body: GenerationTaskReq):
+    logger.info(f'{request.ip}:{request.port} - "{request.method} {request.path}" - {body}')
+    key = (body.prompt, body.max_tokens)
+    try:
+        if cache is None:
+            raise MissCacheError()
+        outputs = cache.get(key)
+        output = random.choice(outputs)
+        logger.info('Cache hit')
+    except MissCacheError:
+        inputs = tokenizer(body.prompt, truncation=True, max_length=512)
+        inputs['max_tokens'] = body.max_tokens
+        inputs['top_k'] = body.top_k
+        inputs['top_p'] = body.top_p
+        inputs['temperature'] = body.temperature
+        try:
+            uid = id(body)
+            engine.submit(uid, inputs)
+            output = await engine.wait(uid)
+            assert isinstance(output, Tensor)
+            output = tokenizer.decode(output, skip_special_tokens=True)
+            if cache is not None:
+                cache.add(key, output)
+        except QueueFullError as e:
+            return json({'detail': e.args[0]}, status=406)
+
+    return json({'text': output})
+
+
+@app.after_server_stop
+def shutdown(*_):
+    engine.shutdown()
+
+
+def get_model_fn(model_name: str):
+    model_map = {
+        'opt-125m': opt_125M,
+        'opt-6.7b': opt_6B,
+        'opt-30b': opt_30B,
+        'opt-175b': opt_175B
+    }
+    return model_map[model_name]
+
+
+def print_args(args: argparse.Namespace):
+    print('\n==> Args:')
+    for k, v in args.__dict__.items():
+        print(f'{k} = {v}')
+
+
+FIXED_CACHE_KEYS = [
+    ('Question: What is the name of the largest continent on earth?\nAnswer: Asia\n\nQuestion: What is at the center of the solar system?\nAnswer:', 64),
+    ('A chat between a salesman and a student.\n\nSalesman: Hi boy, are you looking for a new phone?\nStudent: Yes, my phone is not functioning well.\nSalesman: What is your budget? \nStudent: I have received my scholarship so I am fine with any phone.\nSalesman: Great, then perhaps this latest flagship phone is just right for you.', 64),
+    ("English: I am happy today.\nChinese: 我今天很开心。\n\nEnglish: I am going to play basketball.\nChinese: 我一会去打篮球。\n\nEnglish: Let's celebrate our anniversary.\nChinese:", 64)
+]
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('model', choices=['opt-125m', 'opt-6.7b', 'opt-30b', 'opt-175b'])
+    parser.add_argument('--tp', type=int, default=1)
+    parser.add_argument('--master_host', default='localhost')
+    parser.add_argument('--master_port', type=int, default=19990)
+    parser.add_argument('--rpc_port', type=int, default=19980)
+    parser.add_argument('--max_batch_size', type=int, default=8)
+    parser.add_argument('--pipe_size', type=int, default=1)
+    parser.add_argument('--queue_size', type=int, default=0)
+    parser.add_argument('--http_host', default='0.0.0.0')
+    parser.add_argument('--http_port', type=int, default=7070)
+    parser.add_argument('--checkpoint', default=None)
+    parser.add_argument('--cache_size', type=int, default=0)
+    parser.add_argument('--cache_list_size', type=int, default=1)
+    args = parser.parse_args()
+    print_args(args)
+    model_kwargs = {}
+    if args.checkpoint is not None:
+        model_kwargs['checkpoint'] = args.checkpoint
+
+    logger = logging.getLogger(__name__)
+    tokenizer = GPT2Tokenizer.from_pretrained('facebook/opt-30b')
+    if args.cache_size > 0:
+        cache = ListCache(args.cache_size, args.cache_list_size, fixed_keys=FIXED_CACHE_KEYS)
+    else:
+        cache = None
+    engine = launch_engine(args.tp, 1, args.master_host, args.master_port, args.rpc_port, get_model_fn(args.model),
+                           batch_manager=BatchManagerForGeneration(max_batch_size=args.max_batch_size,
+                                                                   pad_token_id=tokenizer.pad_token_id),
+                           pipe_size=args.pipe_size,
+                           queue_size=args.queue_size,
+                           **model_kwargs)
+    app.run(args.http_host, args.http_port)

examples/tutorial/opt/inference/requirements.txt

@@ -0,0 +1,8 @@
+fastapi==0.85.1
+locust==2.11.0
+pydantic==1.10.2
+sanic==22.9.0
+sanic_ext==22.9.0
+torch>=1.10.0
+transformers==4.23.1
+uvicorn==0.19.0

examples/tutorial/opt/inference/script/process-opt-175b/README.md

@@ -0,0 +1,46 @@
+# Process OPT-175B weights
+
+You should download the pre-trained weights following the [doc](https://github.com/facebookresearch/metaseq/tree/main/projects/OPT) before reading this.
+
+First, install `metaseq` and `git clone https://github.com/facebookresearch/metaseq.git`.
+
+Then, `cd metaseq`.
+
+To consolidate checkpoints to eliminate FSDP:
+
+```shell
+bash metaseq/scripts/reshard_mp_launch_no_slurm.sh <directory_where_all_the_shards_are>/checkpoint_last <output_dir>/ 8 1
+```
+
+You will get 8 files in `<output_dir>`, and you should have the following checksums:
+```
+7e71cb65c4be784aa0b2889ac6039ee8  reshard-model_part-0-shard0.pt
+c8123da04f2c25a9026ea3224d5d5022  reshard-model_part-1-shard0.pt
+45e5d10896382e5bc4a7064fcafd2b1e  reshard-model_part-2-shard0.pt
+abb7296c4d2fc17420b84ca74fc3ce64  reshard-model_part-3-shard0.pt
+05dcc7ac6046f4d3f90b3d1068e6da15  reshard-model_part-4-shard0.pt
+d24dd334019060ce1ee7e625fcf6b4bd  reshard-model_part-5-shard0.pt
+fb1615ce0bbe89cc717f3e5079ee2655  reshard-model_part-6-shard0.pt
+2f3124432d2dbc6aebfca06be4b791c2  reshard-model_part-7-shard0.pt
+```
+
+Copy `flat-meta.json` to `<output_dir>`.
+
+Then cd to this dir, and we unflatten parameters.
+
+```shell
+bash unflat.sh <output_dir>/ <new_output_dir>/
+```
+
+Finally, you will get 8 files in `<new_output_dir>` with following checksums:
+```
+6169c59d014be95553c89ec01b8abb62  reshard-model_part-0.pt
+58868105da3d74a528a548fdb3a8cff6  reshard-model_part-1.pt
+69b255dc5a49d0eba9e4b60432cda90b  reshard-model_part-2.pt
+002c052461ff9ffb0cdac3d5906f41f2  reshard-model_part-3.pt
+6d57f72909320d511ffd5f1c668b2beb  reshard-model_part-4.pt
+93c8c4041cdc0c7907cc7afcf15cec2a  reshard-model_part-5.pt
+5d63b8750d827a1aa7c8ae5b02a3a2ca  reshard-model_part-6.pt
+f888bd41e009096804fe9a4b48c7ffe8  reshard-model_part-7.pt
+```
+

examples/tutorial/opt/inference/script/process-opt-175b/convert_ckpt.py

@@ -0,0 +1,55 @@
+import argparse
+import json
+import os
+import re
+from collections import defaultdict
+
+import numpy as np
+import torch
+
+
+def load_json(path: str):
+    with open(path) as f:
+        return json.load(f)
+
+
+def parse_shape_info(flat_dir: str):
+    data = load_json(os.path.join(flat_dir, 'shape.json'))
+    flat_info = defaultdict(lambda: defaultdict(list))
+    for k, shape in data.items():
+        matched = re.match(r'decoder.layers.\d+', k)
+        if matched is None:
+            flat_key = 'flat_param_0'
+        else:
+            flat_key = f'{matched[0]}.flat_param_0'
+        flat_info[flat_key]['names'].append(k)
+        flat_info[flat_key]['shapes'].append(shape)
+        flat_info[flat_key]['numels'].append(int(np.prod(shape)))
+    return flat_info
+
+
+def convert(flat_dir: str, output_dir: str, part: int):
+    flat_path = os.path.join(flat_dir, f'reshard-model_part-{part}-shard0.pt')
+    output_path = os.path.join(output_dir, f'reshard-model_part-{part}.pt')
+    flat_meta = load_json(os.path.join(flat_dir, 'flat-meta.json'))
+    flat_sd = torch.load(flat_path)
+    print(f'Loaded flat state dict from {flat_path}')
+    output_sd = {}
+    for flat_key, param_meta in flat_meta.items():
+        flat_param = flat_sd['model'][flat_key]
+        assert sum(param_meta['numels']) == flat_param.numel(
+        ), f'flat {flat_key} {flat_param.numel()} vs {sum(param_meta["numels"])}'
+        for name, shape, param in zip(param_meta['names'], param_meta['shapes'], flat_param.split(param_meta['numels'])):
+            output_sd[name] = param.view(shape)
+
+    torch.save(output_sd, output_path)
+    print(f'Saved unflat state dict to {output_path}')
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('flat_dir')
+    parser.add_argument('output_dir')
+    parser.add_argument('part', type=int)
+    args = parser.parse_args()
+    convert(args.flat_dir, args.output_dir, args.part)

examples/tutorial/opt/inference/script/process-opt-175b/unflat.sh

@@ -0,0 +1,7 @@
+#!/usr/bin/env sh
+
+for i in $(seq 0 7); do
+    python convert_ckpt.py $1 $2 ${i} &
+done
+
+wait $(jobs -p)

examples/tutorial/opt/inference/script/processing_ckpt_66b.py

@@ -0,0 +1,55 @@
+import os
+import torch
+from multiprocessing import Pool
+
+# download pytorch model ckpt in https://huggingface.co/facebook/opt-66b/tree/main
+# you can use whether wget or git lfs
+
+path = "/path/to/your/ckpt"
+new_path = "/path/to/the/processed/ckpt/"
+
+assert os.path.isdir(path)
+files = []
+for filename in os.listdir(path):
+    filepath = os.path.join(path, filename)
+    if os.path.isfile(filepath):
+        files.append(filepath)
+
+with Pool(14) as pool:
+    ckpts = pool.map(torch.load, files)
+
+restored = {}
+for ckpt in ckpts:
+    for k,v in ckpt.items():
+        if(k[0] == 'm'):
+            k = k[6:]      
+        if(k == "lm_head.weight"):
+            k = "head.dense.weight"
+        if(k == "decoder.final_layer_norm.weight"):
+            k = "decoder.layer_norm.weight"
+        if(k == "decoder.final_layer_norm.bias"):
+            k = "decoder.layer_norm.bias"
+        restored[k] = v
+restored["decoder.version"] = "0.0"
+
+
+split_num = len(restored.keys()) // 60
+count = 0
+file_count = 1
+tmp = {}
+for k,v in restored.items():
+    print(k)
+    tmp[k] = v
+    count = count + 1    
+    if(count == split_num):
+        filename = str(file_count) + "-restored.pt"
+        torch.save(tmp, os.path.join(new_path, filename))
+        file_count = file_count + 1
+        count = 0
+        tmp = {}
+
+filename = str(file_count) + "-restored.pt"
+torch.save(tmp, os.path.join(new_path, filename))
+
+
+

examples/tutorial/opt/opt/README.md

@@ -0,0 +1,53 @@
+<!---
+Copyright 2020 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+-->
+# Train OPT model with Colossal-AI
+
+## OPT
+Meta recently released [Open Pretrained Transformer (OPT)](https://github.com/facebookresearch/metaseq), a 175-Billion parameter AI language model, which stimulates AI programmers to perform various downstream tasks and application deployments.
+
+The following example of [Colossal-AI](https://github.com/hpcaitech/ColossalAI) demonstrates fine-tuning Casual Language Modelling at low cost.
+
+We are using the pre-training weights of the OPT model provided by Hugging Face Hub on the raw WikiText-2 (no tokens were replaced before
+the tokenization). This training script is adapted from the [HuggingFace Language Modelling examples](https://github.com/huggingface/transformers/tree/main/examples/pytorch/language-modeling).
+
+## Our Modifications
+We adapt the OPT training code to ColossalAI by leveraging Gemini and ZeRO DDP.
+
+## Quick Start
+You can launch training by using the following bash script
+
+```bash
+bash ./run_clm.sh <batch-size-per-gpu> <mem-cap> <model> <gpu-num>
+```
+
+- batch-size-per-gpu: number of samples fed to each GPU, default is 16
+- mem-cap: limit memory usage within a value in GB, default is 0 (no limit)
+- model: the size of the OPT model, default is `6.7b`. Acceptable values include `125m`, `350m`, `1.3b`, `2.7b`, `6.7`, `13b`, `30b`, `66b`. For `175b`, you can request
+the pretrained weights from [OPT weight downloading page](https://github.com/facebookresearch/metaseq/tree/main/projects/OPT).
+- gpu-num: the number of GPUs to use, default is 1.
+
+## Remarkable Performance
+On a single GPU, Colossal-AI’s automatic strategy provides remarkable performance gains from the ZeRO Offloading strategy by Microsoft DeepSpeed.
+Users can experience up to a 40% speedup, at a variety of model scales. However, when using a traditional deep learning training framework like PyTorch, a single GPU can no longer support the training of models at such a scale.
+
+<p align="center">
+<img src="https://raw.githubusercontent.com/hpcaitech/public_assets/main/colossalai/img/OPT.png" width=1000/>
+</p>
+
+Adopting the distributed training strategy with 8 GPUs is as simple as adding a `-nprocs 8` to the training command of Colossal-AI!
+
+More details about behind the scenes can be found on the corresponding [blog](https://medium.com/@yangyou_berkeley/colossal-ai-seamlessly-accelerates-large-models-at-low-costs-with-hugging-face-4d1a887e500d),
+and a detailed tutorial will be added in [Documentation](https://www.colossalai.org/docs/get_started/installation) very soon.

examples/tutorial/opt/opt/benchmark.sh

@@ -0,0 +1,21 @@
+export BS=16
+export MEMCAP=0
+export MODEL="6.7b"
+export GPUNUM=1
+
+for MODEL in "6.7b" "13b" "1.3b"
+do
+for GPUNUM in 8 1
+do
+for BS in 16 24 32 8
+do
+for MEMCAP in 0 40
+do
+pkill -9 torchrun
+pkill -9 python
+
+bash ./run_clm.sh $BS $MEMCAP $MODEL $GPUNUM
+done
+done
+done
+done

examples/tutorial/opt/opt/colossalai_zero.py

@@ -0,0 +1,6 @@
+from colossalai.zero.shard_utils import TensorShardStrategy
+
+zero = dict(model_config=dict(shard_strategy=TensorShardStrategy(),
+                              tensor_placement_policy="auto",
+                              reuse_fp16_shard=True),
+            optimizer_config=dict(gpu_margin_mem_ratio=0.8, initial_scale=16384))

examples/tutorial/opt/opt/context.py

@@ -0,0 +1,32 @@
+import torch.distributed as dist
+
+from colossalai.context import ParallelMode
+from colossalai.core import global_context as gpc
+
+
+class barrier_context():
+    """
+    This context manager is used to allow one process to execute while blocking all
+    other processes in the same process group. This is often useful when downloading is required
+    as we only want to download in one process to prevent file corruption.
+    Args:
+        executor_rank (int): the process rank to execute without blocking, all other processes will be blocked
+        parallel_mode (ParallelMode): the parallel mode corresponding to a process group
+    Usage:
+        with barrier_context():
+            dataset = CIFAR10(root='./data', download=True)
+    """
+
+    def __init__(self, executor_rank: int = 0, parallel_mode: ParallelMode = ParallelMode.GLOBAL):
+        # the class name is lowercase by convention
+        current_rank = gpc.get_local_rank(parallel_mode=parallel_mode)
+        self.should_block = current_rank != executor_rank
+        self.group = gpc.get_group(parallel_mode=parallel_mode)
+
+    def __enter__(self):
+        if self.should_block:
+            dist.barrier(group=self.group)
+
+    def __exit__(self, exc_type, exc_value, exc_traceback):
+        if not self.should_block:
+            dist.barrier(group=self.group)

examples/tutorial/opt/opt/requirements.txt

@@ -0,0 +1,6 @@
+colossalai
+torch >= 1.8.1
+datasets >= 1.8.0
+sentencepiece != 0.1.92
+protobuf
+accelerate == 0.13.2

examples/tutorial/opt/opt/run_clm.py

@@ -0,0 +1,596 @@
+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Fine-tuning the library models for causal language modeling (GPT, GPT-2, CTRL, ...)
+on a text file or a dataset without using HuggingFace Trainer.
+
+Here is the full list of checkpoints on the hub that can be fine-tuned by this script:
+https://huggingface.co/models?filter=text-generation
+"""
+# You can also adapt this script on your own causal language modeling task. Pointers for this are left as comments.
+
+import math
+import os
+import time
+from itertools import chain
+
+import datasets
+import torch
+import torch.distributed as dist
+from accelerate.utils import set_seed
+from context import barrier_context
+from datasets import load_dataset
+from packaging import version
+from torch.utils.data import DataLoader
+from tqdm.auto import tqdm
+
+import colossalai
+import transformers
+from colossalai.context import ParallelMode
+from colossalai.core import global_context as gpc
+from colossalai.logging import disable_existing_loggers, get_dist_logger
+from colossalai.nn.optimizer import HybridAdam
+from colossalai.nn.parallel import ZeroDDP
+from colossalai.tensor import ProcessGroup
+from colossalai.utils import get_current_device, get_dataloader
+from colossalai.utils.model.colo_init_context import ColoInitContext
+from colossalai.zero import ZeroOptimizer
+from transformers import (
+    CONFIG_MAPPING,
+    MODEL_MAPPING,
+    AutoConfig,
+    AutoTokenizer,
+    GPT2Tokenizer,
+    OPTForCausalLM,
+    SchedulerType,
+    default_data_collator,
+    get_scheduler,
+)
+from transformers.utils.versions import require_version
+
+require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/language-modeling/requirements.txt")
+
+MODEL_CONFIG_CLASSES = list(MODEL_MAPPING.keys())
+MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
+
+
+def get_time_stamp():
+    torch.cuda.synchronize()
+    return time.time()
+
+
+def parse_args():
+    parser = colossalai.get_default_parser()
+    parser.add_argument(
+        "--dataset_name",
+        type=str,
+        default=None,
+        help="The name of the dataset to use (via the datasets library).",
+    )
+    parser.add_argument(
+        "--dataset_config_name",
+        type=str,
+        default=None,
+        help="The configuration name of the dataset to use (via the datasets library).",
+    )
+    parser.add_argument("--train_file",
+                        type=str,
+                        default=None,
+                        help="A csv or a json file containing the training data.")
+    parser.add_argument("--validation_file",
+                        type=str,
+                        default=None,
+                        help="A csv or a json file containing the validation data.")
+    parser.add_argument(
+        "--validation_split_percentage",
+        default=5,
+        help="The percentage of the train set used as validation set in case there's no validation split",
+    )
+    parser.add_argument(
+        "--model_name_or_path",
+        type=str,
+        help="Path to pretrained model or model identifier from huggingface.co/models.",
+        required=True,
+    )
+    parser.add_argument(
+        "--config_name",
+        type=str,
+        default=None,
+        help="Pretrained config name or path if not the same as model_name",
+    )
+    parser.add_argument(
+        "--tokenizer_name",
+        type=str,
+        default=None,
+        help="Pretrained tokenizer name or path if not the same as model_name",
+    )
+    parser.add_argument(
+        "--use_slow_tokenizer",
+        action="store_true",
+        help="If passed, will use a slow tokenizer (not backed by the 🤗 Tokenizers library).",
+    )
+    parser.add_argument(
+        "--per_device_train_batch_size",
+        type=int,
+        default=8,
+        help="Batch size (per device) for the training dataloader.",
+    )
+    parser.add_argument(
+        "--per_device_eval_batch_size",
+        type=int,
+        default=8,
+        help="Batch size (per device) for the evaluation dataloader.",
+    )
+    parser.add_argument(
+        "--learning_rate",
+        type=float,
+        default=5e-5,
+        help="Initial learning rate (after the potential warmup period) to use.",
+    )
+    parser.add_argument("--weight_decay", type=float, default=0.0, help="Weight decay to use.")
+    parser.add_argument("--num_train_epochs", type=int, default=3, help="Total number of training epochs to perform.")
+    parser.add_argument(
+        "--max_train_steps",
+        type=int,
+        default=None,
+        help="Total number of training steps to perform. If provided, overrides num_train_epochs.",
+    )
+    parser.add_argument(
+        "--gradient_accumulation_steps",
+        type=int,
+        default=1,
+        help="Number of updates steps to accumulate before performing a backward/update pass.",
+    )
+    parser.add_argument(
+        "--lr_scheduler_type",
+        type=SchedulerType,
+        default="linear",
+        help="The scheduler type to use.",
+        choices=["linear", "cosine", "cosine_with_restarts", "polynomial", "constant", "constant_with_warmup"],
+    )
+    parser.add_argument("--num_warmup_steps",
+                        type=int,
+                        default=0,
+                        help="Number of steps for the warmup in the lr scheduler.")
+    parser.add_argument("--output_dir", type=str, default=None, help="Where to store the final model.")
+    parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
+    parser.add_argument(
+        "--model_type",
+        type=str,
+        default=None,
+        help="Model type to use if training from scratch.",
+        choices=MODEL_TYPES,
+    )
+    parser.add_argument(
+        "--block_size",
+        type=int,
+        default=None,
+        help=("Optional input sequence length after tokenization. The training dataset will be truncated in block of"
+              " this size for training. Default to the model max input length for single sentence inputs (take into"
+              " account special tokens)."),
+    )
+    parser.add_argument(
+        "--preprocessing_num_workers",
+        type=int,
+        default=None,
+        help="The number of processes to use for the preprocessing.",
+    )
+    parser.add_argument("--overwrite_cache",
+                        type=bool,
+                        default=False,
+                        help="Overwrite the cached training and evaluation sets")
+    parser.add_argument("--no_keep_linebreaks",
+                        action="store_true",
+                        help="Do not keep line breaks when using TXT files.")
+    parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
+    parser.add_argument("--hub_model_id",
+                        type=str,
+                        help="The name of the repository to keep in sync with the local `output_dir`.")
+    parser.add_argument("--hub_token", type=str, help="The token to use to push to the Model Hub.")
+    parser.add_argument(
+        "--checkpointing_steps",
+        type=str,
+        default=None,
+        help="Whether the various states should be saved at the end of every n steps, or 'epoch' for each epoch.",
+    )
+    parser.add_argument(
+        "--resume_from_checkpoint",
+        type=str,
+        default=None,
+        help="If the training should continue from a checkpoint folder.",
+    )
+    parser.add_argument(
+        "--with_tracking",
+        action="store_true",
+        help="Whether to enable experiment trackers for logging.",
+    )
+    parser.add_argument(
+        "--report_to",
+        type=str,
+        default="all",
+        help=('The integration to report the results and logs to. Supported platforms are `"tensorboard"`,'
+              ' `"wandb"` and `"comet_ml"`. Use `"all"` (default) to report to all integrations.'
+              "Only applicable when `--with_tracking` is passed."),
+    )
+
+    parser.add_argument("--mem_cap", type=int, default=0, help="use mem cap")
+    parser.add_argument("--init_in_cpu", action='store_true', default=False, help="init training model in cpu")
+    args = parser.parse_args()
+
+    # Sanity checks
+    if args.dataset_name is None and args.train_file is None and args.validation_file is None:
+        raise ValueError("Need either a dataset name or a training/validation file.")
+    else:
+        if args.train_file is not None:
+            extension = args.train_file.split(".")[-1]
+            assert extension in ["csv", "json", "txt"], "`train_file` should be a csv, json or txt file."
+        if args.validation_file is not None:
+            extension = args.validation_file.split(".")[-1]
+            assert extension in ["csv", "json", "txt"], "`validation_file` should be a csv, json or txt file."
+
+    if args.push_to_hub:
+        assert args.output_dir is not None, "Need an `output_dir` to create a repo when `--push_to_hub` is passed."
+
+    return args
+
+
+def colo_memory_cap(size_in_GB):
+    from colossalai.utils import colo_device_memory_capacity, colo_set_process_memory_fraction, get_current_device
+    cuda_capacity = colo_device_memory_capacity(get_current_device())
+    if size_in_GB * (1024**3) < cuda_capacity:
+        colo_set_process_memory_fraction(size_in_GB * (1024**3) / cuda_capacity)
+        print("Using {} GB of GPU memory".format(size_in_GB))
+
+
+def main():
+    args = parse_args()
+    disable_existing_loggers()
+    colossalai.launch_from_torch(config=dict())
+    logger = get_dist_logger()
+    is_main_process = dist.get_rank() == 0
+
+    if is_main_process:
+        datasets.utils.logging.set_verbosity_warning()
+        transformers.utils.logging.set_verbosity_info()
+    else:
+        datasets.utils.logging.set_verbosity_error()
+        transformers.utils.logging.set_verbosity_error()
+
+    if args.mem_cap > 0:
+        colo_memory_cap(args.mem_cap)
+
+    # If passed along, set the training seed now.
+    if args.seed is not None:
+        set_seed(args.seed)
+        logger.info(f"Rank {dist.get_rank()}: random seed is set to {args.seed}")
+
+    # Handle the repository creation
+    with barrier_context():
+        if args.output_dir is not None:
+            os.makedirs(args.output_dir, exist_ok=True)
+
+    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
+    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
+    # (the dataset will be downloaded automatically from the datasets Hub).
+    #
+    # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
+    # 'text' is found. You can easily tweak this behavior (see below).
+    #
+    # In distributed training, the load_dataset function guarantee that only one local process can concurrently
+    # download the dataset.
+    logger.info("Start preparing dataset", ranks=[0])
+    if args.dataset_name is not None:
+        # Downloading and loading a dataset from the hub.
+        raw_datasets = load_dataset(args.dataset_name, args.dataset_config_name)
+        if "validation" not in raw_datasets.keys():
+            raw_datasets["validation"] = load_dataset(
+                args.dataset_name,
+                args.dataset_config_name,
+                split=f"train[:{args.validation_split_percentage}%]",
+            )
+            raw_datasets["train"] = load_dataset(
+                args.dataset_name,
+                args.dataset_config_name,
+                split=f"train[{args.validation_split_percentage}%:]",
+            )
+    else:
+        data_files = {}
+        dataset_args = {}
+        if args.train_file is not None:
+            data_files["train"] = args.train_file
+        if args.validation_file is not None:
+            data_files["validation"] = args.validation_file
+        extension = args.train_file.split(".")[-1]
+        if extension == "txt":
+            extension = "text"
+            dataset_args["keep_linebreaks"] = not args.no_keep_linebreaks
+        raw_datasets = load_dataset(extension, data_files=data_files, **dataset_args)
+        # If no validation data is there, validation_split_percentage will be used to divide the dataset.
+        if "validation" not in raw_datasets.keys():
+            raw_datasets["validation"] = load_dataset(
+                extension,
+                data_files=data_files,
+                split=f"train[:{args.validation_split_percentage}%]",
+                **dataset_args,
+            )
+            raw_datasets["train"] = load_dataset(
+                extension,
+                data_files=data_files,
+                split=f"train[{args.validation_split_percentage}%:]",
+                **dataset_args,
+            )
+    logger.info("Dataset is prepared", ranks=[0])
+
+    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
+    # https://huggingface.co/docs/datasets/loading_datasets.html.
+
+    # Load pretrained model and tokenizer
+    #
+    # In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
+    # download model & vocab.
+    if args.config_name:
+        config = AutoConfig.from_pretrained(args.config_name)
+    elif args.model_name_or_path:
+        config = AutoConfig.from_pretrained(args.model_name_or_path)
+    else:
+        config = CONFIG_MAPPING[args.model_type]()
+        logger.warning("You are instantiating a new config instance from scratch.")
+    logger.info("Model config has been created", ranks=[0])
+
+    if args.model_name_or_path == 'facebook/opt-13b':
+        tokenizer = GPT2Tokenizer.from_pretrained(args.model_name_or_path)
+    else:
+        print(f'load model from {args.model_name_or_path}')
+        tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, use_fast=not args.use_slow_tokenizer)
+    logger.info(f"{tokenizer.__class__.__name__} has been created", ranks=[0])
+
+    if args.init_in_cpu:
+        init_dev = torch.device('cpu')
+    else:
+        init_dev = get_current_device()
+
+    # build model
+    if args.model_name_or_path is None or args.model_name_or_path == 'facebook/opt-13b':
+        # currently, there has a bug in pretrained opt-13b
+        # we can not import it until huggingface fix it
+        logger.info("Train a new model from scratch", ranks=[0])
+        with ColoInitContext(device=init_dev):
+            model = OPTForCausalLM(config)
+    else:
+        logger.info("Finetune a pre-trained model", ranks=[0])
+        with ColoInitContext(device=init_dev):
+            model = OPTForCausalLM.from_pretrained(args.model_name_or_path,
+                                                   from_tf=bool(".ckpt" in args.model_name_or_path),
+                                                   config=config,
+                                                   local_files_only=False)
+
+    # enable graident checkpointing
+    model.gradient_checkpointing_enable()
+
+    PLACEMENT_POLICY = 'auto'
+    cai_version = colossalai.__version__
+    logger.info(f'using Colossal-AI version {cai_version}')
+    if version.parse(cai_version) > version.parse("0.1.10"):
+        from colossalai.nn.parallel import GeminiDDP
+        model = GeminiDDP(model, device=get_current_device(), placement_policy=PLACEMENT_POLICY, pin_memory=True)
+    elif version.parse(cai_version) <= version.parse("0.1.10") and version.parse(cai_version) >= version.parse("0.1.9"):
+        from colossalai.gemini import ChunkManager, GeminiManager
+        pg = ProcessGroup()
+        chunk_size = ChunkManager.search_chunk_size(model, 64 * 1024**2, 32)
+        chunk_manager = ChunkManager(chunk_size,
+                                     pg,
+                                     enable_distributed_storage=True,
+                                     init_device=GeminiManager.get_default_device(PLACEMENT_POLICY))
+        gemini_manager = GeminiManager(PLACEMENT_POLICY, chunk_manager)
+        model = ZeroDDP(model, gemini_manager)
+
+    logger.info(f'{model.__class__.__name__} has been created', ranks=[0])
+
+    # Preprocessing the datasets.
+    # First we tokenize all the texts.
+    column_names = raw_datasets["train"].column_names
+    text_column_name = "text" if "text" in column_names else column_names[0]
+
+    def tokenize_function(examples):
+        return tokenizer(examples[text_column_name])
+
+    with barrier_context(executor_rank=0, parallel_mode=ParallelMode.DATA):
+        tokenized_datasets = raw_datasets.map(
+            tokenize_function,
+            batched=True,
+            num_proc=args.preprocessing_num_workers,
+            remove_columns=column_names,
+            load_from_cache_file=not args.overwrite_cache,
+            desc="Running tokenizer on dataset",
+        )
+
+    if args.block_size is None:
+        block_size = tokenizer.model_max_length
+        if block_size > 1024:
+            logger.warning(
+                f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
+                "Picking 1024 instead. You can change that default value by passing --block_size xxx.")
+        block_size = 1024
+    else:
+        if args.block_size > tokenizer.model_max_length:
+            logger.warning(f"The block_size passed ({args.block_size}) is larger than the maximum length for the model"
+                           f"({tokenizer.model_max_length}). Using block_size={tokenizer.model_max_length}.")
+        block_size = min(args.block_size, tokenizer.model_max_length)
+
+    # Main data processing function that will concatenate all texts from our dataset and generate chunks of block_size.
+    def group_texts(examples):
+        # Concatenate all texts.
+        concatenated_examples = {k: list(chain(*examples[k])) for k in examples.keys()}
+        total_length = len(concatenated_examples[list(examples.keys())[0]])
+        # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
+        # customize this part to your needs.
+        if total_length >= block_size:
+            total_length = (total_length // block_size) * block_size
+        # Split by chunks of max_len.
+        result = {
+            k: [t[i:i + block_size] for i in range(0, total_length, block_size)
+               ] for k, t in concatenated_examples.items()
+        }
+        result["labels"] = result["input_ids"].copy()
+        return result
+
+    # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a remainder
+    # for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value might be slower
+    # to preprocess.
+    #
+    # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
+    # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
+
+    with barrier_context(executor_rank=0, parallel_mode=ParallelMode.DATA):
+        lm_datasets = tokenized_datasets.map(
+            group_texts,
+            batched=True,
+            num_proc=args.preprocessing_num_workers,
+            load_from_cache_file=not args.overwrite_cache,
+            desc=f"Grouping texts in chunks of {block_size}",
+        )
+
+    train_dataset = lm_datasets["train"]
+    eval_dataset = lm_datasets["validation"]
+
+    # Log a few random samples from the training set:
+    # for index in random.sample(range(len(train_dataset)), 3):
+    #     logger.info(f"Sample {index} of the training set: {train_dataset[index]}.")
+
+    # DataLoaders creation:
+    train_dataloader = get_dataloader(train_dataset,
+                                      shuffle=True,
+                                      add_sampler=True,
+                                      collate_fn=default_data_collator,
+                                      batch_size=args.per_device_train_batch_size)
+    eval_dataloader = DataLoader(eval_dataset,
+                                 collate_fn=default_data_collator,
+                                 batch_size=args.per_device_eval_batch_size)
+    logger.info("Dataloaders have been created", ranks=[0])
+
+    # Optimizer
+    # Split weights in two groups, one with weight decay and the other not.
+    no_decay = ["bias", "LayerNorm.weight"]
+    optimizer_grouped_parameters = [
+        {
+            "params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
+            "weight_decay": args.weight_decay,
+        },
+        {
+            "params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)],
+            "weight_decay": 0.0,
+        },
+    ]
+
+    optimizer = HybridAdam(optimizer_grouped_parameters, lr=args.learning_rate)
+    optimizer = ZeroOptimizer(optimizer, model, initial_scale=2**14)
+
+    # Scheduler and math around the number of training steps.
+    overrode_max_train_steps = False
+    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
+    if args.max_train_steps is None:
+        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+        overrode_max_train_steps = True
+
+    lr_scheduler = get_scheduler(
+        name=args.lr_scheduler_type,
+        optimizer=optimizer,
+        num_warmup_steps=args.num_warmup_steps,
+        num_training_steps=args.max_train_steps,
+    )
+
+    # We need to recalculate our total training steps as the size of the training dataloader may have changed.
+    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
+    if overrode_max_train_steps:
+        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+    # Afterwards we recalculate our number of training epochs
+    args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
+
+    # Train!
+    total_batch_size = args.per_device_train_batch_size * gpc.get_world_size(ParallelMode.DATA)
+
+    logger.info("***** Running training *****", ranks=[0])
+    logger.info(f"  Num examples = {len(train_dataset)}", ranks=[0])
+    logger.info(f"  Num Epochs = {args.num_train_epochs}", ranks=[0])
+    logger.info(f"  Instantaneous batch size per device = {args.per_device_train_batch_size}", ranks=[0])
+    logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}", ranks=[0])
+    logger.info(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}", ranks=[0])
+    logger.info(f"  Total optimization steps = {args.max_train_steps}", ranks=[0])
+
+    # Only show the progress bar once on each machine.
+    progress_bar = tqdm(range(args.max_train_steps), disable=not is_main_process)
+    completed_steps = 0
+    starting_epoch = 0
+    global_step = 0
+
+    for epoch in range(starting_epoch, args.num_train_epochs):
+
+        if completed_steps >= args.max_train_steps:
+            break
+
+        model.train()
+        for step, batch in enumerate(train_dataloader):
+            batch = {k: v.cuda() for k, v in batch.items()}
+            outputs = model(**batch)
+            loss = outputs['loss']
+            optimizer.backward(loss)
+
+            if step % args.gradient_accumulation_steps == 0 or step == len(train_dataloader) - 1:
+                optimizer.step()
+                lr_scheduler.step()
+                optimizer.zero_grad()
+                progress_bar.update(1)
+                completed_steps += 1
+
+            global_step += 1
+            logger.info("Global step {} finished".format(global_step + 1), ranks=[0])
+
+            if completed_steps >= args.max_train_steps:
+                break
+
+        model.eval()
+        losses = []
+        for step, batch in enumerate(eval_dataloader):
+            with torch.no_grad():
+                batch = {k: v.cuda() for k, v in batch.items()}
+                outputs = model(**batch)
+
+        loss = outputs['loss'].unsqueeze(0)
+        losses.append(loss)
+
+        losses = torch.cat(losses)
+        losses = losses[:len(eval_dataset)]
+        try:
+            eval_loss = torch.mean(losses)
+            perplexity = math.exp(eval_loss)
+        except OverflowError:
+            perplexity = float("inf")
+
+        logger.info(f"Epoch {epoch}: perplexity: {perplexity} eval_loss: {eval_loss}", ranks=[0])
+
+    if args.output_dir is not None:
+        model_state = model.state_dict()
+        if is_main_process:
+            torch.save(model_state, args.output_dir + '/epoch_{}_model.pth'.format(completed_steps))
+        dist.barrier()
+        # load_state = torch.load(args.output_dir + '/epoch_{}_model.pth'.format(completed_steps))
+        # model.load_state_dict(load_state, strict=False)
+
+    logger.info("Training finished", ranks=[0])
+
+
+if __name__ == "__main__":
+    main()

examples/tutorial/opt/opt/run_clm.sh

@@ -0,0 +1,22 @@
+set -x
+export BS=${1:-16}
+export MEMCAP=${2:-0}
+export MODEL=${3:-"125m"}
+export GPUNUM=${4:-1}
+
+# make directory for logs
+mkdir -p ./logs
+
+export MODLE_PATH="facebook/opt-${MODEL}"
+
+# HF_DATASETS_OFFLINE=1 TRANSFORMERS_OFFLINE=1
+torchrun \
+  --nproc_per_node ${GPUNUM} \
+  --master_port 19198 \
+  run_clm.py \
+  --dataset_name wikitext \
+  --dataset_config_name wikitext-2-raw-v1 \
+  --output_dir $PWD \
+  --mem_cap ${MEMCAP} \
+  --model_name_or_path ${MODLE_PATH} \
+  --per_device_train_batch_size ${BS} 2>&1 | tee ./logs/colo_${MODEL}_bs_${BS}_cap_${MEMCAP}_gpu_${GPUNUM}.log

examples/tutorial/opt/zero/README.md

@@ -0,0 +1,16 @@
+## Overview
+This example shows how to use ColossalAI to run huggingface GPT training with Gemini and ZeRO DDP.
+
+## GPT
+We use the huggingface transformers GPT2 model. The input data is randonly generated.
+
+## Our Modifications
+We adapt the OPT training code to ColossalAI by leveraging Gemini and ZeRO DDP.
+
+## Quick Start
+You can launch training by using the following bash script
+
+```bash
+pip install -r requirements.txt
+bash run.sh
+```

examples/tutorial/opt/zero/requirements.txt

@@ -0,0 +1,3 @@
+colossalai >= 0.1.10
+torch >= 1.8.1
+transformers >= 4.231

examples/tutorial/opt/zero/run.sh

@@ -0,0 +1 @@
+env OMP_NUM_THREADS=16 torchrun --standalone --nproc_per_node=4 train_gpt_demo.py --tp_degree=2 --placement='cpu' 2>&1 | tee run.log

examples/tutorial/opt/zero/train_gpt_demo.py

@@ -0,0 +1,241 @@
+from functools import partial
+from time import time
+
+import psutil
+import torch
+import torch.nn as nn
+from packaging import version
+
+import colossalai
+from colossalai.logging import disable_existing_loggers, get_dist_logger
+from colossalai.nn.optimizer import HybridAdam
+from colossalai.nn.parallel import ZeroDDP
+from colossalai.tensor import ColoParameter, ComputePattern, ComputeSpec, ProcessGroup, ShardSpec
+from colossalai.utils import get_current_device
+from colossalai.utils.model.colo_init_context import ColoInitContext
+from colossalai.zero import ZeroOptimizer
+from transformers import GPT2Config, GPT2LMHeadModel
+
+
+def parse_args():
+    parser = colossalai.get_default_parser()
+    parser.add_argument(
+        "--tp_degree",
+        type=int,
+        default=1,
+        help="Tensor Parallelism Degree.",
+    )
+    parser.add_argument(
+        "--placement",
+        type=str,
+        default='cpu',
+        help="Placement Policy for Gemini.",
+    )
+    args = parser.parse_args()
+    return args
+
+
+## Parameter Sharding Strategies for Tensor Parallelism
+def split_param_single_dim_tp1d(dim: int, param: ColoParameter, pg: ProcessGroup):
+    spec = (ShardSpec([dim], [pg.tp_world_size()]), ComputeSpec(ComputePattern.TP1D))
+    if param.process_group.tp_world_size() == 1:
+        param.set_process_group(pg)
+    param.set_tensor_spec(*spec)
+
+
+def split_param_row_tp1d(param: ColoParameter, pg: ProcessGroup):
+    split_param_single_dim_tp1d(0, param, pg)
+
+
+def split_param_col_tp1d(param: ColoParameter, pg: ProcessGroup):
+    split_param_single_dim_tp1d(-1, param, pg)
+
+
+## Define the Model and Loss Based on Huggingface transformers GPT2LMHeadModel
+class GPTLMModel(nn.Module):
+
+    def __init__(self,
+                 hidden_size=768,
+                 num_layers=12,
+                 num_attention_heads=12,
+                 max_seq_len=1024,
+                 vocab_size=50257,
+                 checkpoint=False):
+        super().__init__()
+        self.checkpoint = checkpoint
+        self.model = GPT2LMHeadModel(
+            GPT2Config(n_embd=hidden_size,
+                       n_layer=num_layers,
+                       n_head=num_attention_heads,
+                       n_positions=max_seq_len,
+                       n_ctx=max_seq_len,
+                       vocab_size=vocab_size))
+        if checkpoint:
+            self.model.gradient_checkpointing_enable()
+
+    def forward(self, input_ids, attention_mask):
+        # Only return lm_logits
+        return self.model(input_ids=input_ids, attention_mask=attention_mask, use_cache=not self.checkpoint)[0]
+
+
+class GPTLMLoss(nn.Module):
+
+    def __init__(self):
+        super().__init__()
+        self.loss_fn = nn.CrossEntropyLoss()
+
+    def forward(self, logits, labels):
+        shift_logits = logits[..., :-1, :].contiguous()
+        shift_labels = labels[..., 1:].contiguous()
+        # Flatten the tokens
+        return self.loss_fn(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+
+
+## Randomly Generated Data
+def get_data(batch_size, seq_len, vocab_size):
+    input_ids = torch.randint(0, vocab_size, (batch_size, seq_len), device=torch.cuda.current_device())
+    attention_mask = torch.ones_like(input_ids)
+    return input_ids, attention_mask
+
+
+def gpt2_medium(checkpoint=False):
+    return GPTLMModel(hidden_size=1024, num_layers=24, num_attention_heads=16, checkpoint=checkpoint)
+
+
+def gpt2_xl(checkpoint=True):
+    return GPTLMModel(hidden_size=1600, num_layers=48, num_attention_heads=32, checkpoint=checkpoint)
+
+
+def gpt2_10b(checkpoint=True):
+    return GPTLMModel(hidden_size=4096, num_layers=50, num_attention_heads=16, checkpoint=checkpoint)
+
+
+def get_cpu_mem():
+    return psutil.Process().memory_info().rss / 1024**2
+
+
+def get_gpu_mem():
+    return torch.cuda.memory_allocated() / 1024**2
+
+
+def get_mem_info(prefix=''):
+    return f'{prefix}GPU memory usage: {get_gpu_mem():.2f} MB, CPU memory usage: {get_cpu_mem():.2f} MB'
+
+
+def get_tflops(model_numel, batch_size, seq_len, step_time):
+    return model_numel * batch_size * seq_len * 8 / 1e12 / (step_time + 1e-12)
+
+
+# Tensor Parallel
+def tensor_parallelize(model: torch.nn.Module, pg: ProcessGroup):
+    """tensor_parallelize
+    Sharding the Model Parameters.
+
+    Args:
+        model (torch.nn.Module): a torch module to be sharded
+    """
+    for mn, module in model.named_modules():
+        for pn, param in module.named_parameters(recurse=False):
+            # set process group for all parameters
+            param.set_process_group(pg)
+
+            if 'mlp.c_fc' in mn:
+                if 'weight' in pn or 'bias' in pn:
+                    split_param_col_tp1d(param, pg)    # colmn slice
+                    # keep the shape of the output from c_fc
+                    param.compute_spec.set_output_replicate(False)
+            elif 'mlp.c_proj' in mn:
+                if 'weight' in pn:
+                    split_param_row_tp1d(param, pg)    # row slice
+            elif 'wte' in mn or 'wpe' in mn:
+                split_param_col_tp1d(param, pg)    # colmn slice
+            elif 'c_attn' in mn or 'c_proj' in mn:
+                split_param_col_tp1d(param, pg)    # colmn slice
+
+
+# Gemini + ZeRO DDP
+def gemini_zero_dpp(model: torch.nn.Module, pg: ProcessGroup, placememt_policy: str = "auto"):
+    cai_version = colossalai.__version__
+    if version.parse(cai_version) > version.parse("0.1.10"):
+        from colossalai.nn.parallel import GeminiDDP
+        model = GeminiDDP(model,
+                          device=get_current_device(),
+                          placement_policy=placememt_policy,
+                          pin_memory=True,
+                          search_range_mb=32)
+    elif version.parse(cai_version) <= version.parse("0.1.10") and version.parse(cai_version) >= version.parse("0.1.9"):
+        from colossalai.gemini import ChunkManager, GeminiManager
+        chunk_size = ChunkManager.search_chunk_size(model, 64 * 1024**2, 32)
+        gemini_manager = GeminiManager(placememt_policy, chunk_manager)
+        chunk_manager = ChunkManager(chunk_size,
+                                     pg,
+                                     enable_distributed_storage=True,
+                                     init_device=GeminiManager.get_default_device(placememt_policy))
+        model = ZeroDDP(model, gemini_manager)
+    else:
+        raise NotImplemented(f"CAI version {cai_version} is not supported")
+    return model
+
+
+def main():
+    args = parse_args()
+
+    BATCH_SIZE = 8
+    SEQ_LEN = 1024
+    VOCAB_SIZE = 50257
+    NUM_STEPS = 10
+
+    disable_existing_loggers()
+    colossalai.launch_from_torch(config={})
+
+    pg = ProcessGroup(tp_degree=args.tp_degree)
+
+    logger = get_dist_logger()
+    logger.info(get_mem_info(), ranks=[0])
+
+    # build GPT model
+    with ColoInitContext(device=get_current_device()):
+        model = gpt2_medium(checkpoint=True)
+
+    numel = sum([p.numel() for p in model.parameters()])
+    logger.info(f'Model numel: {numel}', ranks=[0])
+    get_tflops_func = partial(get_tflops, numel, BATCH_SIZE, SEQ_LEN)
+
+    # Tensor Parallelism (TP)
+    tensor_parallelize(model, pg)
+    # Gemini + ZeRO DP, Note it must be used after TP
+    model = gemini_zero_dpp(model, pg, args.placement)
+    logger.info(get_mem_info(prefix='After init model, '), ranks=[0])
+
+    # build criterion
+    criterion = GPTLMLoss()
+
+    # build optimizer
+    optimizer = HybridAdam(model.parameters(), lr=1e-3)
+    optimizer = ZeroOptimizer(optimizer, model, initial_scale=2**5)
+    logger.info(get_mem_info(prefix='After init optim, '), ranks=[0])
+
+    torch.cuda.synchronize()
+    model.train()
+    for n in range(NUM_STEPS):
+        # we just use randomly generated data here
+        input_ids, attn_mask = get_data(BATCH_SIZE, SEQ_LEN, VOCAB_SIZE)
+        optimizer.zero_grad()
+        start = time()
+        outputs = model(input_ids, attn_mask)
+        loss = criterion(outputs, input_ids)
+        logger.info(get_mem_info(prefix=f'[{n+1}/{NUM_STEPS}] Forward '), ranks=[0])
+        optimizer.backward(loss)
+        logger.info(get_mem_info(prefix=f'[{n+1}/{NUM_STEPS}] Backward '), ranks=[0])
+        optimizer.step()
+        logger.info(get_mem_info(prefix=f'[{n+1}/{NUM_STEPS}] Optimizer step '), ranks=[0])
+        step_time = time() - start
+        logger.info(
+            f'[{n+1}/{NUM_STEPS}] Loss:{loss.item():.3f}, Step time: {step_time:.3f}s, TFLOPS: {get_tflops_func(step_time):.3f}',
+            ranks=[0])
+
+    torch.cuda.synchronize()
+
+
+if __name__ == '__main__':
+    main()

examples/tutorial/sequence_parallel/README.md

@@ -0,0 +1,143 @@
+# Handson 2: Sequence Parallelism with BERT
+
+In this example, we implemented BERT with sequence parallelism. Sequence parallelism splits the input tensor and intermediate 
+activation along the sequence dimension. This method can achieve better memory efficiency and allows us to train with larger batch size and longer sequence length.
+
+Paper: [Sequence Parallelism: Long Sequence Training from System Perspective](https://arxiv.org/abs/2105.13120)
+
+## How to Prepare WikiPedia Dataset
+
+First, let's prepare the WikiPedia dataset from scratch. To generate a preprocessed dataset, we need four items:
+1. raw WikiPedia dataset
+2. wikipedia extractor (extract data from the raw dataset)
+3. vocabulary file
+4. preprocessing scripts (generate final data from extracted data)
+
+For the preprocessing script, we thank Megatron-LM for providing a preprocessing script to generate the corpus file.
+
+```python
+# download raw data 
+mkdir data && cd ./data
+wget https://dumps.wikimedia.org/enwiki/latest/enwiki-latest-pages-articles.xml.bz2
+
+# install wiki extractor
+git clone https://github.com/FrankLeeeee/wikiextractor.git
+pip install ./wikiextractor
+
+# extractmodule 
+wikiextractor --json enwiki-latest-pages-articles.xml.bz2
+cat text/*/* > ./corpus.json
+cd ..
+
+# download vocab file
+mkdir vocab && cd ./vocab
+wget https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-vocab.txt
+cd ..
+
+# preprocess some data 
+git clone https://github.com/NVIDIA/Megatron-LM.git
+cd ./Megatron-LM
+python tools/preprocess_data.py \
+    --input ../data/corpus.json \
+    --output-prefix my-bert \
+    --vocab ../vocab/bert-large-uncased-vocab.txt \
+    --dataset-impl mmap \
+    --tokenizer-type BertWordPieceLowerCase \
+    --split-sentences \
+    --workers 24
+```
+
+After running the preprocessing scripts, you will obtain two files:
+1. my-bert_text_sentence.bin
+2. my-bert_text_sentence.idx
+
+If you happen to encouter `index out of range` problem when running Megatron's script,
+this is probably because that a sentence starts with a punctuation and cannot be tokenized. A work-around is to update `Encoder.encode` method with the code below:
+
+```python
+class Encoder(object):
+    def __init__(self, args):
+        ...
+
+    def initializer(self):
+        ...
+
+    def encode(self, json_line):
+        data = json.loads(json_line)
+        ids = {}
+        for key in self.args.json_keys:
+            text = data[key]
+            doc_ids = []
+
+            # lsg: avoid sentences which start with a punctuation
+            # as it cannot be tokenized by splitter
+            if len(text) > 0 and text[0] in string.punctuation:
+                text = text[1:]
+
+            for sentence in Encoder.splitter.tokenize(text):
+                sentence_ids = Encoder.tokenizer.tokenize(sentence)
+                if len(sentence_ids) > 0:
+                    doc_ids.append(sentence_ids)
+            if len(doc_ids) > 0 and self.args.append_eod:
+                doc_ids[-1].append(Encoder.tokenizer.eod)
+            ids[key] = doc_ids
+        return ids, len(json_line)
+```
+
+## How to Train with Sequence Parallelism
+
+We provided `train.py` for you to execute training. Before invoking the script, there are several 
+steps to perform.
+
+### Step 1. Set data path and vocab path
+
+At the top of `config.py`, you can see two global variables `DATA_PATH` and `VOCAB_FILE_PATH`. 
+
+```python
+DATA_PATH = <data-path>
+VOCAB_FILE_PATH = <vocab-path>
+```
+
+`DATA_PATH` refers to the path to the data file generated by Megatron's script. For example, in the section above, you should get two data files (my-bert_text_sentence.bin and my-bert_text_sentence.idx). You just need to `DATA_PATH` to the path to the bin file without the file extension.
+
+For example, if your my-bert_text_sentence.bin is /home/Megatron-LM/my-bert_text_sentence.bin, then you should set
+
+```python
+DATA_PATH = '/home/Megatron-LM/my-bert_text_sentence'
+```
+
+The `VOCAB_FILE_PATH` refers to the path to the vocabulary downloaded when you prepare the dataset 
+(e.g. bert-large-uncased-vocab.txt).
+
+### Step 3. Make Dataset Helper
+
+Build BERT dataset helper. Requirements are `CUDA`, `g++`, `pybind11` and `make`.
+
+```python
+cd ./data/datasets
+make
+```
+
+### Step 3. Configure your parameters
+
+In the `config.py` provided, a set of parameters are defined including training scheme, model, etc.
+You can also modify the ColossalAI setting. For example, if you wish to parallelize over the 
+sequence dimension on 8 GPUs. You can change `size=4` to `size=8`. If you wish to use pipeline parallelism, you can set `pipeline=<num_of_pipeline_stages>`.
+
+### Step 4. Invoke parallel training
+
+Lastly, you can start training with sequence parallelism. How you invoke `train.py` depends on your 
+machine setting.
+
+- If you are using a single machine with multiple GPUs, PyTorch launch utility can easily let you
+  start your script. A sample command is like below:
+
+  ```bash
+    python -m torch.distributed.launch --nproc_per_node <num_gpus_on_this_machine> --master_addr localhost --master_port 29500 train.py
+  ```
+
+- If you are using multiple machines with multiple GPUs, we suggest that you refer to `colossalai
+  launch_from_slurm` or `colossalai.launch_from_openmpi` as it is easier to use SLURM and OpenMPI 
+  to start multiple processes over multiple nodes. If you have your own launcher, you can fall back 
+  to the default `colossalai.launch` function.
+  

examples/tutorial/sequence_parallel/config.py

@@ -0,0 +1,40 @@
+from colossalai.amp import AMP_TYPE
+
+DATA_PATH = ''
+VOCAB_FILE_PATH = ''
+
+# hyper-parameters
+TRAIN_ITERS = 1000000
+DECAY_ITERS = 990000
+WARMUP_FRACTION = 0.01
+GLOBAL_BATCH_SIZE = 32    # dp world size * sentences per GPU
+EVAL_ITERS = 10
+EVAL_INTERVAL = 10
+LR = 0.0001
+MIN_LR = 1e-05
+WEIGHT_DECAY = 0.01
+SEQ_LENGTH = 512
+
+# BERT config
+DEPTH = 12
+NUM_ATTENTION_HEADS = 12
+HIDDEN_SIZE = 768
+
+# model config
+ADD_BINARY_HEAD = False
+
+# random seed
+SEED = 1234
+
+# pipeline config
+# only enabled when pipeline > 1
+NUM_MICRO_BATCHES = 4
+
+# colossalai config
+parallel = dict(pipeline=1, tensor=dict(size=4, mode='sequence'))
+
+fp16 = dict(mode=AMP_TYPE.NAIVE, verbose=True)
+
+clip_grad_norm = 1.0
+
+gradient_handler = [dict(type='SequenceParallelGradientHandler')]

examples/tutorial/sequence_parallel/data/__init__.py

@@ -0,0 +1,102 @@
+from colossalai.context.parallel_context import ParallelContext
+from colossalai.core import global_context as gpc
+from colossalai.logging import get_dist_logger
+from colossalai.context import ParallelMode
+from .datasets.data_samplers import build_pretraining_data_loader
+from .datasets.builder import build_train_valid_test_datasets
+import torch
+
+
+def cyclic_iter(iter):
+    while True:
+        for x in iter:
+            yield x
+
+
+def build_train_valid_test_data_iterators(train_iters,
+                                          global_batch_size,
+                                          eval_interval,
+                                          eval_iters,
+                                          dataloader_type='single',
+                                          **kwargs
+                                          ):
+    (train_dataloader, valid_dataloader, test_dataloader) = (None, None, None)
+
+    logger = get_dist_logger()
+    logger.info('> building train, validation, and test datasets ...', ranks=[0])
+
+    # Backward compatibility, assume fixed batch size.
+    # if iteration > 0 and consumed_train_samples == 0:
+    #     assert train_samples is None, \
+    #         'only backward compatibility support for iteration-based training'
+    #     consumed_train_samples = iteration * global_batch_size
+    # if iteration > 0 and consumed_valid_samples == 0:
+    #     if train_samples is None:
+    #         consumed_valid_samples = (iteration // eval_interval) * \
+    #             eval_iters * global_batch_size
+
+    # Data loader only on rank 0 of each model parallel group.
+    if not gpc.is_initialized(ParallelMode.TENSOR) or gpc.get_local_rank(ParallelMode.TENSOR) == 0:
+
+        # Number of train/valid/test samples.
+        train_samples = train_iters * global_batch_size
+        eval_iters_ = (train_iters // eval_interval + 1) * eval_iters
+        test_iters = eval_iters
+        train_val_test_num_samples = [train_samples,
+                                      eval_iters_ * global_batch_size,
+                                      test_iters * global_batch_size]
+        logger.info(' > datasets target sizes (minimum size):')
+        logger.info('    train:      {}'.format(train_val_test_num_samples[0]), ranks=[0])
+        logger.info('    validation: {}'.format(train_val_test_num_samples[1]), ranks=[0])
+        logger.info('    test:       {}'.format(train_val_test_num_samples[2]), ranks=[0])
+
+        # Build the datasets.
+        train_ds, valid_ds, test_ds = build_train_valid_test_datasets(
+            train_valid_test_num_samples=train_val_test_num_samples, **kwargs)
+
+        # Build dataloaders.
+        dp_size = gpc.get_world_size(ParallelMode.DATA)
+        train_dataloader = build_pretraining_data_loader(
+            train_ds, consumed_samples=0, micro_batch_size=global_batch_size//dp_size)
+        valid_dataloader = build_pretraining_data_loader(
+            valid_ds, consumed_samples=0, micro_batch_size=global_batch_size//dp_size)
+        test_dataloader = build_pretraining_data_loader(test_ds, 0, micro_batch_size=global_batch_size//dp_size)
+
+        # Flags to know if we need to do training/validation/testing.
+        do_train = train_dataloader is not None and train_iters > 0
+        do_valid = valid_dataloader is not None and eval_iters > 0
+        do_test = test_dataloader is not None and eval_iters > 0
+        # Need to broadcast num_tokens and num_type_tokens.
+        flags = torch.cuda.LongTensor(
+            [int(do_train), int(do_valid), int(do_test)])
+    else:
+        flags = torch.cuda.LongTensor([0, 0, 0])
+
+    # Broadcast num tokens.
+    torch.distributed.broadcast(flags,
+                                gpc.get_ranks_in_group(ParallelMode.TENSOR)[0],
+                                group=gpc.get_group(ParallelMode.TENSOR))
+
+    # Build iterators.
+    dl_type = dataloader_type
+    assert dl_type in ['single', 'cyclic']
+
+    if train_dataloader is not None:
+        train_data_iterator = iter(train_dataloader) if dl_type == 'single' \
+            else iter(cyclic_iter(train_dataloader))
+    else:
+        train_data_iterator = None
+
+    if valid_dataloader is not None:
+        valid_data_iterator = iter(valid_dataloader) if dl_type == 'single' \
+            else iter(cyclic_iter(valid_dataloader))
+    else:
+        valid_data_iterator = None
+
+    if test_dataloader is not None:
+        test_data_iterator = iter(test_dataloader) if dl_type == 'single' \
+            else iter(cyclic_iter(test_dataloader))
+    else:
+        test_data_iterator = None
+
+    return train_data_iterator, valid_data_iterator, test_data_iterator

examples/tutorial/sequence_parallel/data/bert_helper.py

@@ -0,0 +1,165 @@
+from colossalai.core import global_context as gpc
+from colossalai.context import ParallelMode
+import torch
+
+_MAX_DATA_DIM = 5
+
+
+def _build_key_size_numel_dictionaries(keys, data):
+    """Build the size on rank 0 and broadcast."""
+    max_dim = _MAX_DATA_DIM
+    sizes = [0 for _ in range(max_dim) for _ in keys]
+
+    # Pack the sizes on rank zero.
+    if not gpc.is_initialized(ParallelMode.TENSOR) or gpc.get_local_rank(ParallelMode.TENSOR) == 0:
+        offset = 0
+        for key in keys:
+            assert data[key].dim() < max_dim, 'you should increase MAX_DATA_DIM'
+            size = data[key].size()
+            for i, s in enumerate(size):
+                sizes[i + offset] = s
+            offset += max_dim
+
+    # Move to GPU and broadcast.
+    sizes_cuda = torch.cuda.LongTensor(sizes)
+    torch.distributed.broadcast(sizes_cuda, gpc.get_ranks_in_group(ParallelMode.TENSOR)[0],
+                                group=gpc.get_group(ParallelMode.TENSOR))
+
+    # Move back to cpu and unpack.
+    sizes_cpu = sizes_cuda.cpu()
+    key_size = {}
+    key_numel = {}
+    total_numel = 0
+    offset = 0
+    for key in keys:
+        i = 0
+        size = []
+        numel = 1
+        while sizes_cpu[offset + i] > 0:
+            this_size = sizes_cpu[offset + i]
+            size.append(this_size)
+            numel *= this_size
+            i += 1
+        key_size[key] = size
+        key_numel[key] = numel
+        total_numel += numel
+        offset += max_dim
+
+    return key_size, key_numel, total_numel
+
+
+def broadcast_data(keys, data, datatype):
+    """Broadcast data from rank zero of each model parallel group to the
+    members of the same model parallel group.
+
+    Arguments:
+        keys: list of keys in the data dictionary to be broadcasted
+        data: data dictionary of string keys and cpu tensor values.
+        datatype: torch data type of all tensors in data associated
+                  with keys.
+    """
+    # Build (key, size) and (key, number of elements) dictionaries along
+    # with the total number of elements on all ranks.
+    key_size, key_numel, total_numel = _build_key_size_numel_dictionaries(keys,
+                                                                          data)
+
+    # Pack on rank zero.
+    if not gpc.is_initialized(ParallelMode.TENSOR) or gpc.get_local_rank(ParallelMode.TENSOR) == 0:
+        # Check that all keys have the same data type.
+        # Flatten the data associated with the keys
+        flatten_data = torch.cat(
+            [data[key].contiguous().view(-1) for key in keys], dim=0).cuda()
+    else:
+        flatten_data = torch.empty(total_numel,
+                                   device=torch.cuda.current_device(),
+                                   dtype=datatype)
+
+    # Broadcast
+    torch.distributed.broadcast(flatten_data,
+                                gpc.get_ranks_in_group(ParallelMode.TENSOR)[0],
+                                group=gpc.get_group(ParallelMode.TENSOR))
+
+    # Unpack
+    output = {}
+    offset = 0
+    for key in keys:
+        size = key_size[key]
+        numel = key_numel[key]
+        output[key] = flatten_data.narrow(0, offset, numel).view(size)
+        offset += numel
+
+    return output
+
+
+def get_batch(data_iterator):
+    """Build the batch."""
+
+    # Items and their type.
+    keys = ['text', 'types', 'labels', 'is_random', 'loss_mask', 'padding_mask']
+    datatype = torch.int64
+
+    # Broadcast data.
+    if data_iterator is not None:
+        data = next(data_iterator)
+    else:
+        data = None
+    data_b = broadcast_data(keys, data, datatype)
+
+    # Unpack.
+    tokens = data_b['text'].long()
+    types = data_b['types'].long()
+    sentence_order = data_b['is_random'].long()
+    loss_mask = data_b['loss_mask'].float()
+    lm_labels = data_b['labels'].long()
+    padding_mask = data_b['padding_mask'].long()
+
+    return tokens, types, sentence_order, loss_mask, lm_labels, padding_mask
+
+
+def get_batch_for_sequence_parallel(data_iterator):
+    """Build the batch."""
+
+    # Items and their type.
+    keys = ['text', 'types', 'labels', 'is_random', 'loss_mask', 'padding_mask']
+    datatype = torch.int64
+
+    # Broadcast data.
+    if data_iterator is not None:
+        data = next(data_iterator)
+    else:
+        data = None
+
+    # unpack
+    data_b = broadcast_data(keys, data, datatype)
+
+    # # get tensor parallel local rank
+    global_rank = torch.distributed.get_rank()
+    local_world_size = 1 if not gpc.is_initialized(ParallelMode.TENSOR) else gpc.get_world_size(ParallelMode.TENSOR)
+    local_rank = global_rank % local_world_size
+    seq_length = data_b['text'].size(1)
+    sub_seq_length = seq_length // local_world_size
+    sub_seq_start = local_rank * sub_seq_length
+    sub_seq_end = (local_rank+1) * sub_seq_length
+    #
+    # # Unpack.
+    tokens = data_b['text'][:, sub_seq_start:sub_seq_end].long()
+    types = data_b['types'][:, sub_seq_start:sub_seq_end].long()
+    sentence_order = data_b['is_random'].long()
+    loss_mask = data_b['loss_mask'][:, sub_seq_start:sub_seq_end].float()
+    lm_labels = data_b['labels'][:, sub_seq_start:sub_seq_end].long()
+    padding_mask = data_b['padding_mask'].long()
+
+    return tokens, types, sentence_order, loss_mask, lm_labels, padding_mask
+
+
+class SequenceParallelDataIterator:
+
+    def __init__(self, data_iter):
+        self.data_iter = data_iter
+    
+
+    def __iter__(self):
+        return self.data_iter
+
+    def __next__(self):
+        return get_batch_for_sequence_parallel(self.data_iter)

examples/tutorial/sequence_parallel/data/datasets/Makefile

@@ -0,0 +1,9 @@
+CXXFLAGS += -O3 -Wall -shared -std=c++11 -fPIC -fdiagnostics-color
+CPPFLAGS += $(shell python3 -m pybind11 --includes)
+LIBNAME = helpers
+LIBEXT = $(shell python3-config --extension-suffix)
+
+default: $(LIBNAME)$(LIBEXT)
+
+%$(LIBEXT): %.cpp
+	$(CXX) $(CXXFLAGS) $(CPPFLAGS) $< -o $@

examples/tutorial/sequence_parallel/data/datasets/__init__.py

@@ -0,0 +1 @@
+from . import indexed_dataset

examples/tutorial/sequence_parallel/data/datasets/bert_dataset.py

@@ -0,0 +1,225 @@
+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""BERT Style dataset."""
+
+from colossalai.logging import get_dist_logger
+import numpy as np
+import torch
+from torch.utils.data import Dataset
+
+from ..tokenizer import get_tokenizer
+from .dataset_utils import (get_a_and_b_segments, truncate_segments, create_tokens_and_tokentypes,
+                            create_masked_lm_predictions, pad_and_convert_to_numpy)
+from colossalai.core import global_context as gpc
+from colossalai.context import ParallelMode
+import time
+import os
+from . import helpers
+
+
+class BertDataset(Dataset):
+
+    def __init__(self, name, indexed_dataset, data_prefix, num_epochs, max_num_samples, masked_lm_prob, max_seq_length,
+                 short_seq_prob, seed, binary_head):
+
+        # Params to store.
+        self.name = name
+        self.seed = seed
+        self.masked_lm_prob = masked_lm_prob
+        self.max_seq_length = max_seq_length
+        self.binary_head = binary_head
+
+        # Dataset.
+        self.indexed_dataset = indexed_dataset
+
+        # Build the samples mapping.
+        self.samples_mapping = get_samples_mapping_(
+            self.indexed_dataset,
+            data_prefix,
+            num_epochs,
+            max_num_samples,
+            self.max_seq_length - 3,    # account for added tokens,
+            short_seq_prob,
+            self.seed,
+            self.name,
+            self.binary_head)
+
+        # Vocab stuff.
+        tokenizer = get_tokenizer()
+        self.vocab_id_list = list(tokenizer.inv_vocab.keys())
+        self.vocab_id_to_token_dict = tokenizer.inv_vocab
+        self.cls_id = tokenizer.cls
+        self.sep_id = tokenizer.sep
+        self.mask_id = tokenizer.mask
+        self.pad_id = tokenizer.pad
+
+    def __len__(self):
+        return self.samples_mapping.shape[0]
+
+    def __getitem__(self, idx):
+        start_idx, end_idx, seq_length = self.samples_mapping[idx]
+        sample = [self.indexed_dataset[i] for i in range(start_idx, end_idx)]
+        # Note that this rng state should be numpy and not python since
+        # python randint is inclusive whereas the numpy one is exclusive.
+        # We % 2**32 since numpy requires the seed to be between 0 and 2**32 - 1
+        np_rng = np.random.RandomState(seed=((self.seed + idx) % 2**32))
+        return build_training_sample(
+            sample,
+            seq_length,
+            self.max_seq_length,    # needed for padding
+            self.vocab_id_list,
+            self.vocab_id_to_token_dict,
+            self.cls_id,
+            self.sep_id,
+            self.mask_id,
+            self.pad_id,
+            self.masked_lm_prob,
+            np_rng,
+            self.binary_head)
+
+
+def get_samples_mapping_(indexed_dataset, data_prefix, num_epochs, max_num_samples, max_seq_length, short_seq_prob,
+                         seed, name, binary_head):
+    logger = get_dist_logger()
+    if not num_epochs:
+        if not max_num_samples:
+            raise ValueError("Need to specify either max_num_samples "
+                             "or num_epochs")
+        num_epochs = np.iinfo(np.int32).max - 1
+    if not max_num_samples:
+        max_num_samples = np.iinfo(np.int64).max - 1
+
+    # Filename of the index mapping
+    indexmap_filename = data_prefix
+    indexmap_filename += '_{}_indexmap'.format(name)
+    if num_epochs != (np.iinfo(np.int32).max - 1):
+        indexmap_filename += '_{}ep'.format(num_epochs)
+    if max_num_samples != (np.iinfo(np.int64).max - 1):
+        indexmap_filename += '_{}mns'.format(max_num_samples)
+    indexmap_filename += '_{}msl'.format(max_seq_length)
+    indexmap_filename += '_{:0.2f}ssp'.format(short_seq_prob)
+    indexmap_filename += '_{}s'.format(seed)
+    indexmap_filename += '.npy'
+
+    # Build the indexed mapping if not exist.
+    if torch.distributed.get_rank() == 0 and \
+       not os.path.isfile(indexmap_filename):
+        print(' > WARNING: could not find index map file {}, building '
+              'the indices on rank 0 ...'.format(indexmap_filename))
+
+        # Make sure the types match the helpers input types.
+        assert indexed_dataset.doc_idx.dtype == np.int64
+        assert indexed_dataset.sizes.dtype == np.int32
+
+        # Build samples mapping
+        verbose = torch.distributed.get_rank() == 0
+        start_time = time.time()
+        logger.info('\n > building samples index mapping for {} ...'.format(name), ranks=[0])
+        # First compile and then import.
+        samples_mapping = helpers.build_mapping(indexed_dataset.doc_idx, indexed_dataset.sizes, num_epochs,
+                                                max_num_samples, max_seq_length, short_seq_prob, seed, verbose,
+                                                2 if binary_head else 1)
+        logger.info('\n > done building samples index maping', ranks=[0])
+        np.save(indexmap_filename, samples_mapping, allow_pickle=True)
+        logger.info('\n > saved the index mapping in {}'.format(indexmap_filename), ranks=[0])
+        # Make sure all the ranks have built the mapping
+        logger.info('\n > elapsed time to build and save samples mapping '
+                    '(seconds): {:4f}'.format(time.time() - start_time),
+                    ranks=[0])
+    # This should be a barrier but nccl barrier assumes
+    # device_index=rank which is not the case for model
+    # parallel case
+    counts = torch.cuda.LongTensor([1])
+    torch.distributed.all_reduce(counts, group=gpc.get_group(ParallelMode.DATA))
+    if gpc.is_initialized(ParallelMode.PIPELINE):
+        torch.distributed.all_reduce(counts, group=gpc.get_group(ParallelMode.PIPELINE))
+    assert counts[0].item() == (torch.distributed.get_world_size() //
+                                torch.distributed.get_world_size(group=gpc.get_group(ParallelMode.SEQUENCE)))
+
+    # Load indexed dataset.
+    start_time = time.time()
+    samples_mapping = np.load(indexmap_filename, allow_pickle=True, mmap_mode='r')
+    logger.info('\n > loading indexed mapping from {}'.format(indexmap_filename) +
+                '\n    loaded indexed file in {:3.3f} seconds'.format(time.time() - start_time) +
+                '\n    total number of samples: {}'.format(samples_mapping.shape[0]),
+                ranks=[0])
+
+    return samples_mapping
+
+
+def build_training_sample(sample, target_seq_length, max_seq_length, vocab_id_list, vocab_id_to_token_dict, cls_id,
+                          sep_id, mask_id, pad_id, masked_lm_prob, np_rng, binary_head):
+    """Build training sample.
+
+    Arguments:
+        sample: A list of sentences in which each sentence is a list token ids.
+        target_seq_length: Desired sequence length.
+        max_seq_length: Maximum length of the sequence. All values are padded to
+            this length.
+        vocab_id_list: List of vocabulary ids. Used to pick a random id.
+        vocab_id_to_token_dict: A dictionary from vocab ids to text tokens.
+        cls_id: Start of example id.
+        sep_id: Separator id.
+        mask_id: Mask token id.
+        pad_id: Padding token id.
+        masked_lm_prob: Probability to mask tokens.
+        np_rng: Random number genenrator. Note that this rng state should be
+              numpy and not python since python randint is inclusive for
+              the opper bound whereas the numpy one is exclusive.
+    """
+
+    if binary_head:
+        # We assume that we have at least two sentences in the sample
+        assert len(sample) > 1
+    assert target_seq_length <= max_seq_length
+
+    # Divide sample into two segments (A and B).
+    if binary_head:
+        tokens_a, tokens_b, is_next_random = get_a_and_b_segments(sample, np_rng)
+    else:
+        tokens_a = []
+        for j in range(len(sample)):
+            tokens_a.extend(sample[j])
+        tokens_b = []
+        is_next_random = False
+
+    # Truncate to `target_sequence_length`.
+    max_num_tokens = target_seq_length
+    truncated = truncate_segments(tokens_a, tokens_b, len(tokens_a), len(tokens_b), max_num_tokens, np_rng)
+
+    # Build tokens and toketypes.
+    tokens, tokentypes = create_tokens_and_tokentypes(tokens_a, tokens_b, cls_id, sep_id)
+
+    # Masking.
+    max_predictions_per_seq = masked_lm_prob * max_num_tokens
+    (tokens, masked_positions, masked_labels,
+     _) = create_masked_lm_predictions(tokens, vocab_id_list, vocab_id_to_token_dict, masked_lm_prob, cls_id, sep_id,
+                                       mask_id, max_predictions_per_seq, np_rng)
+
+    # Padding.
+    tokens_np, tokentypes_np, labels_np, padding_mask_np, loss_mask_np \
+        = pad_and_convert_to_numpy(tokens, tokentypes, masked_positions,
+                                   masked_labels, pad_id, max_seq_length)
+
+    train_sample = {
+        'text': tokens_np,
+        'types': tokentypes_np,
+        'labels': labels_np,
+        'is_random': int(is_next_random),
+        'loss_mask': loss_mask_np,
+        'padding_mask': padding_mask_np,
+        'truncated': int(truncated)
+    }
+    return train_sample

examples/tutorial/sequence_parallel/data/datasets/blendable_dataset.py

@@ -0,0 +1,62 @@
+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Blendable dataset."""
+
+import time
+
+import numpy as np
+import torch
+
+
+class BlendableDataset(torch.utils.data.Dataset):
+
+    def __init__(self, datasets, weights):
+
+        self.datasets = datasets
+        num_datasets = len(datasets)
+        assert num_datasets == len(weights)
+
+        self.size = 0
+        for dataset in self.datasets:
+            self.size += len(dataset)
+
+        # Normalize weights.
+        weights = np.array(weights, dtype=np.float64)
+        sum_weights = np.sum(weights)
+        assert sum_weights > 0.0
+        weights /= sum_weights
+
+        # Build indices.
+        start_time = time.time()
+        assert num_datasets < 255
+        self.dataset_index = np.zeros(self.size, dtype=np.uint8)
+        self.dataset_sample_index = np.zeros(self.size, dtype=np.int64)
+
+        from . import helpers
+        helpers.build_blending_indices(self.dataset_index,
+                                       self.dataset_sample_index,
+                                       weights, num_datasets, self.size,
+                                       torch.distributed.get_rank() == 0)
+        print('> elapsed time for building blendable dataset indices: '
+              '{:.2f} (sec)'.format(time.time() - start_time))
+
+    def __len__(self):
+        return self.size
+
+    def __getitem__(self, idx):
+        dataset_idx = self.dataset_index[idx]
+        sample_idx = self.dataset_sample_index[idx]
+        return self.datasets[dataset_idx][sample_idx]

examples/tutorial/sequence_parallel/data/datasets/builder.py

@@ -0,0 +1,152 @@
+from .blendable_dataset import BlendableDataset
+from .dataset_utils import get_datasets_weights_and_num_samples, get_indexed_dataset_, get_train_valid_test_split_
+from .bert_dataset import BertDataset
+from colossalai.logging import get_dist_logger
+
+DSET_TYPE_BERT = 'standard_bert'
+DSET_TYPE_ICT = 'ict'
+DSET_TYPE_T5 = 't5'
+
+DSET_TYPES = [DSET_TYPE_BERT, DSET_TYPE_ICT, DSET_TYPE_T5]
+
+
+def _build_train_valid_test_datasets(data_prefix, data_impl, splits_string,
+                                     train_valid_test_num_samples,
+                                     max_seq_length, masked_lm_prob,
+                                     short_seq_prob, seed, skip_warmup,
+                                     binary_head,
+                                     dataset_type='standard_bert'):
+
+    if dataset_type not in DSET_TYPES:
+        raise ValueError("Invalid dataset_type: ", dataset_type)
+
+    # Indexed dataset.
+    indexed_dataset = get_indexed_dataset_(data_prefix,
+                                           data_impl,
+                                           skip_warmup)
+
+    # Get start and end indices of train/valid/train into doc-idx
+    # Note that doc-idx is designed to be num-docs + 1 so we can
+    # easily iterate over it.
+    total_num_of_documents = indexed_dataset.doc_idx.shape[0] - 1
+    splits = get_train_valid_test_split_(splits_string, total_num_of_documents)
+
+    logger = get_dist_logger()
+
+    # Print stats about the splits.
+    logger.info('\n > dataset split:', ranks=[0])
+
+    def print_split_stats(name, index):
+        start_index = indexed_dataset.doc_idx[splits[index]]
+        end_index = indexed_dataset.doc_idx[splits[index + 1]]
+        logger.info('\n    {}:'.format(name) +
+                    '\n     document indices in [{}, {}) total of {} documents'.format(
+                        splits[index], splits[index + 1],
+                        splits[index + 1] - splits[index]) +
+                    '\n     sentence indices in [{}, {}) total of {} sentences'.format(
+                        start_index, end_index,
+                        end_index - start_index),
+                    ranks=[0])
+    print_split_stats('train', 0)
+    print_split_stats('validation', 1)
+    print_split_stats('test', 2)
+
+    def build_dataset(index, name):
+        dataset = None
+        if splits[index + 1] > splits[index]:
+            # Get the pointer to the original doc-idx so we can set it later.
+            doc_idx_ptr = indexed_dataset.get_doc_idx()
+            # Slice the doc-idx
+            start_index = splits[index]
+            # Add +1 so we can index into the dataset to get the upper bound.
+            end_index = splits[index + 1] + 1
+            # New doc_idx view.
+            indexed_dataset.set_doc_idx(doc_idx_ptr[start_index:end_index])
+            # Build the dataset accordingly.
+            kwargs = dict(
+                name=name,
+                data_prefix=data_prefix,
+                num_epochs=None,
+                max_num_samples=train_valid_test_num_samples[index],
+                max_seq_length=max_seq_length,
+                seed=seed,
+            )
+
+            if dataset_type != DSET_TYPE_BERT:
+                raise NotImplementedError("Only BERT dataset is supported")
+            else:
+                dataset = BertDataset(
+                    indexed_dataset=indexed_dataset,
+                    masked_lm_prob=masked_lm_prob,
+                    short_seq_prob=short_seq_prob,
+                    binary_head=binary_head,
+                    **kwargs
+                )
+
+            # Set the original pointer so dataset remains the main dataset.
+            indexed_dataset.set_doc_idx(doc_idx_ptr)
+            # Checks.
+            assert indexed_dataset.doc_idx[0] == 0
+            assert indexed_dataset.doc_idx.shape[0] == \
+                (total_num_of_documents + 1)
+        return dataset
+
+    train_dataset = build_dataset(0, 'train')
+    valid_dataset = build_dataset(1, 'valid')
+    test_dataset = build_dataset(2, 'test')
+
+    return (train_dataset, valid_dataset, test_dataset)
+
+
+def build_train_valid_test_datasets(data_prefix, data_impl, splits_string,
+                                    train_valid_test_num_samples,
+                                    max_seq_length, masked_lm_prob,
+                                    short_seq_prob, seed, skip_warmup,
+                                    binary_head,
+                                    dataset_type='standard_bert'):
+
+    if len(data_prefix) == 1:
+        return _build_train_valid_test_datasets(data_prefix[0],
+                                                data_impl, splits_string,
+                                                train_valid_test_num_samples,
+                                                max_seq_length, masked_lm_prob,
+                                                short_seq_prob, seed,
+                                                skip_warmup,
+                                                binary_head,
+                                                dataset_type=dataset_type)
+    # Blending dataset.
+    # Parse the values.
+    output = get_datasets_weights_and_num_samples(data_prefix,
+                                                  train_valid_test_num_samples)
+    prefixes, weights, datasets_train_valid_test_num_samples = output
+
+    # Build individual datasets.
+    train_datasets = []
+    valid_datasets = []
+    test_datasets = []
+    for i in range(len(prefixes)):
+        train_ds, valid_ds, test_ds = _build_train_valid_test_datasets(
+            prefixes[i], data_impl, splits_string,
+            datasets_train_valid_test_num_samples[i],
+            max_seq_length, masked_lm_prob, short_seq_prob,
+            seed, skip_warmup, binary_head, dataset_type=dataset_type)
+        if train_ds:
+            train_datasets.append(train_ds)
+        if valid_ds:
+            valid_datasets.append(valid_ds)
+        if test_ds:
+            test_datasets.append(test_ds)
+
+        # Blend.
+    blending_train_dataset = None
+    if train_datasets:
+        blending_train_dataset = BlendableDataset(train_datasets, weights)
+    blending_valid_dataset = None
+    if valid_datasets:
+        blending_valid_dataset = BlendableDataset(valid_datasets, weights)
+    blending_test_dataset = None
+    if test_datasets:
+        blending_test_dataset = BlendableDataset(test_datasets, weights)
+
+    return (blending_train_dataset, blending_valid_dataset,
+            blending_test_dataset)

examples/tutorial/sequence_parallel/data/datasets/data_samplers.py

@@ -0,0 +1,153 @@
+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Dataloaders."""
+
+import torch
+import random
+from colossalai.core import global_context as gpc
+from colossalai.context import ParallelMode
+
+
+def build_pretraining_data_loader(dataset, consumed_samples, micro_batch_size, dataloader_type='single', num_workers=0):
+    """Build dataloader given an input dataset."""
+
+    if dataset is None:
+        return None
+
+    # Megatron sampler
+    if dataloader_type == 'single':
+        batch_sampler = MegatronPretrainingSampler(total_samples=len(dataset),
+                                                   consumed_samples=consumed_samples,
+                                                   micro_batch_size=micro_batch_size,
+                                                   data_parallel_rank=gpc.get_local_rank(ParallelMode.DATA),
+                                                   data_parallel_size=gpc.get_world_size(ParallelMode.DATA))
+    elif dataloader_type == 'cyclic':
+        batch_sampler = MegatronPretrainingRandomSampler(total_samples=len(dataset),
+                                                         consumed_samples=consumed_samples,
+                                                         micro_batch_size=micro_batch_size,
+                                                         data_parallel_rank=gpc.get_local_rank(ParallelMode.DATA),
+                                                         data_parallel_size=gpc.get_world_size(ParallelMode.DATA))
+    else:
+        raise Exception('{} dataloader type is not supported.'.format(dataloader_type))
+
+    # Torch dataloader.
+    return torch.utils.data.DataLoader(dataset, batch_sampler=batch_sampler, num_workers=num_workers, pin_memory=True)
+
+
+class MegatronPretrainingSampler:
+
+    def __init__(self,
+                 total_samples,
+                 consumed_samples,
+                 micro_batch_size,
+                 data_parallel_rank,
+                 data_parallel_size,
+                 drop_last=True):
+        # Keep a copy of input params for later use.
+        self.total_samples = total_samples
+        self.consumed_samples = consumed_samples
+        self.micro_batch_size = micro_batch_size
+        self.data_parallel_rank = data_parallel_rank
+        self.micro_batch_times_data_parallel_size = \
+            self.micro_batch_size * data_parallel_size
+        self.drop_last = drop_last
+
+        # Sanity checks.
+        assert self.total_samples > 0, \
+            'no sample to consume: {}'.format(self.total_samples)
+        assert self.consumed_samples < self.total_samples, \
+            'no samples left to consume: {}, {}'.format(self.consumed_samples,
+                                                        self.total_samples)
+        assert self.micro_batch_size > 0
+        assert data_parallel_size > 0
+        assert self.data_parallel_rank < data_parallel_size, \
+            'data_parallel_rank should be smaller than data size: {}, ' \
+            '{}'.format(self.data_parallel_rank, data_parallel_size)
+
+    def __len__(self):
+        return self.total_samples
+
+    def get_start_end_idx(self):
+        start_idx = self.data_parallel_rank * self.micro_batch_size
+        end_idx = start_idx + self.micro_batch_size
+        return start_idx, end_idx
+
+    def __iter__(self):
+        batch = []
+        # Last batch will be dropped if drop_last is not set False
+        for idx in range(self.consumed_samples, self.total_samples):
+            batch.append(idx)
+            if len(batch) == self.micro_batch_times_data_parallel_size:
+                start_idx, end_idx = self.get_start_end_idx()
+                yield batch[start_idx:end_idx]
+                batch = []
+
+        # Check the last partial batch and see drop_last is set
+        if len(batch) > 0 and not self.drop_last:
+            start_idx, end_idx = self.get_start_end_idx()
+            yield batch[start_idx:end_idx]
+
+
+class MegatronPretrainingRandomSampler:
+
+    def __init__(self, total_samples, consumed_samples, micro_batch_size, data_parallel_rank, data_parallel_size):
+        # Keep a copy of input params for later use.
+        self.total_samples = total_samples
+        self.consumed_samples = consumed_samples
+        self.micro_batch_size = micro_batch_size
+        self.data_parallel_rank = data_parallel_rank
+        self.data_parallel_size = data_parallel_size
+        self.micro_batch_times_data_parallel_size = \
+            self.micro_batch_size * data_parallel_size
+        self.last_batch_size = \
+            self.total_samples % self.micro_batch_times_data_parallel_size
+
+        # Sanity checks.
+        assert self.total_samples > 0, \
+            'no sample to consume: {}'.format(self.total_samples)
+        assert self.micro_batch_size > 0
+        assert data_parallel_size > 0
+        assert self.data_parallel_rank < data_parallel_size, \
+            'data_parallel_rank should be smaller than data size: {}, ' \
+            '{}'.format(self.data_parallel_rank, data_parallel_size)
+
+    def __len__(self):
+        return self.total_samples
+
+    def __iter__(self):
+        active_total_samples = self.total_samples - self.last_batch_size
+        self.epoch = self.consumed_samples // active_total_samples
+        current_epoch_samples = self.consumed_samples % active_total_samples
+        assert current_epoch_samples % self.micro_batch_times_data_parallel_size == 0
+
+        # data sharding and random sampling
+        bucket_size = (self.total_samples // self.micro_batch_times_data_parallel_size) \
+            * self.micro_batch_size
+        bucket_offset = current_epoch_samples // self.data_parallel_size
+        start_idx = self.data_parallel_rank * bucket_size
+
+        g = torch.Generator()
+        g.manual_seed(self.epoch)
+        random_idx = torch.randperm(bucket_size, generator=g).tolist()
+        idx_range = [start_idx + x for x in random_idx[bucket_offset:]]
+
+        batch = []
+        # Last batch if not complete will be dropped.
+        for idx in idx_range:
+            batch.append(idx)
+            if len(batch) == self.micro_batch_size:
+                self.consumed_samples += self.micro_batch_times_data_parallel_size
+                yield batch
+                batch = []

examples/tutorial/sequence_parallel/data/datasets/dataset_utils.py

@@ -0,0 +1,592 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors, and NVIDIA.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+# Most of the code here has been copied from:
+#   https://github.com/google-research/albert/blob/master/create_pretraining_data.py
+# with some modifications.
+
+import math
+import time
+import collections
+from colossalai.logging import get_dist_logger
+import numpy as np
+from .blendable_dataset import BlendableDataset
+from .indexed_dataset import make_dataset as make_indexed_dataset
+
+DSET_TYPE_STD = 'standard_bert'
+DSET_TYPE_ICT = 'ict'
+
+DSET_TYPES = [DSET_TYPE_ICT, DSET_TYPE_STD]
+
+
+def get_datasets_weights_and_num_samples(data_prefix,
+                                         train_valid_test_num_samples):
+
+    # The data prefix should be in the format of:
+    #   weight-1, data-prefix-1, weight-2, data-prefix-2, ..
+    assert len(data_prefix) % 2 == 0
+    num_datasets = len(data_prefix) // 2
+    weights = [0]*num_datasets
+    prefixes = [0]*num_datasets
+    for i in range(num_datasets):
+        weights[i] = float(data_prefix[2*i])
+        prefixes[i] = (data_prefix[2*i+1]).strip()
+    # Normalize weights
+    weight_sum = 0.0
+    for weight in weights:
+        weight_sum += weight
+    assert weight_sum > 0.0
+    weights = [weight / weight_sum for weight in weights]
+
+    # Add 0.5% (the 1.005 factor) so in case the bleding dataset does
+    # not uniformly distribute the number of samples, we still have
+    # samples left to feed to the network.
+    datasets_train_valid_test_num_samples = []
+    for weight in weights:
+        datasets_train_valid_test_num_samples.append(
+            [int(math.ceil(val * weight * 1.005))
+             for val in train_valid_test_num_samples])
+
+    return prefixes, weights, datasets_train_valid_test_num_samples
+
+
+def compile_helper():
+    """Compile helper function ar runtime. Make sure this
+    is invoked on a single process."""
+    import os
+    import subprocess
+    path = os.path.abspath(os.path.dirname(__file__))
+    ret = subprocess.run(['make', '-C', path])
+    if ret.returncode != 0:
+        print("Making C++ dataset helpers module failed, exiting.")
+        import sys
+        sys.exit(1)
+
+
+def get_a_and_b_segments(sample, np_rng):
+    """Divide sample into a and b segments."""
+
+    # Number of sentences in the sample.
+    n_sentences = len(sample)
+    # Make sure we always have two sentences.
+    assert n_sentences > 1, 'make sure each sample has at least two sentences.'
+
+    # First part:
+    # `a_end` is how many sentences go into the `A`.
+    a_end = 1
+    if n_sentences >= 3:
+        # Note that randin in numpy is exclusive.
+        a_end = np_rng.randint(1, n_sentences)
+    tokens_a = []
+    for j in range(a_end):
+        tokens_a.extend(sample[j])
+
+    # Second part:
+    tokens_b = []
+    for j in range(a_end, n_sentences):
+        tokens_b.extend(sample[j])
+
+    # Random next:
+    is_next_random = False
+    if np_rng.random() < 0.5:
+        is_next_random = True
+        tokens_a, tokens_b = tokens_b, tokens_a
+
+    return tokens_a, tokens_b, is_next_random
+
+
+def truncate_segments(tokens_a, tokens_b, len_a, len_b, max_num_tokens, np_rng):
+    """Truncates a pair of sequences to a maximum sequence length."""
+    #print(len_a, len_b, max_num_tokens)
+    assert len_a > 0
+    if len_a + len_b <= max_num_tokens:
+        return False
+    while len_a + len_b > max_num_tokens:
+        if len_a > len_b:
+            len_a -= 1
+            tokens = tokens_a
+        else:
+            len_b -= 1
+            tokens = tokens_b
+        if np_rng.random() < 0.5:
+            del tokens[0]
+        else:
+            tokens.pop()
+    return True
+
+
+def create_tokens_and_tokentypes(tokens_a, tokens_b, cls_id, sep_id):
+    """Merge segments A and B, add [CLS] and [SEP] and build tokentypes."""
+
+    tokens = []
+    tokentypes = []
+    # [CLS].
+    tokens.append(cls_id)
+    tokentypes.append(0)
+    # Segment A.
+    for token in tokens_a:
+        tokens.append(token)
+        tokentypes.append(0)
+    # [SEP].
+    tokens.append(sep_id)
+    tokentypes.append(0)
+    # Segment B.
+    for token in tokens_b:
+        tokens.append(token)
+        tokentypes.append(1)
+    if tokens_b:
+        # [SEP].
+        tokens.append(sep_id)
+        tokentypes.append(1)
+
+    return tokens, tokentypes
+
+
+MaskedLmInstance = collections.namedtuple("MaskedLmInstance",
+                                          ["index", "label"])
+
+
+def is_start_piece(piece):
+    """Check if the current word piece is the starting piece (BERT)."""
+    # When a word has been split into
+    # WordPieces, the first token does not have any marker and any subsequence
+    # tokens are prefixed with ##. So whenever we see the ## token, we
+    # append it to the previous set of word indexes.
+    return not piece.startswith("##")
+
+
+def create_masked_lm_predictions(tokens,
+                                 vocab_id_list, vocab_id_to_token_dict,
+                                 masked_lm_prob,
+                                 cls_id, sep_id, mask_id,
+                                 max_predictions_per_seq,
+                                 np_rng,
+                                 max_ngrams=3,
+                                 do_whole_word_mask=True,
+                                 favor_longer_ngram=False,
+                                 do_permutation=False):
+    """Creates the predictions for the masked LM objective.
+    Note: Tokens here are vocab ids and not text tokens."""
+
+    cand_indexes = []
+    # Note(mingdachen): We create a list for recording if the piece is
+    # the starting piece of current token, where 1 means true, so that
+    # on-the-fly whole word masking is possible.
+    token_boundary = [0] * len(tokens)
+
+    for (i, token) in enumerate(tokens):
+        if token == cls_id or token == sep_id:
+            token_boundary[i] = 1
+            continue
+        # Whole Word Masking means that if we mask all of the wordpieces
+        # corresponding to an original word.
+        #
+        # Note that Whole Word Masking does *not* change the training code
+        # at all -- we still predict each WordPiece independently, softmaxed
+        # over the entire vocabulary.
+        if (do_whole_word_mask and len(cand_indexes) >= 1 and
+                not is_start_piece(vocab_id_to_token_dict[token])):
+            cand_indexes[-1].append(i)
+        else:
+            cand_indexes.append([i])
+            if is_start_piece(vocab_id_to_token_dict[token]):
+                token_boundary[i] = 1
+
+    output_tokens = list(tokens)
+
+    masked_lm_positions = []
+    masked_lm_labels = []
+
+    if masked_lm_prob == 0:
+        return (output_tokens, masked_lm_positions,
+                masked_lm_labels, token_boundary)
+
+    num_to_predict = min(max_predictions_per_seq,
+                         max(1, int(round(len(tokens) * masked_lm_prob))))
+
+    # Note(mingdachen):
+    # By default, we set the probabilities to favor shorter ngram sequences.
+    ngrams = np.arange(1, max_ngrams + 1, dtype=np.int64)
+    pvals = 1. / np.arange(1, max_ngrams + 1)
+    pvals /= pvals.sum(keepdims=True)
+
+    if favor_longer_ngram:
+        pvals = pvals[::-1]
+
+    ngram_indexes = []
+    for idx in range(len(cand_indexes)):
+        ngram_index = []
+        for n in ngrams:
+            ngram_index.append(cand_indexes[idx:idx + n])
+        ngram_indexes.append(ngram_index)
+
+    np_rng.shuffle(ngram_indexes)
+
+    masked_lms = []
+    covered_indexes = set()
+    for cand_index_set in ngram_indexes:
+        if len(masked_lms) >= num_to_predict:
+            break
+        if not cand_index_set:
+            continue
+        # Note(mingdachen):
+        # Skip current piece if they are covered in lm masking or previous ngrams.
+        for index_set in cand_index_set[0]:
+            for index in index_set:
+                if index in covered_indexes:
+                    continue
+
+        n = np_rng.choice(ngrams[:len(cand_index_set)],
+                          p=pvals[:len(cand_index_set)] /
+                          pvals[:len(cand_index_set)].sum(keepdims=True))
+        index_set = sum(cand_index_set[n - 1], [])
+        n -= 1
+        # Note(mingdachen):
+        # Repeatedly looking for a candidate that does not exceed the
+        # maximum number of predictions by trying shorter ngrams.
+        while len(masked_lms) + len(index_set) > num_to_predict:
+            if n == 0:
+                break
+            index_set = sum(cand_index_set[n - 1], [])
+            n -= 1
+        # If adding a whole-word mask would exceed the maximum number of
+        # predictions, then just skip this candidate.
+        if len(masked_lms) + len(index_set) > num_to_predict:
+            continue
+        is_any_index_covered = False
+        for index in index_set:
+            if index in covered_indexes:
+                is_any_index_covered = True
+                break
+        if is_any_index_covered:
+            continue
+        for index in index_set:
+            covered_indexes.add(index)
+
+            masked_token = None
+            # 80% of the time, replace with [MASK]
+            if np_rng.random() < 0.8:
+                masked_token = mask_id
+            else:
+                # 10% of the time, keep original
+                if np_rng.random() < 0.5:
+                    masked_token = tokens[index]
+                # 10% of the time, replace with random word
+                else:
+                    masked_token = vocab_id_list[np_rng.randint(0, len(vocab_id_list))]
+
+            output_tokens[index] = masked_token
+
+            masked_lms.append(MaskedLmInstance(index=index, label=tokens[index]))
+    assert len(masked_lms) <= num_to_predict
+
+    np_rng.shuffle(ngram_indexes)
+
+    select_indexes = set()
+    if do_permutation:
+        for cand_index_set in ngram_indexes:
+            if len(select_indexes) >= num_to_predict:
+                break
+            if not cand_index_set:
+                continue
+            # Note(mingdachen):
+            # Skip current piece if they are covered in lm masking or previous ngrams.
+            for index_set in cand_index_set[0]:
+                for index in index_set:
+                    if index in covered_indexes or index in select_indexes:
+                        continue
+
+            n = np.random.choice(ngrams[:len(cand_index_set)],
+                                 p=pvals[:len(cand_index_set)] /
+                                 pvals[:len(cand_index_set)].sum(keepdims=True))
+            index_set = sum(cand_index_set[n - 1], [])
+            n -= 1
+
+            while len(select_indexes) + len(index_set) > num_to_predict:
+                if n == 0:
+                    break
+                index_set = sum(cand_index_set[n - 1], [])
+                n -= 1
+            # If adding a whole-word mask would exceed the maximum number of
+            # predictions, then just skip this candidate.
+            if len(select_indexes) + len(index_set) > num_to_predict:
+                continue
+            is_any_index_covered = False
+            for index in index_set:
+                if index in covered_indexes or index in select_indexes:
+                    is_any_index_covered = True
+                    break
+            if is_any_index_covered:
+                continue
+            for index in index_set:
+                select_indexes.add(index)
+        assert len(select_indexes) <= num_to_predict
+
+        select_indexes = sorted(select_indexes)
+        permute_indexes = list(select_indexes)
+        np_rng.shuffle(permute_indexes)
+        orig_token = list(output_tokens)
+
+        for src_i, tgt_i in zip(select_indexes, permute_indexes):
+            output_tokens[src_i] = orig_token[tgt_i]
+            masked_lms.append(MaskedLmInstance(index=src_i, label=orig_token[src_i]))
+
+    masked_lms = sorted(masked_lms, key=lambda x: x.index)
+
+    for p in masked_lms:
+        masked_lm_positions.append(p.index)
+        masked_lm_labels.append(p.label)
+
+    return (output_tokens, masked_lm_positions, masked_lm_labels, token_boundary)
+
+
+def pad_and_convert_to_numpy(tokens, tokentypes, masked_positions,
+                             masked_labels, pad_id, max_seq_length):
+    """Pad sequences and convert them to numpy."""
+
+    # Some checks.
+    num_tokens = len(tokens)
+    padding_length = max_seq_length - num_tokens
+    assert padding_length >= 0
+    assert len(tokentypes) == num_tokens
+    assert len(masked_positions) == len(masked_labels)
+
+    # Tokens and token types.
+    filler = [pad_id] * padding_length
+    tokens_np = np.array(tokens + filler, dtype=np.int64)
+    tokentypes_np = np.array(tokentypes + filler, dtype=np.int64)
+
+    # Padding mask.
+    padding_mask_np = np.array([1] * num_tokens + [0] * padding_length,
+                               dtype=np.int64)
+
+    # Lables and loss mask.
+    labels = [-1] * max_seq_length
+    loss_mask = [0] * max_seq_length
+    for i in range(len(masked_positions)):
+        assert masked_positions[i] < num_tokens
+        labels[masked_positions[i]] = masked_labels[i]
+        loss_mask[masked_positions[i]] = 1
+    labels_np = np.array(labels, dtype=np.int64)
+    loss_mask_np = np.array(loss_mask, dtype=np.int64)
+
+    return tokens_np, tokentypes_np, labels_np, padding_mask_np, loss_mask_np
+
+
+def build_train_valid_test_datasets(data_prefix, data_impl, splits_string,
+                                    train_valid_test_num_samples,
+                                    max_seq_length, masked_lm_prob,
+                                    short_seq_prob, seed, skip_warmup,
+                                    binary_head,
+                                    dataset_type='standard_bert'):
+
+    if len(data_prefix) == 1:
+        return _build_train_valid_test_datasets(data_prefix[0],
+                                                data_impl, splits_string,
+                                                train_valid_test_num_samples,
+                                                max_seq_length, masked_lm_prob,
+                                                short_seq_prob, seed,
+                                                skip_warmup,
+                                                binary_head,
+                                                dataset_type=dataset_type)
+    # Blending dataset.
+    # Parse the values.
+    output = get_datasets_weights_and_num_samples(data_prefix,
+                                                  train_valid_test_num_samples)
+    prefixes, weights, datasets_train_valid_test_num_samples = output
+
+    # Build individual datasets.
+    train_datasets = []
+    valid_datasets = []
+    test_datasets = []
+    for i in range(len(prefixes)):
+        train_ds, valid_ds, test_ds = _build_train_valid_test_datasets(
+            prefixes[i], data_impl, splits_string,
+            datasets_train_valid_test_num_samples[i],
+            max_seq_length, masked_lm_prob, short_seq_prob,
+            seed, skip_warmup, binary_head, dataset_type=dataset_type)
+        if train_ds:
+            train_datasets.append(train_ds)
+        if valid_ds:
+            valid_datasets.append(valid_ds)
+        if test_ds:
+            test_datasets.append(test_ds)
+
+        # Blend.
+    blending_train_dataset = None
+    if train_datasets:
+        blending_train_dataset = BlendableDataset(train_datasets, weights)
+    blending_valid_dataset = None
+    if valid_datasets:
+        blending_valid_dataset = BlendableDataset(valid_datasets, weights)
+    blending_test_dataset = None
+    if test_datasets:
+        blending_test_dataset = BlendableDataset(test_datasets, weights)
+
+    return (blending_train_dataset, blending_valid_dataset,
+            blending_test_dataset)
+
+
+def _build_train_valid_test_datasets(data_prefix, data_impl, splits_string,
+                                     train_valid_test_num_samples,
+                                     max_seq_length, masked_lm_prob,
+                                     short_seq_prob, seed, skip_warmup,
+                                     binary_head,
+                                     dataset_type='standard_bert'):
+    logger = get_dist_logger()
+
+    if dataset_type not in DSET_TYPES:
+        raise ValueError("Invalid dataset_type: ", dataset_type)
+
+    # Indexed dataset.
+    indexed_dataset = get_indexed_dataset_(data_prefix,
+                                           data_impl,
+                                           skip_warmup)
+
+    if dataset_type == DSET_TYPE_ICT:
+        args = get_args()
+        title_dataset = get_indexed_dataset_(args.titles_data_path,
+                                             data_impl,
+                                             skip_warmup)
+
+    # Get start and end indices of train/valid/train into doc-idx
+    # Note that doc-idx is designed to be num-docs + 1 so we can
+    # easily iterate over it.
+    total_num_of_documents = indexed_dataset.doc_idx.shape[0] - 1
+    splits = get_train_valid_test_split_(splits_string, total_num_of_documents)
+
+    # Print stats about the splits.
+    logger.info('\n > dataset split:')
+
+    def print_split_stats(name, index):
+        start_index = indexed_dataset.doc_idx[splits[index]]
+        end_index = indexed_dataset.doc_idx[splits[index + 1]]
+        logger.info('\n    {}:'.format(name) +
+                    '\n     document indices in [{}, {}) total of {} documents'.format(
+                        splits[index],
+                        splits[index + 1],
+                        splits[index + 1] - splits[index]) +
+                    '\n     sentence indices in [{}, {}) total of {} sentences'.format(
+                        start_index,
+                        end_index,
+                        end_index - start_index),
+                    ranks=[0])
+    print_split_stats('train', 0)
+    print_split_stats('validation', 1)
+    print_split_stats('test', 2)
+
+    def build_dataset(index, name):
+        from .bert_dataset import BertDataset
+        dataset = None
+        if splits[index + 1] > splits[index]:
+            # Get the pointer to the original doc-idx so we can set it later.
+            doc_idx_ptr = indexed_dataset.get_doc_idx()
+            # Slice the doc-idx
+            start_index = splits[index]
+            # Add +1 so we can index into the dataset to get the upper bound.
+            end_index = splits[index + 1] + 1
+            # New doc_idx view.
+            indexed_dataset.set_doc_idx(doc_idx_ptr[start_index:end_index])
+            # Build the dataset accordingly.
+            kwargs = dict(
+                name=name,
+                data_prefix=data_prefix,
+                num_epochs=None,
+                max_num_samples=train_valid_test_num_samples[index],
+                max_seq_length=max_seq_length,
+                seed=seed,
+                binary_head=binary_head
+            )
+
+            if dataset_type == DSET_TYPE_ICT:
+                args = get_args()
+                dataset = ICTDataset(
+                    block_dataset=indexed_dataset,
+                    title_dataset=title_dataset,
+                    query_in_block_prob=args.query_in_block_prob,
+                    use_one_sent_docs=args.use_one_sent_docs,
+                    **kwargs
+                )
+            else:
+                dataset = BertDataset(
+                    indexed_dataset=indexed_dataset,
+                    masked_lm_prob=masked_lm_prob,
+                    short_seq_prob=short_seq_prob,
+                    **kwargs
+                )
+
+            # Set the original pointer so dataset remains the main dataset.
+            indexed_dataset.set_doc_idx(doc_idx_ptr)
+            # Checks.
+            assert indexed_dataset.doc_idx[0] == 0
+            assert indexed_dataset.doc_idx.shape[0] == \
+                (total_num_of_documents + 1)
+        return dataset
+
+    train_dataset = build_dataset(0, 'train')
+    valid_dataset = build_dataset(1, 'valid')
+    test_dataset = build_dataset(2, 'test')
+
+    return (train_dataset, valid_dataset, test_dataset)
+
+
+def get_indexed_dataset_(data_prefix, data_impl, skip_warmup):
+    logger = get_dist_logger()
+    start_time = time.time()
+    indexed_dataset = make_indexed_dataset(data_prefix,
+                                           data_impl,
+                                           skip_warmup)
+    assert indexed_dataset.sizes.shape[0] == indexed_dataset.doc_idx[-1]
+    logger.info('\n > building dataset index ...', ranks=[0])
+    logger.info('\n > finished creating indexed dataset in {:4f} '
+                'seconds'.format(time.time() - start_time), ranks=[0])
+    logger.info('\n > indexed dataset stats:' +
+                '\n    number of documents: {}'.format(
+                    indexed_dataset.doc_idx.shape[0] - 1) +
+                '\n    number of sentences: {}'.format(
+                    indexed_dataset.sizes.shape[0]),
+                ranks=[0]
+                )
+
+    return indexed_dataset
+
+
+def get_train_valid_test_split_(splits_string, size):
+    """ Get dataset splits from comma or '/' separated string list."""
+
+    splits = []
+    if splits_string.find(',') != -1:
+        splits = [float(s) for s in splits_string.split(',')]
+    elif splits_string.find('/') != -1:
+        splits = [float(s) for s in splits_string.split('/')]
+    else:
+        splits = [float(splits_string)]
+    while len(splits) < 3:
+        splits.append(0.)
+    splits = splits[:3]
+    splits_sum = sum(splits)
+    assert splits_sum > 0.0
+    splits = [split / splits_sum for split in splits]
+    splits_index = [0]
+    for index, split in enumerate(splits):
+        splits_index.append(splits_index[index] +
+                            int(round(split * float(size))))
+    diff = splits_index[-1] - size
+    for index in range(1, len(splits_index)):
+        splits_index[index] -= diff
+    assert len(splits_index) == 4
+    assert splits_index[-1] == size
+    return splits_index

examples/tutorial/sequence_parallel/data/datasets/helpers.cpp

@@ -0,0 +1,717 @@
+/*
+ coding=utf-8
+ Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+
+     http://www.apache.org/licenses/LICENSE-2.0
+
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License.
+ */
+
+
+/* Helper methods for fast index mapping builds */
+
+#include <algorithm>
+#include <iostream>
+#include <limits>
+#include <math.h>
+#include <stdexcept>
+#include <pybind11/pybind11.h>
+#include <pybind11/numpy.h>
+#include <random>
+
+namespace py = pybind11;
+using namespace std;
+
+const int32_t LONG_SENTENCE_LEN = 512;
+
+
+void build_blending_indices(py::array_t<uint8_t>& dataset_index,
+			    py::array_t<int64_t>& dataset_sample_index,
+			    const py::array_t<double>& weights,
+			    const int32_t num_datasets,
+			    const int64_t size, const bool verbose) {
+  /* Given multiple datasets and a weighting array, build samples
+   such that it follows those wieghts.*/
+
+  if (verbose) {
+    std::cout << "> building indices for blendable datasets ..." << std::endl;
+  }
+
+  // Get the pointer access without the checks.
+  auto dataset_index_ptr = dataset_index.mutable_unchecked<1>();
+  auto dataset_sample_index_ptr = dataset_sample_index.mutable_unchecked<1>();
+  auto weights_ptr = weights.unchecked<1>();
+
+  // Initialize buffer for number of samples used for each dataset.
+  int64_t current_samples[num_datasets];
+  for(int64_t i = 0; i < num_datasets; ++i) {
+    current_samples[i] = 0;
+  }
+
+  // For each sample:
+  for(int64_t sample_idx = 0; sample_idx < size; ++sample_idx) {
+
+    // Determine where the max error in sampling is happening.
+    auto sample_idx_double = std::max(static_cast<double>(sample_idx), 1.0);
+    int64_t max_error_index = 0;
+    double max_error = weights_ptr[0] * sample_idx_double -
+      static_cast<double>(current_samples[0]);
+    for (int64_t dataset_idx = 1; dataset_idx < num_datasets; ++dataset_idx) {
+      double error = weights_ptr[dataset_idx] * sample_idx_double -
+	static_cast<double>(current_samples[dataset_idx]);
+      if (error > max_error) {
+	max_error = error;
+	max_error_index = dataset_idx;
+      }
+    }
+
+    // Populate the indices.
+    dataset_index_ptr[sample_idx] = static_cast<uint8_t>(max_error_index);
+    dataset_sample_index_ptr[sample_idx] = current_samples[max_error_index];
+
+    // Update the total samples.
+    current_samples[max_error_index] += 1;
+    
+  }
+
+  // print info
+  if (verbose) {
+    std::cout << " > sample ratios:" << std::endl;
+    for (int64_t dataset_idx = 0; dataset_idx < num_datasets; ++dataset_idx) {
+      auto ratio = static_cast<double>(current_samples[dataset_idx]) /
+	static_cast<double>(size);
+      std::cout << "   dataset " << dataset_idx << ", input: " <<
+	weights_ptr[dataset_idx] << ", achieved: " << ratio << std::endl; 
+    }
+  }
+
+}
+
+
+py::array build_sample_idx(const py::array_t<int32_t>& sizes_,
+			   const py::array_t<int32_t>& doc_idx_,
+			   const int32_t seq_length,
+			   const int32_t num_epochs,
+			   const int64_t tokens_per_epoch) {
+    /* Sample index (sample_idx) is used for gpt2 like dataset for which
+       the documents are flattened and the samples are built based on this
+       1-D flatten array. It is a 2D array with sizes [number-of-samples + 1, 2]
+       where [..., 0] contains the index into `doc_idx` and [..., 1] is the
+       starting offset in that document.*/
+
+    // Consistency checks.
+    assert(seq_length > 1);
+    assert(num_epochs > 0);
+    assert(tokens_per_epoch > 1);
+
+    // Remove bound checks.
+    auto sizes = sizes_.unchecked<1>();
+    auto doc_idx = doc_idx_.unchecked<1>();
+
+    // Mapping and it's length (1D).
+    int64_t num_samples = (num_epochs * tokens_per_epoch - 1) / seq_length;
+    int32_t* sample_idx = new int32_t[2*(num_samples+1)];
+
+    cout << "    using:" << endl << std::flush;
+    cout << "     number of documents:       " <<
+      doc_idx_.shape(0) / num_epochs << endl << std::flush;
+    cout << "     number of epochs:          " << num_epochs <<
+      endl << std::flush;
+    cout << "     sequence length:           " << seq_length <<
+      endl << std::flush;
+    cout << "     total number of samples:   " << num_samples <<
+      endl << std::flush;
+
+    // Index into sample_idx.
+    int64_t sample_index = 0;
+    // Index into doc_idx.
+    int64_t doc_idx_index = 0;
+    // Begining offset for each document.
+    int32_t doc_offset = 0;
+    // Start with first document and no offset.
+    sample_idx[2 * sample_index] = doc_idx_index;
+    sample_idx[2 * sample_index + 1] = doc_offset;
+    ++sample_index;
+
+    while (sample_index <= num_samples) {
+        // Start with a fresh sequence.
+      int32_t remaining_seq_length = seq_length + 1;
+      while (remaining_seq_length != 0) {
+            // Get the document length.
+	auto doc_id = doc_idx[doc_idx_index];
+	auto doc_length = sizes[doc_id] - doc_offset;
+	// And add it to the current sequence.
+	remaining_seq_length -= doc_length;
+	// If we have more than a full sequence, adjust offset and set
+	// remaining length to zero so we return from the while loop.
+	// Note that -1 here is for the same reason we have -1 in
+	// `_num_epochs` calculations.
+	if (remaining_seq_length <= 0) {
+	  doc_offset += (remaining_seq_length + doc_length - 1);
+	  remaining_seq_length = 0;
+	} else {
+	  // Otherwise, start from the begining of the next document.
+	  ++doc_idx_index;
+	  doc_offset = 0;
+	}
+      }
+      // Record the sequence.
+      sample_idx[2 * sample_index] = doc_idx_index;
+      sample_idx[2 * sample_index + 1] = doc_offset;
+      ++sample_index;
+    }
+
+    // Method to deallocate memory.
+    py::capsule free_when_done(sample_idx, [](void *mem_) {
+	int32_t *mem = reinterpret_cast<int32_t*>(mem_);
+	delete[] mem;
+      });
+
+    // Return the numpy array.
+    const auto byte_size = sizeof(int32_t);
+    return py::array(std::vector<int64_t>{num_samples+1, 2}, // shape
+                     {2*byte_size, byte_size}, // C-style contiguous strides
+                     sample_idx, // the data pointer
+                     free_when_done); // numpy array references
+    
+}
+
+
+inline int32_t get_target_sample_len(const int32_t short_seq_ratio,
+				     const int32_t max_length,
+				     std::mt19937& rand32_gen) {
+    /* Training sample length. */
+    if (short_seq_ratio == 0) {
+      return max_length;
+    }
+    const auto random_number = rand32_gen();
+    if ((random_number % short_seq_ratio) == 0) {
+      return 2 + random_number % (max_length - 1);
+    }
+    return max_length;
+}
+
+
+template<typename DocIdx>
+py::array build_mapping_impl(const py::array_t<int64_t>& docs_,
+                             const py::array_t<int32_t>& sizes_,
+                             const int32_t num_epochs,
+                             const uint64_t max_num_samples,
+                             const int32_t max_seq_length,
+                             const double short_seq_prob,
+                             const int32_t seed,
+			     const bool verbose,
+			     const int32_t min_num_sent) {
+    /* Build a mapping of (start-index, end-index, sequence-length) where
+       start and end index are the indices of the sentences in the sample
+       and sequence-length is the target sequence length.
+    */
+
+    // Consistency checks.
+    assert(num_epochs > 0);
+    assert(max_seq_length > 1);
+    assert(short_seq_prob >= 0.0);
+    assert(short_seq_prob <= 1.0);
+    assert(seed > 0);
+
+    // Remove bound checks.
+    auto docs = docs_.unchecked<1>();
+    auto sizes = sizes_.unchecked<1>();
+
+    // For efficiency, convert probability to ratio. Note: rand() generates int.
+    int32_t short_seq_ratio = 0;
+    if (short_seq_prob > 0) {
+      short_seq_ratio = static_cast<int32_t>(round(1.0 / short_seq_prob));
+    }
+
+    if (verbose) {
+        const auto sent_start_index = docs[0];
+	const auto sent_end_index = docs[docs_.shape(0) - 1];
+	const auto num_sentences = sent_end_index - sent_start_index;
+	cout << "    using:" << endl << std::flush;
+	cout << "     number of documents:            " << docs_.shape(0) - 1 <<
+	  endl << std::flush;
+	cout << "     sentences range:                [" << sent_start_index <<
+	", " << sent_end_index << ")" << endl << std::flush;
+	cout << "     total number of sentences:      " << num_sentences <<
+	  endl << std::flush;
+	cout << "     number of epochs:               " << num_epochs <<
+	  endl << std::flush;
+	cout << "     maximum number of samples:      " << max_num_samples <<
+	  endl << std::flush;
+	cout << "     maximum sequence length:        " << max_seq_length <<
+	  endl << std::flush;
+	cout << "     short sequence probability:     " << short_seq_prob <<
+	endl << std::flush;
+	cout << "     short sequence ration (1/prob): " << short_seq_ratio <<
+	  endl << std::flush;
+	cout << "     seed:                           " << seed << endl <<
+	  std::flush;
+    }
+
+    // Mapping and it's length (1D).
+    int64_t num_samples = -1;
+    DocIdx* maps = NULL;
+
+    // Perform two iterations, in the first iteration get the size
+    // and allocate memory and in the second iteration populate the map.
+    bool second = false;
+    for (int32_t iteration=0; iteration<2; ++iteration) {
+
+        // Set the seed so both iterations produce the same results.
+        std::mt19937 rand32_gen(seed);
+
+        // Set the flag on second iteration.
+        second = (iteration == 1);
+
+        // Counters:
+        uint64_t empty_docs = 0;
+        uint64_t one_sent_docs = 0;
+	uint64_t long_sent_docs = 0;
+
+        // Current map index.
+        uint64_t map_index = 0;
+
+        // For each epoch:
+        for (int32_t epoch=0; epoch<num_epochs; ++epoch) {
+            if (map_index >= max_num_samples) {
+	        if (verbose && (!second)) {
+		  cout << "    reached " << max_num_samples << " samples after "
+		       << epoch << " epochs ..." << endl << std::flush;
+		}
+                break;
+            }
+            // For each document:
+            for (int32_t doc=0; doc<(docs.shape(0) - 1); ++doc) {
+
+                // Document sentences are in [sent_index_first, sent_index_last)
+                const auto sent_index_first = docs[doc];
+                const auto sent_index_last = docs[doc + 1];
+
+                // At the begining of the document previous index is the
+		// start index.
+                auto prev_start_index = sent_index_first;
+
+                // Remaining documents.
+                auto num_remain_sent = sent_index_last - sent_index_first;
+
+                // Some bookkeeping
+                if ((epoch == 0) && (!second)) {
+                    if (num_remain_sent == 0) {
+		        ++empty_docs;
+                    }
+                    if (num_remain_sent == 1) {
+		        ++one_sent_docs;
+                    }
+                }
+
+		// Detect documents with long sentences.
+		bool contains_long_sentence = false;
+		if (num_remain_sent > 1) {
+		    for (auto sent_index=sent_index_first;
+			 sent_index < sent_index_last; ++sent_index) {
+		        if (sizes[sent_index] > LONG_SENTENCE_LEN){
+			    if ((epoch == 0) && (!second)) {
+			        ++long_sent_docs;
+			    }
+			    contains_long_sentence = true;
+			    break;
+			}
+		    }
+		}
+
+                // If we have more than two sentences.
+                if ((num_remain_sent >= min_num_sent) && (!contains_long_sentence)) {
+
+                    // Set values.
+                    auto seq_len = int32_t{0};
+                    auto num_sent = int32_t{0};
+                    auto target_seq_len = get_target_sample_len(short_seq_ratio,
+								max_seq_length,
+								rand32_gen);
+
+                    // Loop through sentences.
+                    for (auto sent_index=sent_index_first;
+                         sent_index < sent_index_last; ++sent_index) {
+
+		        // Add the size and number of sentences.
+		        seq_len += sizes[sent_index];
+		        ++num_sent;
+			--num_remain_sent;
+
+			// If we have reached the target length.
+			// and if not only one sentence is left in the document.
+			// and if we have at least two sentneces.
+			// and if we have reached end of the document.
+			if (((seq_len >= target_seq_len) &&
+			     (num_remain_sent > 1) &&
+			     (num_sent >= min_num_sent) ) || (num_remain_sent == 0)) {
+
+			    // Check for overflow.
+			    if ((3 * map_index + 2) >
+				std::numeric_limits<int64_t>::max()) {
+			        cout << "number of samples exceeded maximum "
+				     << "allowed by type int64: "
+				     << std::numeric_limits<int64_t>::max()
+				     << endl;
+				throw std::overflow_error("Number of samples");
+			    }
+
+			    // Populate the map.
+			    if (second) {
+			        const auto map_index_0 = 3 * map_index;
+				maps[map_index_0] = static_cast<DocIdx>(prev_start_index);
+				maps[map_index_0 + 1] = static_cast<DocIdx>(sent_index + 1);
+				maps[map_index_0 + 2] = static_cast<DocIdx>(target_seq_len);
+			    }
+
+			    // Update indices / counters.
+			    ++map_index;
+			    prev_start_index = sent_index + 1;
+			    target_seq_len = get_target_sample_len(short_seq_ratio,
+								   max_seq_length,
+								   rand32_gen);
+			    seq_len = 0;
+			    num_sent = 0;
+			}
+
+                    } // for (auto sent_index=sent_index_first; ...
+                } // if (num_remain_sent > 1) {
+            } // for (int doc=0; doc < num_docs; ++doc) {
+        } // for (int epoch=0; epoch < num_epochs; ++epoch) {
+
+        if (!second) {
+	    if (verbose) {
+	        cout << "   number of empty documents: " << empty_docs <<
+		  endl << std::flush;
+		cout << "   number of documents with one sentence: " <<
+		  one_sent_docs << endl << std::flush;
+		cout << "   number of documents with long sentences: " <<
+		  long_sent_docs << endl << std::flush;
+		cout << "   will create mapping for " << map_index <<
+		  " samples" << endl << std::flush;
+	    }
+	    assert(maps == NULL);
+	    assert(num_samples < 0);
+            maps = new DocIdx[3*map_index];
+            num_samples = static_cast<int64_t>(map_index);
+        }
+
+    } // for (int iteration=0; iteration < 2; ++iteration) {
+
+    // Shuffle.
+    // We need a 64 bit random number generator as we might have more
+    // than 2 billion samples.
+    std::mt19937_64 rand64_gen(seed + 1);
+    for (auto i=(num_samples - 1); i > 0; --i) {
+      const auto j = static_cast<int64_t>(rand64_gen() % (i + 1));
+      const auto i0 = 3 * i;
+      const auto j0 = 3 * j;
+      // Swap values.
+      swap(maps[i0], maps[j0]);
+      swap(maps[i0 + 1], maps[j0 + 1]);
+      swap(maps[i0 + 2], maps[j0 + 2]);
+    }
+
+    // Method to deallocate memory.
+    py::capsule free_when_done(maps, [](void *mem_) {
+            DocIdx *mem = reinterpret_cast<DocIdx*>(mem_);
+	    delete[] mem;
+        });
+
+    // Return the numpy array.
+    const auto byte_size = sizeof(DocIdx);
+    return py::array(std::vector<int64_t>{num_samples, 3}, // shape
+                     {3*byte_size, byte_size}, // C-style contiguous strides
+                     maps, // the data pointer
+                     free_when_done); // numpy array references
+
+}
+
+
+py::array build_mapping(const py::array_t<int64_t>& docs_,
+                        const py::array_t<int>& sizes_,
+                        const int num_epochs,
+                        const uint64_t max_num_samples,
+                        const int max_seq_length,
+                        const double short_seq_prob,
+                        const int seed,
+			const bool verbose,
+			const int32_t min_num_sent) {
+
+    if (sizes_.size() > std::numeric_limits<uint32_t>::max()) {
+        if (verbose) {
+	   cout << "    using uint64 for data mapping..." << endl << std::flush;
+	}
+	return build_mapping_impl<uint64_t>(docs_, sizes_, num_epochs,
+					    max_num_samples, max_seq_length,
+					    short_seq_prob, seed, verbose,
+					    min_num_sent);
+    } else {
+       if (verbose) {
+	   cout << "    using uint32 for data mapping..." << endl << std::flush;
+       }
+       return build_mapping_impl<uint32_t>(docs_, sizes_, num_epochs,
+					   max_num_samples, max_seq_length,
+					   short_seq_prob, seed, verbose,
+					   min_num_sent);
+    }
+}
+
+template<typename DocIdx>
+py::array build_blocks_mapping_impl(const py::array_t<int64_t>& docs_,
+                                    const py::array_t<int32_t>& sizes_,
+                                    const py::array_t<int32_t>& titles_sizes_,
+                                    const int32_t num_epochs,
+                                    const uint64_t max_num_samples,
+                                    const int32_t max_seq_length,
+                                    const int32_t seed,
+                                    const bool verbose,
+                                    const bool use_one_sent_blocks) {
+    /* Build a mapping of (start-index, end-index, sequence-length) where
+       start and end index are the indices of the sentences in the sample
+       and sequence-length is the target sequence length.
+    */
+
+    // Consistency checks.
+    assert(num_epochs > 0);
+    assert(max_seq_length > 1);
+    assert(seed > 0);
+
+    // Remove bound checks.
+    auto docs = docs_.unchecked<1>();
+    auto sizes = sizes_.unchecked<1>();
+    auto titles_sizes = titles_sizes_.unchecked<1>();
+
+    if (verbose) {
+        const auto sent_start_index = docs[0];
+        const auto sent_end_index = docs[docs_.shape(0) - 1];
+        const auto num_sentences = sent_end_index - sent_start_index;
+        cout << "    using:" << endl << std::flush;
+        cout << "     number of documents:            " << docs_.shape(0) - 1 <<
+          endl << std::flush;
+        cout << "     sentences range:                [" << sent_start_index <<
+        ", " << sent_end_index << ")" << endl << std::flush;
+        cout << "     total number of sentences:      " << num_sentences <<
+          endl << std::flush;
+        cout << "     number of epochs:               " << num_epochs <<
+          endl << std::flush;
+        cout << "     maximum number of samples:      " << max_num_samples <<
+          endl << std::flush;
+        cout << "     maximum sequence length:        " << max_seq_length <<
+          endl << std::flush;
+        cout << "     seed:                           " << seed << endl <<
+          std::flush;
+    }
+
+    // Mapping and its length (1D).
+    int64_t num_samples = -1;
+    DocIdx* maps = NULL;
+
+    // Acceptable number of sentences per block.
+    int min_num_sent = 2;
+    if (use_one_sent_blocks) {
+        min_num_sent = 1;
+    }
+
+    // Perform two iterations, in the first iteration get the size
+    // and allocate memory and in the second iteration populate the map.
+    bool second = false;
+    for (int32_t iteration=0; iteration<2; ++iteration) {
+
+        // Set the flag on second iteration.
+        second = (iteration == 1);
+
+        // Current map index.
+        uint64_t map_index = 0;
+
+        uint64_t empty_docs = 0;
+        uint64_t one_sent_docs = 0;
+        uint64_t long_sent_docs = 0;
+        // For each epoch:
+        for (int32_t epoch=0; epoch<num_epochs; ++epoch) {
+            // assign every block a unique id
+            int32_t block_id = 0;
+
+            if (map_index >= max_num_samples) {
+                if (verbose && (!second)) {
+                cout << "    reached " << max_num_samples << " samples after "
+                     << epoch << " epochs ..." << endl << std::flush;
+                }
+                break;
+            }
+            // For each document:
+            for (int32_t doc=0; doc<(docs.shape(0) - 1); ++doc) {
+
+                // Document sentences are in [sent_index_first, sent_index_last)
+                const auto sent_index_first = docs[doc];
+                const auto sent_index_last = docs[doc + 1];
+                const auto target_seq_len = max_seq_length - titles_sizes[doc];
+
+                // At the begining of the document previous index is the
+                // start index.
+                auto prev_start_index = sent_index_first;
+
+                // Remaining documents.
+                auto num_remain_sent = sent_index_last - sent_index_first;
+
+                // Some bookkeeping
+                if ((epoch == 0) && (!second)) {
+                    if (num_remain_sent == 0) {
+		                ++empty_docs;
+                    }
+                    if (num_remain_sent == 1) {
+		                ++one_sent_docs;
+                    }
+                }
+                // Detect documents with long sentences.
+                bool contains_long_sentence = false;
+                if (num_remain_sent >= min_num_sent) {
+                    for (auto sent_index=sent_index_first;
+                    sent_index < sent_index_last; ++sent_index) {
+                        if (sizes[sent_index] > LONG_SENTENCE_LEN){
+                            if ((epoch == 0) && (!second)) {
+                                ++long_sent_docs;
+                            }
+                            contains_long_sentence = true;
+                            break;
+                        }
+                    }
+                }
+                // If we have enough sentences and no long sentences.
+                if ((num_remain_sent >= min_num_sent) && (!contains_long_sentence)) {
+
+                    // Set values.
+                    auto seq_len = int32_t{0};
+                    auto num_sent = int32_t{0};
+
+                    // Loop through sentences.
+                    for (auto sent_index=sent_index_first;
+                         sent_index < sent_index_last; ++sent_index) {
+
+                            // Add the size and number of sentences.
+                            seq_len += sizes[sent_index];
+                            ++num_sent;
+                            --num_remain_sent;
+
+                        // If we have reached the target length.
+                        // and there are an acceptable number of sentences left
+                        // and if we have at least the minimum number of sentences.
+                        // or if we have reached end of the document.
+                        if (((seq_len >= target_seq_len) &&
+                             (num_remain_sent >= min_num_sent) &&
+                             (num_sent >= min_num_sent) ) || (num_remain_sent == 0)) {
+
+                            // Populate the map.
+                            if (second) {
+                                const auto map_index_0 = 4 * map_index;
+                                // Each sample has 4 items: the starting sentence index, ending sentence index,
+                                // the index of the document from which the block comes (used for fetching titles)
+                                // and the unique id of the block (used for creating block indexes)
+
+                                maps[map_index_0] = static_cast<DocIdx>(prev_start_index);
+                                maps[map_index_0 + 1] = static_cast<DocIdx>(sent_index + 1);
+                                maps[map_index_0 + 2] = static_cast<DocIdx>(doc);
+                                maps[map_index_0 + 3] = static_cast<DocIdx>(block_id);
+                            }
+
+                            // Update indices / counters.
+                            ++map_index;
+                            ++block_id;
+                            prev_start_index = sent_index + 1;
+                            seq_len = 0;
+                            num_sent = 0;
+                        }
+                    } // for (auto sent_index=sent_index_first; ...
+                } // if (num_remain_sent > 1) {
+            } // for (int doc=0; doc < num_docs; ++doc) {
+        } // for (int epoch=0; epoch < num_epochs; ++epoch) {
+
+        if (!second) {
+            if (verbose) {
+	        cout << "   number of empty documents: " << empty_docs <<
+              endl << std::flush;
+            cout << "   number of documents with one sentence: " <<
+              one_sent_docs << endl << std::flush;
+            cout << "   number of documents with long sentences: " <<
+              long_sent_docs << endl << std::flush;
+            cout << "   will create mapping for " << map_index <<
+              " samples" << endl << std::flush;
+            }
+            assert(maps == NULL);
+            assert(num_samples < 0);
+            maps = new DocIdx[4*map_index];
+            num_samples = static_cast<int64_t>(map_index);
+        }
+
+    } // for (int iteration=0; iteration < 2; ++iteration) {
+
+    // Shuffle.
+    // We need a 64 bit random number generator as we might have more
+    // than 2 billion samples.
+    std::mt19937_64 rand64_gen(seed + 1);
+    for (auto i=(num_samples - 1); i > 0; --i) {
+        const auto j = static_cast<int64_t>(rand64_gen() % (i + 1));
+        const auto i0 = 4 * i;
+        const auto j0 = 4 * j;
+        // Swap values.
+        swap(maps[i0], maps[j0]);
+        swap(maps[i0 + 1], maps[j0 + 1]);
+        swap(maps[i0 + 2], maps[j0 + 2]);
+        swap(maps[i0 + 3], maps[j0 + 3]);
+    }
+
+    // Method to deallocate memory.
+    py::capsule free_when_done(maps, [](void *mem_) {
+            DocIdx *mem = reinterpret_cast<DocIdx*>(mem_);
+	    delete[] mem;
+        });
+
+    // Return the numpy array.
+    const auto byte_size = sizeof(DocIdx);
+    return py::array(std::vector<int64_t>{num_samples, 4}, // shape
+                     {4*byte_size, byte_size}, // C-style contiguous strides
+                     maps, // the data pointer
+                     free_when_done); // numpy array references
+
+}
+
+py::array build_blocks_mapping(const py::array_t<int64_t>& docs_,
+                               const py::array_t<int>& sizes_,
+                               const py::array_t<int>& titles_sizes_,
+                               const int num_epochs,
+                               const uint64_t max_num_samples,
+                               const int max_seq_length,
+                               const int seed,
+                    const bool verbose,
+                    const bool use_one_sent_blocks) {
+
+    if (sizes_.size() > std::numeric_limits<uint32_t>::max()) {
+        if (verbose) {
+	   cout << "    using uint64 for data mapping..." << endl << std::flush;
+	}
+	return build_blocks_mapping_impl<uint64_t>(docs_, sizes_, titles_sizes_,
+	                    num_epochs, max_num_samples, max_seq_length, seed, verbose, use_one_sent_blocks);
+    } else {
+       if (verbose) {
+	   cout << "    using uint32 for data mapping..." << endl << std::flush;
+       }
+       return build_blocks_mapping_impl<uint32_t>(docs_, sizes_, titles_sizes_,
+                        num_epochs, max_num_samples, max_seq_length, seed, verbose, use_one_sent_blocks);
+    }
+}
+
+PYBIND11_MODULE(helpers, m) {
+    m.def("build_mapping", &build_mapping);
+    m.def("build_blocks_mapping", &build_blocks_mapping);
+    m.def("build_sample_idx", &build_sample_idx);
+    m.def("build_blending_indices", &build_blending_indices);
+}

examples/tutorial/sequence_parallel/data/datasets/ict_dataset.py

@@ -0,0 +1,156 @@
+import itertools
+import random
+
+import numpy as np
+from torch.utils.data import Dataset
+
+from megatron import get_tokenizer
+from megatron import get_args
+from megatron.data.dataset_utils import get_indexed_dataset_
+from megatron.data.realm_dataset_utils import get_block_samples_mapping
+
+def make_attention_mask(source_block, target_block):
+    """
+    Returns a 2-dimensional (2-D) attention mask
+    :param source_block: 1-D array
+    :param target_block: 1-D array
+    """
+    mask = (target_block[None, :] >= 1) * (source_block[:, None] >= 1)
+    mask = mask.astype(np.int64)
+    # (source_length, target_length)
+    return mask
+
+def get_ict_dataset(use_titles=True, query_in_block_prob=1):
+    """Get a dataset which uses block samples mappings to get ICT/block indexing data (via get_block())
+    rather than for training, since it is only built with a single epoch sample mapping.
+    """
+    args = get_args()
+    block_dataset = get_indexed_dataset_(args.data_path, 'mmap', True)
+    titles_dataset = get_indexed_dataset_(args.titles_data_path, 'mmap', True)
+
+    kwargs = dict(
+        name='full',
+        block_dataset=block_dataset,
+        title_dataset=titles_dataset,
+        data_prefix=args.data_path,
+        num_epochs=1,
+        max_num_samples=None,
+        max_seq_length=args.seq_length,
+        seed=1,
+        query_in_block_prob=query_in_block_prob,
+        use_titles=use_titles,
+        use_one_sent_docs=args.use_one_sent_docs
+    )
+    dataset = ICTDataset(**kwargs)
+    return dataset
+
+
+class ICTDataset(Dataset):
+    """Dataset containing sentences and their blocks for an inverse cloze task."""
+    def __init__(self, name, block_dataset, title_dataset, data_prefix,
+                 num_epochs, max_num_samples, max_seq_length, query_in_block_prob,
+                 seed, use_titles=True, use_one_sent_docs=False, binary_head=False):
+        self.name = name
+        self.seed = seed
+        self.max_seq_length = max_seq_length
+        self.query_in_block_prob = query_in_block_prob
+        self.block_dataset = block_dataset
+        self.title_dataset = title_dataset
+        self.rng = random.Random(self.seed)
+        self.use_titles = use_titles
+        self.use_one_sent_docs = use_one_sent_docs
+
+        self.samples_mapping = get_block_samples_mapping(
+            block_dataset, title_dataset, data_prefix, num_epochs,
+            max_num_samples, max_seq_length, seed, name, use_one_sent_docs)
+        self.tokenizer = get_tokenizer()
+        self.vocab_id_list = list(self.tokenizer.inv_vocab.keys())
+        self.vocab_id_to_token_list = self.tokenizer.inv_vocab
+        self.cls_id = self.tokenizer.cls
+        self.sep_id = self.tokenizer.sep
+        self.mask_id = self.tokenizer.mask
+        self.pad_id = self.tokenizer.pad
+
+    def __len__(self):
+        return len(self.samples_mapping)
+
+    def __getitem__(self, idx):
+        """Get an ICT example of a pseudo-query and the block of text from which it was extracted"""
+        sample_data = self.samples_mapping[idx]
+        start_idx, end_idx, doc_idx, block_idx = sample_data.as_tuple()
+
+        if self.use_titles:
+            title = self.title_dataset[int(doc_idx)]
+            title_pad_offset = 3 + len(title)
+        else:
+            title = None
+            title_pad_offset = 2
+        block = [self.block_dataset[i] for i in range(start_idx, end_idx)]
+        assert len(block) > 1 or self.use_one_sent_docs or self.query_in_block_prob == 1
+
+        # randint() is inclusive for Python rng
+        rand_sent_idx = self.rng.randint(0, len(block) - 1)
+
+        # keep the query in the context query_in_block_prob fraction of the time.
+        if self.rng.random() < self.query_in_block_prob:
+            query = block[rand_sent_idx].copy()
+        else:
+            query = block.pop(rand_sent_idx)
+
+        # still need to truncate because blocks are concluded when
+        # the sentence lengths have exceeded max_seq_length.
+        query = query[:self.max_seq_length - 2]
+        block = list(itertools.chain(*block))[:self.max_seq_length - title_pad_offset]
+
+        query_tokens, query_pad_mask = self.concat_and_pad_tokens(query)
+        context_tokens, context_pad_mask = self.concat_and_pad_tokens(block, title)
+
+        query_mask = make_attention_mask(query_tokens, query_tokens)
+        context_mask = make_attention_mask(context_tokens, context_tokens)
+
+        block_data = sample_data.as_array()
+
+        sample = {
+            'query_tokens': query_tokens,
+            'query_mask': query_mask,
+            'query_pad_mask': query_pad_mask,
+            'context_tokens': context_tokens,
+            'context_mask': context_mask,
+            'context_pad_mask': context_pad_mask,
+            'block_data': block_data,
+        }
+
+        return sample
+
+    def get_block(self, start_idx, end_idx, doc_idx):
+        """Get the IDs for an evidence block plus the title of the corresponding document"""
+        block = [self.block_dataset[i] for i in range(start_idx, end_idx)]
+        title = self.title_dataset[int(doc_idx)]
+
+        block = list(itertools.chain(*block))[:self.max_seq_length - (3 + len(title))]
+        block_tokens, block_pad_mask = self.concat_and_pad_tokens(block, title)
+
+        return block_tokens, block_pad_mask
+
+    def get_null_block(self):
+        """Get empty block and title - used in REALM pretraining"""
+        block, title = [], []
+        block_tokens, block_pad_mask = self.concat_and_pad_tokens(block, title)
+
+        return block_tokens, block_pad_mask
+
+    def concat_and_pad_tokens(self, tokens, title=None):
+        """Concat with special tokens and pad sequence to self.max_seq_length"""
+        tokens = list(tokens)
+        if title is None:
+            tokens = [self.cls_id] + tokens + [self.sep_id]
+        else:
+            title = list(title)
+            tokens = [self.cls_id] + title + [self.sep_id] + tokens + [self.sep_id]
+        assert len(tokens) <= self.max_seq_length
+
+        num_pad = self.max_seq_length - len(tokens)
+        pad_mask = [1] * len(tokens) + [0] * num_pad
+        tokens += [self.pad_id] * num_pad
+
+        return np.array(tokens), np.array(pad_mask)

examples/tutorial/sequence_parallel/data/datasets/indexed_dataset.py

@@ -0,0 +1,569 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+# copied from fairseq/fairseq/data/indexed_dataset.py
+# Removed IndexedRawTextDataset since it relied on Fairseq dictionary
+# other slight modifications to remove fairseq dependencies
+# Added document index to index file and made it accessible.
+#    An empty sentence no longer separates documents.
+
+from functools import lru_cache
+import os
+import shutil
+import struct
+from itertools import accumulate
+
+import numpy as np
+import torch
+
+
+def __best_fitting_dtype(vocab_size=None):
+    if vocab_size is not None and vocab_size < 65500:
+        return np.uint16
+    else:
+        return np.int32
+
+
+def get_available_dataset_impl():
+    return ['lazy', 'cached', 'mmap']
+
+
+def infer_dataset_impl(path):
+    if IndexedDataset.exists(path):
+        with open(index_file_path(path), 'rb') as f:
+            magic = f.read(8)
+            if magic == IndexedDataset._HDR_MAGIC:
+                return 'cached'
+            elif magic == MMapIndexedDataset.Index._HDR_MAGIC[:8]:
+                return 'mmap'
+            else:
+                return None
+    else:
+        print(f"Dataset does not exist: {path}")
+        print("Path should be a basename that both .idx and .bin can be appended to get full filenames.")
+        return None
+
+
+def make_builder(out_file, impl, vocab_size=None):
+    if impl == 'mmap':
+        return MMapIndexedDatasetBuilder(out_file, dtype=__best_fitting_dtype(vocab_size))
+    else:
+        return IndexedDatasetBuilder(out_file)
+
+
+def make_dataset(path, impl, skip_warmup=False):
+    if not IndexedDataset.exists(path):
+        print(f"Dataset does not exist: {path}")
+        print("Path should be a basename that both .idx and .bin can be appended to get full filenames.")
+        return None
+    if impl == 'infer':
+        impl = infer_dataset_impl(path)
+    if impl == 'lazy' and IndexedDataset.exists(path):
+        return IndexedDataset(path)
+    elif impl == 'cached' and IndexedDataset.exists(path):
+        return IndexedCachedDataset(path)
+    elif impl == 'mmap' and MMapIndexedDataset.exists(path):
+        return MMapIndexedDataset(path, skip_warmup)
+    print(f"Unknown dataset implementation: {impl}")
+    return None
+
+
+def dataset_exists(path, impl):
+    if impl == 'mmap':
+        return MMapIndexedDataset.exists(path)
+    else:
+        return IndexedDataset.exists(path)
+
+
+def read_longs(f, n):
+    a = np.empty(n, dtype=np.int64)
+    f.readinto(a)
+    return a
+
+
+def write_longs(f, a):
+    f.write(np.array(a, dtype=np.int64))
+
+
+dtypes = {
+    1: np.uint8,
+    2: np.int8,
+    3: np.int16,
+    4: np.int32,
+    5: np.int64,
+    6: np.float,
+    7: np.double,
+    8: np.uint16
+}
+
+
+def code(dtype):
+    for k in dtypes.keys():
+        if dtypes[k] == dtype:
+            return k
+    raise ValueError(dtype)
+
+
+def index_file_path(prefix_path):
+    return prefix_path + '.idx'
+
+
+def data_file_path(prefix_path):
+    return prefix_path + '.bin'
+
+
+def create_doc_idx(sizes):
+    doc_idx = [0]
+    for i, s in enumerate(sizes):
+        if s == 0:
+            doc_idx.append(i + 1)
+    return doc_idx
+
+
+class IndexedDataset(torch.utils.data.Dataset):
+    """Loader for IndexedDataset"""
+    _HDR_MAGIC = b'TNTIDX\x00\x00'
+
+    def __init__(self, path):
+        super().__init__()
+        self.path = path
+        self.data_file = None
+        self.read_index(path)
+
+    def read_index(self, path):
+        with open(index_file_path(path), 'rb') as f:
+            magic = f.read(8)
+            assert magic == self._HDR_MAGIC, (
+                'Index file doesn\'t match expected format. '
+                'Make sure that --dataset-impl is configured properly.'
+            )
+            version = f.read(8)
+            assert struct.unpack('<Q', version) == (1,)
+            code, self.element_size = struct.unpack('<QQ', f.read(16))
+            self.dtype = dtypes[code]
+            self._len, self.s = struct.unpack('<QQ', f.read(16))
+            self.doc_count = struct.unpack('<Q', f.read(8))
+            self.dim_offsets = read_longs(f, self._len + 1)
+            self.data_offsets = read_longs(f, self._len + 1)
+            self.sizes = read_longs(f, self.s)
+            self.doc_idx = read_longs(f, self.doc_count)
+
+    def read_data(self, path):
+        self.data_file = open(data_file_path(path), 'rb', buffering=0)
+
+    def check_index(self, i):
+        if i < 0 or i >= self._len:
+            raise IndexError('index out of range')
+
+    def __del__(self):
+        if self.data_file:
+            self.data_file.close()
+
+    # @lru_cache(maxsize=8)
+    def __getitem__(self, idx):
+        if not self.data_file:
+            self.read_data(self.path)
+        if isinstance(idx, int):
+            i = idx
+            self.check_index(i)
+            tensor_size = self.sizes[self.dim_offsets[i]:self.dim_offsets[i + 1]]
+            a = np.empty(tensor_size, dtype=self.dtype)
+            self.data_file.seek(self.data_offsets[i] * self.element_size)
+            self.data_file.readinto(a)
+            return a
+        elif isinstance(idx, slice):
+            start, stop, step = idx.indices(len(self))
+            if step != 1:
+                raise ValueError("Slices into indexed_dataset must be contiguous")
+            sizes = self.sizes[self.dim_offsets[start]:self.dim_offsets[stop]]
+            size = sum(sizes)
+            a = np.empty(size, dtype=self.dtype)
+            self.data_file.seek(self.data_offsets[start] * self.element_size)
+            self.data_file.readinto(a)
+            offsets = list(accumulate(sizes))
+            sents = np.split(a, offsets[:-1])
+            return sents
+
+    def __len__(self):
+        return self._len
+
+    def num_tokens(self, index):
+        return self.sizes[index]
+
+    def size(self, index):
+        return self.sizes[index]
+
+    @staticmethod
+    def exists(path):
+        return (
+            os.path.exists(index_file_path(path)) and os.path.exists(data_file_path(path))
+        )
+
+    @property
+    def supports_prefetch(self):
+        return False  # avoid prefetching to save memory
+
+
+class IndexedCachedDataset(IndexedDataset):
+
+    def __init__(self, path):
+        super().__init__(path)
+        self.cache = None
+        self.cache_index = {}
+
+    @property
+    def supports_prefetch(self):
+        return True
+
+    def prefetch(self, indices):
+        if all(i in self.cache_index for i in indices):
+            return
+        if not self.data_file:
+            self.read_data(self.path)
+        indices = sorted(set(indices))
+        total_size = 0
+        for i in indices:
+            total_size += self.data_offsets[i + 1] - self.data_offsets[i]
+        self.cache = np.empty(total_size, dtype=self.dtype)
+        ptx = 0
+        self.cache_index.clear()
+        for i in indices:
+            self.cache_index[i] = ptx
+            size = self.data_offsets[i + 1] - self.data_offsets[i]
+            a = self.cache[ptx: ptx + size]
+            self.data_file.seek(self.data_offsets[i] * self.element_size)
+            self.data_file.readinto(a)
+            ptx += size
+        if self.data_file:
+            # close and delete data file after prefetch so we can pickle
+            self.data_file.close()
+            self.data_file = None
+
+    # @lru_cache(maxsize=8)
+    def __getitem__(self, idx):
+        if isinstance(idx, int):
+            i = idx
+            self.check_index(i)
+            tensor_size = self.sizes[self.dim_offsets[i]:self.dim_offsets[i + 1]]
+            a = np.empty(tensor_size, dtype=self.dtype)
+            ptx = self.cache_index[i]
+            np.copyto(a, self.cache[ptx: ptx + a.size])
+            return a
+        elif isinstance(idx, slice):
+            # Hack just to make this work, can optimizer later if necessary
+            sents = []
+            for i in range(*idx.indices(len(self))):
+                sents.append(self[i])
+            return sents
+
+
+class IndexedDatasetBuilder(object):
+    element_sizes = {
+        np.uint8: 1,
+        np.int8: 1,
+        np.int16: 2,
+        np.int32: 4,
+        np.int64: 8,
+        np.float: 4,
+        np.double: 8
+    }
+
+    def __init__(self, out_file, dtype=np.int32):
+        self.out_file = open(out_file, 'wb')
+        self.dtype = dtype
+        self.data_offsets = [0]
+        self.dim_offsets = [0]
+        self.sizes = []
+        self.element_size = self.element_sizes[self.dtype]
+        self.doc_idx = [0]
+
+    def add_item(self, tensor):
+        bytes = self.out_file.write(np.array(tensor.numpy(), dtype=self.dtype))
+        self.data_offsets.append(self.data_offsets[-1] + bytes / self.element_size)
+        for s in tensor.size():
+            self.sizes.append(s)
+        self.dim_offsets.append(self.dim_offsets[-1] + len(tensor.size()))
+
+    def end_document(self):
+        self.doc_idx.append(len(self.sizes))
+
+    def merge_file_(self, another_file):
+        index = IndexedDataset(another_file)
+        assert index.dtype == self.dtype
+
+        begin = self.data_offsets[-1]
+        for offset in index.data_offsets[1:]:
+            self.data_offsets.append(begin + offset)
+        self.sizes.extend(index.sizes)
+        begin = self.dim_offsets[-1]
+        for dim_offset in index.dim_offsets[1:]:
+            self.dim_offsets.append(begin + dim_offset)
+
+        with open(data_file_path(another_file), 'rb') as f:
+            while True:
+                data = f.read(1024)
+                if data:
+                    self.out_file.write(data)
+                else:
+                    break
+
+    def finalize(self, index_file):
+        self.out_file.close()
+        index = open(index_file, 'wb')
+        index.write(b'TNTIDX\x00\x00')
+        index.write(struct.pack('<Q', 1))
+        index.write(struct.pack('<QQ', code(self.dtype), self.element_size))
+        index.write(struct.pack('<QQ', len(self.data_offsets) - 1, len(self.sizes)))
+        index.write(struct.pack('<Q', len(self.doc_idx)))
+        write_longs(index, self.dim_offsets)
+        write_longs(index, self.data_offsets)
+        write_longs(index, self.sizes)
+        write_longs(index, self.doc_idx)
+        index.close()
+
+
+def _warmup_mmap_file(path):
+    with open(path, 'rb') as stream:
+        while stream.read(100 * 1024 * 1024):
+            pass
+
+
+class MMapIndexedDataset(torch.utils.data.Dataset):
+    class Index(object):
+        _HDR_MAGIC = b'MMIDIDX\x00\x00'
+
+        @classmethod
+        def writer(cls, path, dtype):
+            class _Writer(object):
+                def __enter__(self):
+                    self._file = open(path, 'wb')
+
+                    self._file.write(cls._HDR_MAGIC)
+                    self._file.write(struct.pack('<Q', 1))
+                    self._file.write(struct.pack('<B', code(dtype)))
+
+                    return self
+
+                @staticmethod
+                def _get_pointers(sizes):
+                    dtype_size = dtype().itemsize
+                    address = 0
+                    pointers = []
+
+                    for size in sizes:
+                        pointers.append(address)
+                        address += size * dtype_size
+
+                    return pointers
+
+                def write(self, sizes, doc_idx):
+                    pointers = self._get_pointers(sizes)
+
+                    self._file.write(struct.pack('<Q', len(sizes)))
+                    self._file.write(struct.pack('<Q', len(doc_idx)))
+
+                    sizes = np.array(sizes, dtype=np.int32)
+                    self._file.write(sizes.tobytes(order='C'))
+                    del sizes
+
+                    pointers = np.array(pointers, dtype=np.int64)
+                    self._file.write(pointers.tobytes(order='C'))
+                    del pointers
+
+                    doc_idx = np.array(doc_idx, dtype=np.int64)
+                    self._file.write(doc_idx.tobytes(order='C'))
+
+                def __exit__(self, exc_type, exc_val, exc_tb):
+                    self._file.close()
+
+            return _Writer()
+
+        def __init__(self, path, skip_warmup=False):
+            with open(path, 'rb') as stream:
+                magic_test = stream.read(9)
+                assert self._HDR_MAGIC == magic_test, (
+                    'Index file doesn\'t match expected format. '
+                    'Make sure that --dataset-impl is configured properly.'
+                )
+                version = struct.unpack('<Q', stream.read(8))
+                assert (1,) == version
+
+                dtype_code, = struct.unpack('<B', stream.read(1))
+                self._dtype = dtypes[dtype_code]
+                self._dtype_size = self._dtype().itemsize
+
+                self._len = struct.unpack('<Q', stream.read(8))[0]
+                self._doc_count = struct.unpack('<Q', stream.read(8))[0]
+                offset = stream.tell()
+
+            if not skip_warmup:
+                print("    warming up index mmap file...")
+                _warmup_mmap_file(path)
+
+            self._bin_buffer_mmap = np.memmap(path, mode='r', order='C')
+            self._bin_buffer = memoryview(self._bin_buffer_mmap)
+            print("    reading sizes...")
+            self._sizes = np.frombuffer(
+                self._bin_buffer,
+                dtype=np.int32,
+                count=self._len,
+                offset=offset)
+            print("    reading pointers...")
+            self._pointers = np.frombuffer(self._bin_buffer, dtype=np.int64, count=self._len,
+                                           offset=offset + self._sizes.nbytes)
+            print("    reading document index...")
+            self._doc_idx = np.frombuffer(self._bin_buffer, dtype=np.int64, count=self._doc_count,
+                                          offset=offset + self._sizes.nbytes + self._pointers.nbytes)
+
+        def __del__(self):
+            self._bin_buffer_mmap._mmap.close()
+            del self._bin_buffer_mmap
+
+        @property
+        def dtype(self):
+            return self._dtype
+
+        @property
+        def sizes(self):
+            return self._sizes
+
+        @property
+        def doc_idx(self):
+            return self._doc_idx
+
+        @lru_cache(maxsize=8)
+        def __getitem__(self, i):
+            return self._pointers[i], self._sizes[i]
+
+        def __len__(self):
+            return self._len
+
+    def __init__(self, path, skip_warmup=False):
+        super().__init__()
+
+        self._path = None
+        self._index = None
+        self._bin_buffer = None
+
+        self._do_init(path, skip_warmup)
+
+    def __getstate__(self):
+        return self._path
+
+    def __setstate__(self, state):
+        self._do_init(state)
+
+    def _do_init(self, path, skip_warmup):
+        self._path = path
+        self._index = self.Index(index_file_path(self._path), skip_warmup)
+
+        if not skip_warmup:
+            print("    warming up data mmap file...")
+            _warmup_mmap_file(data_file_path(self._path))
+        print("    creating numpy buffer of mmap...")
+        self._bin_buffer_mmap = np.memmap(data_file_path(self._path), mode='r', order='C')
+        print("    creating memory view of numpy buffer...")
+        self._bin_buffer = memoryview(self._bin_buffer_mmap)
+
+    def __del__(self):
+        self._bin_buffer_mmap._mmap.close()
+        del self._bin_buffer_mmap
+        del self._index
+
+    def __len__(self):
+        return len(self._index)
+
+    # @lru_cache(maxsize=8)
+    def __getitem__(self, idx):
+        if isinstance(idx, int):
+            ptr, size = self._index[idx]
+            np_array = np.frombuffer(self._bin_buffer, dtype=self._index.dtype,
+                                     count=size, offset=ptr)
+            return np_array
+        elif isinstance(idx, slice):
+            start, stop, step = idx.indices(len(self))
+            if step != 1:
+                raise ValueError("Slices into indexed_dataset must be contiguous")
+            ptr = self._index._pointers[start]
+            sizes = self._index._sizes[idx]
+            offsets = list(accumulate(sizes))
+            total_size = sum(sizes)
+            np_array = np.frombuffer(self._bin_buffer, dtype=self._index.dtype,
+                                     count=total_size, offset=ptr)
+            sents = np.split(np_array, offsets[:-1])
+            return sents
+
+    def get(self, idx, offset=0, length=None):
+        """ Retrieves a single item from the dataset with the option to only
+        return a portion of the item.
+
+        get(idx) is the same as [idx] but get() does not support slicing.
+        """
+        ptr, size = self._index[idx]
+        if length is None:
+            length = size - offset
+        ptr += offset * np.dtype(self._index.dtype).itemsize
+        np_array = np.frombuffer(self._bin_buffer, dtype=self._index.dtype,
+                                 count=length, offset=ptr)
+        return np_array
+
+    @property
+    def sizes(self):
+        return self._index.sizes
+
+    @property
+    def doc_idx(self):
+        return self._index.doc_idx
+
+    def get_doc_idx(self):
+        return self._index._doc_idx
+
+    def set_doc_idx(self, doc_idx_):
+        self._index._doc_idx = doc_idx_
+
+    @property
+    def supports_prefetch(self):
+        return False
+
+    @staticmethod
+    def exists(path):
+        return (
+            os.path.exists(index_file_path(path)) and os.path.exists(data_file_path(path))
+        )
+
+
+class MMapIndexedDatasetBuilder(object):
+    def __init__(self, out_file, dtype=np.int64):
+        self._data_file = open(out_file, 'wb')
+        self._dtype = dtype
+        self._sizes = []
+        self._doc_idx = [0]
+
+    def add_item(self, tensor):
+        np_array = np.array(tensor.numpy(), dtype=self._dtype)
+        self._data_file.write(np_array.tobytes(order='C'))
+        self._sizes.append(np_array.size)
+
+    def end_document(self):
+        self._doc_idx.append(len(self._sizes))
+
+    def merge_file_(self, another_file):
+        # Concatenate index
+        index = MMapIndexedDataset.Index(index_file_path(another_file))
+        assert index.dtype == self._dtype
+
+        for size in index.sizes:
+            self._sizes.append(size)
+
+        # Concatenate data
+        with open(data_file_path(another_file), 'rb') as f:
+            shutil.copyfileobj(f, self._data_file)
+
+    def finalize(self, index_file):
+        self._data_file.close()
+
+        with MMapIndexedDataset.Index.writer(index_file, self._dtype) as index:
+            index.write(self._sizes, self._doc_idx)

examples/tutorial/sequence_parallel/data/datasets/test/test_indexed_dataset.py

@@ -0,0 +1,125 @@
+# This file isn't really a formal automated test, it's just a place to
+# put some code used during development and manual testing of
+# indexed_dataset.
+
+from megatron.data import indexed_dataset
+from megatron.tokenizer import build_tokenizer
+import argparse
+import os
+import sys
+
+import torch
+
+script_dir = os.path.dirname(os.path.realpath(__file__))
+sys.path.append(os.path.join(script_dir, "../../../"))
+
+
+def test_indexed_dataset(args):
+    ds = indexed_dataset.make_dataset(args.data, args.dataset_impl)
+    tokenizer = build_tokenizer(args)
+    print(len(ds.doc_idx))
+    print(len(ds))
+    print(ds.doc_idx[-1])
+    if ds.supports_prefetch:
+        # just prefetch the whole thing in test (so assume it is small)
+        ds.prefetch(range(len(ds)))
+    if args.count > len(ds.doc_idx) - 1:
+        args.count = len(ds.doc_idx) - 1
+
+    for i in range(args.count):
+        start = ds.doc_idx[i]
+        end = ds.doc_idx[i + 1]
+        ids = ds[start:end]
+        print(f"Document {i}:")
+        print("--------------")
+        for s in ids:
+            assert len(s) > 0
+            l = s.data.tolist()
+            text = tokenizer.detokenize(l)
+            print(text)
+            print("---")
+
+
+def test_indexed_dataset_get(args):
+    ds = indexed_dataset.make_dataset(args.data, args.dataset_impl)
+    tokenizer = build_tokenizer(args)
+    size = ds.sizes[0]
+    print(f"size: {size}")
+    full = ds.get(0)
+    print(full)
+    # print(tokenizer.detokenize(full.data.tolist()))
+    print("---")
+    end = ds.get(0, offset=size - 10)
+    print(end)
+    # print(tokenizer.detokenize(end.data.tolist()))
+
+    start = ds.get(0, length=10)
+    print(start)
+    # print(tokenizer.detokenize(start.data.tolist()))
+
+    part = ds.get(0, offset=2, length=8)
+    print(part)
+    # print(tokenizer.detokenize(part.data.tolist()))
+
+# def test_albert_dataset(args):
+#     # tokenizer = FullBertTokenizer(args.vocab, do_lower_case=True)
+#     # idataset = indexed_dataset.make_dataset(args.data, args.dataset_impl)
+#     # ds = AlbertDataset(idataset, tokenizer)
+#     ds = AlbertDataset.from_paths(args.vocab, args.data, args.dataset_impl,
+#                                   args.epochs, args.max_num_samples,
+#                                   args.masked_lm_prob, args.seq_length,
+#                                   args.short_seq_prob, args.seed)
+#     truncated = 0
+#     total = 0
+#     for i, s in enumerate(ds):
+#         ids = s['text']
+#         tokens = ds.tokenizer.convert_ids_to_tokens(ids)
+#         print(tokens)
+#         if i >= args.count-1:
+#             exit()
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--data', type=str, help='prefix to data files')
+    parser.add_argument('--dataset-impl', type=str, default='infer',
+                        choices=['lazy', 'cached', 'mmap', 'infer'])
+    parser.add_argument('--count', type=int, default=10,
+                        help='Number of samples/documents to print')
+
+    group = parser.add_argument_group(title='tokenizer')
+    group.add_argument('--tokenizer-type', type=str, required=True,
+                       choices=['BertWordPieceLowerCase',
+                                'GPT2BPETokenizer'],
+                       help='What type of tokenizer to use.')
+    group.add_argument('--vocab-file', type=str, default=None,
+                       help='Path to the vocab file')
+    group.add_argument('--merge-file', type=str, default=None,
+                       help='Path to the BPE merge file (if necessary).')
+
+    parser.add_argument('--epochs', type=int, default=5,
+                        help='Number of epochs to plan for')
+    parser.add_argument('--max-num-samples', type=int, default=None,
+                        help='Maximum number of samples to plan for')
+    parser.add_argument('--masked-lm-prob', type=float, default=0.15,
+                        help='probability of masking tokens')
+    parser.add_argument('--seq-length', type=int, default=512,
+                        help='maximum sequence length')
+    parser.add_argument('--short-seq-prob', type=float, default=0.1,
+                        help='probability of creating a short sequence')
+    parser.add_argument('--seed', type=int, default=1234,
+                        help='random seed')
+    args = parser.parse_args()
+    args.rank = 0
+    args.make_vocab_size_divisible_by = 128
+    args.tensor_model_parallel_size = 1
+
+    if args.dataset_impl == "infer":
+        args.dataset_impl = indexed_dataset.infer_dataset_impl(args.data)
+
+#    test_albert_dataset(args)
+    test_indexed_dataset_get(args)
+
+
+if __name__ == "__main__":
+    main()

examples/tutorial/sequence_parallel/data/datasets/test/test_preprocess_data.sh

@@ -0,0 +1,10 @@
+#!/bin/bash
+
+IMPL=cached
+python ../preprocess_data.py \
+       --input test_samples.json \
+       --vocab vocab.txt \
+       --dataset-impl ${IMPL} \
+       --output-prefix test_samples_${IMPL} \
+       --workers 1 \
+       --log-interval 2

examples/tutorial/sequence_parallel/data/tokenizer/__init__.py

@@ -0,0 +1,38 @@
+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from .tokenizer import build_tokenizer
+
+
+_TOKENIZER = None
+_PADDED_VOCAB_SIZE = -1
+
+
+def initialize_tokenizer(vocab_file, tokenizer_type, vocab_extra_ids=0):
+    tokenizer, padded_vocab_size = build_tokenizer(vocab_file, tokenizer_type, vocab_extra_ids)
+    global _TOKENIZER, _PADDED_VOCAB_SIZE
+    _TOKENIZER = tokenizer
+    _PADDED_VOCAB_SIZE = padded_vocab_size
+
+
+def get_tokenizer():
+    global _TOKENIZER
+    return _TOKENIZER
+
+
+def get_padded_vocab_size():
+    global _PADDED_VOCAB_SIZE
+    return _PADDED_VOCAB_SIZE

examples/tutorial/sequence_parallel/data/tokenizer/bert_tokenization.py

@@ -0,0 +1,431 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Tokenization classes."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import collections
+import re
+import unicodedata
+import six
+
+
+def validate_case_matches_checkpoint(do_lower_case, init_checkpoint):
+    """Checks whether the casing config is consistent with the checkpoint name."""
+
+    # The casing has to be passed in by the user and there is no explicit check
+    # as to whether it matches the checkpoint. The casing information probably
+    # should have been stored in the bert_config.json file, but it's not, so
+    # we have to heuristically detect it to validate.
+
+    if not init_checkpoint:
+        return
+
+    m = re.match("^.*?([A-Za-z0-9_-]+)/bert_model.ckpt", init_checkpoint)
+    if m is None:
+        return
+
+    model_name = m.group(1)
+
+    lower_models = [
+        "uncased_L-24_H-1024_A-16", "uncased_L-12_H-768_A-12",
+        "multilingual_L-12_H-768_A-12", "chinese_L-12_H-768_A-12"
+    ]
+
+    cased_models = [
+        "cased_L-12_H-768_A-12", "cased_L-24_H-1024_A-16",
+        "multi_cased_L-12_H-768_A-12"
+    ]
+
+    is_bad_config = False
+    if model_name in lower_models and not do_lower_case:
+        is_bad_config = True
+        actual_flag = "False"
+        case_name = "lowercased"
+        opposite_flag = "True"
+
+    if model_name in cased_models and do_lower_case:
+        is_bad_config = True
+        actual_flag = "True"
+        case_name = "cased"
+        opposite_flag = "False"
+
+    if is_bad_config:
+        raise ValueError(
+            "You passed in `--do_lower_case=%s` with `--init_checkpoint=%s`. "
+            "However, `%s` seems to be a %s model, so you "
+            "should pass in `--do_lower_case=%s` so that the fine-tuning matches "
+            "how the model was pre-training. If this error is wrong, please "
+            "just comment out this check." % (actual_flag, init_checkpoint,
+                                              model_name, case_name, opposite_flag))
+
+
+def convert_to_unicode(text):
+    """Converts `text` to Unicode (if it's not already), assuming utf-8 input."""
+    if six.PY3:
+        if isinstance(text, str):
+            return text
+        elif isinstance(text, bytes):
+            return text.decode("utf-8", "ignore")
+        else:
+            raise ValueError("Unsupported string type: %s" % (type(text)))
+    elif six.PY2:
+        if isinstance(text, str):
+            return text.decode("utf-8", "ignore")
+        elif isinstance(text, unicode):
+            return text
+        else:
+            raise ValueError("Unsupported string type: %s" % (type(text)))
+    else:
+        raise ValueError("Not running on Python2 or Python 3?")
+
+
+def printable_text(text):
+    """Returns text encoded in a way suitable for print or `tf.logging`."""
+
+    # These functions want `str` for both Python2 and Python3, but in one case
+    # it's a Unicode string and in the other it's a byte string.
+    if six.PY3:
+        if isinstance(text, str):
+            return text
+        elif isinstance(text, bytes):
+            return text.decode("utf-8", "ignore")
+        else:
+            raise ValueError("Unsupported string type: %s" % (type(text)))
+    elif six.PY2:
+        if isinstance(text, str):
+            return text
+        elif isinstance(text, unicode):
+            return text.encode("utf-8")
+        else:
+            raise ValueError("Unsupported string type: %s" % (type(text)))
+    else:
+        raise ValueError("Not running on Python2 or Python 3?")
+
+
+def load_vocab(vocab_file):
+    """Loads a vocabulary file into a dictionary."""
+    vocab = collections.OrderedDict()
+    index = 0
+    with open(vocab_file, "r") as reader:
+        while True:
+            token = convert_to_unicode(reader.readline())
+            if not token:
+                break
+            token = token.strip()
+            vocab[token] = index
+            index += 1
+    return vocab
+
+
+def convert_by_vocab(vocab, items):
+    """Converts a sequence of [tokens|ids] using the vocab."""
+    output = []
+    for item in items:
+        output.append(vocab[item])
+    return output
+
+
+def convert_tokens_to_ids(vocab, tokens):
+    return convert_by_vocab(vocab, tokens)
+
+
+def convert_ids_to_tokens(inv_vocab, ids):
+    return convert_by_vocab(inv_vocab, ids)
+
+
+def whitespace_tokenize(text):
+    """Runs basic whitespace cleaning and splitting on a piece of text."""
+    text = text.strip()
+    if not text:
+        return []
+    tokens = text.split()
+    return tokens
+
+
+class FullTokenizer(object):
+    """Runs end-to-end tokenization."""
+
+    def __init__(self, vocab_file, do_lower_case=True):
+        self.vocab = load_vocab(vocab_file)
+        self.inv_vocab = {v: k for k, v in self.vocab.items()}
+        self.basic_tokenizer = BasicTokenizer(do_lower_case=do_lower_case)
+        self.wordpiece_tokenizer = WordpieceTokenizer(vocab=self.vocab)
+
+    def tokenize(self, text):
+        split_tokens = []
+        for token in self.basic_tokenizer.tokenize(text):
+            for sub_token in self.wordpiece_tokenizer.tokenize(token):
+                split_tokens.append(sub_token)
+
+        return split_tokens
+
+    def convert_tokens_to_ids(self, tokens):
+        return convert_by_vocab(self.vocab, tokens)
+
+    def convert_ids_to_tokens(self, ids):
+        return convert_by_vocab(self.inv_vocab, ids)
+
+    @staticmethod
+    def convert_tokens_to_string(tokens, clean_up_tokenization_spaces=True):
+        """ Converts a sequence of tokens (string) in a single string. """
+
+        def clean_up_tokenization(out_string):
+            """ Clean up a list of simple English tokenization artifacts
+            like spaces before punctuations and abbreviated forms.
+            """
+            out_string = (
+                out_string.replace(" .", ".")
+                    .replace(" ?", "?")
+                    .replace(" !", "!")
+                    .replace(" ,", ",")
+                    .replace(" ' ", "'")
+                    .replace(" n't", "n't")
+                    .replace(" 'm", "'m")
+                    .replace(" 's", "'s")
+                    .replace(" 've", "'ve")
+                    .replace(" 're", "'re")
+            )
+            return out_string
+
+        text = ' '.join(tokens).replace(' ##', '').strip()
+        if clean_up_tokenization_spaces:
+            clean_text = clean_up_tokenization(text)
+            return clean_text
+        else:
+            return text
+
+    def vocab_size(self):
+        return len(self.vocab)
+
+
+class BasicTokenizer(object):
+    """Runs basic tokenization (punctuation splitting, lower casing, etc.)."""
+
+    def __init__(self, do_lower_case=True):
+        """Constructs a BasicTokenizer.
+
+        Args:
+          do_lower_case: Whether to lower case the input.
+        """
+        self.do_lower_case = do_lower_case
+
+    def tokenize(self, text):
+        """Tokenizes a piece of text."""
+        text = convert_to_unicode(text)
+        text = self._clean_text(text)
+
+        # This was added on November 1st, 2018 for the multilingual and Chinese
+        # models. This is also applied to the English models now, but it doesn't
+        # matter since the English models were not trained on any Chinese data
+        # and generally don't have any Chinese data in them (there are Chinese
+        # characters in the vocabulary because Wikipedia does have some Chinese
+        # words in the English Wikipedia.).
+        text = self._tokenize_chinese_chars(text)
+
+        orig_tokens = whitespace_tokenize(text)
+        split_tokens = []
+        for token in orig_tokens:
+            if self.do_lower_case:
+                token = token.lower()
+                token = self._run_strip_accents(token)
+            split_tokens.extend(self._run_split_on_punc(token))
+
+        output_tokens = whitespace_tokenize(" ".join(split_tokens))
+        return output_tokens
+
+    def _run_strip_accents(self, text):
+        """Strips accents from a piece of text."""
+        text = unicodedata.normalize("NFD", text)
+        output = []
+        for char in text:
+            cat = unicodedata.category(char)
+            if cat == "Mn":
+                continue
+            output.append(char)
+        return "".join(output)
+
+    def _run_split_on_punc(self, text):
+        """Splits punctuation on a piece of text."""
+        chars = list(text)
+        i = 0
+        start_new_word = True
+        output = []
+        while i < len(chars):
+            char = chars[i]
+            if _is_punctuation(char):
+                output.append([char])
+                start_new_word = True
+            else:
+                if start_new_word:
+                    output.append([])
+                start_new_word = False
+                output[-1].append(char)
+            i += 1
+
+        return ["".join(x) for x in output]
+
+    def _tokenize_chinese_chars(self, text):
+        """Adds whitespace around any CJK character."""
+        output = []
+        for char in text:
+            cp = ord(char)
+            if self._is_chinese_char(cp):
+                output.append(" ")
+                output.append(char)
+                output.append(" ")
+            else:
+                output.append(char)
+        return "".join(output)
+
+    def _is_chinese_char(self, cp):
+        """Checks whether CP is the codepoint of a CJK character."""
+        # This defines a "chinese character" as anything in the CJK Unicode block:
+        #   https://en.wikipedia.org/wiki/CJK_Unified_Ideographs_(Unicode_block)
+        #
+        # Note that the CJK Unicode block is NOT all Japanese and Korean characters,
+        # despite its name. The modern Korean Hangul alphabet is a different block,
+        # as is Japanese Hiragana and Katakana. Those alphabets are used to write
+        # space-separated words, so they are not treated specially and handled
+        # like the all of the other languages.
+        if ((cp >= 0x4E00 and cp <= 0x9FFF) or  #
+            (cp >= 0x3400 and cp <= 0x4DBF) or  #
+            (cp >= 0x20000 and cp <= 0x2A6DF) or  #
+            (cp >= 0x2A700 and cp <= 0x2B73F) or  #
+            (cp >= 0x2B740 and cp <= 0x2B81F) or  #
+            (cp >= 0x2B820 and cp <= 0x2CEAF) or
+            (cp >= 0xF900 and cp <= 0xFAFF) or  #
+                (cp >= 0x2F800 and cp <= 0x2FA1F)):  #
+            return True
+
+        return False
+
+    def _clean_text(self, text):
+        """Performs invalid character removal and whitespace cleanup on text."""
+        output = []
+        for char in text:
+            cp = ord(char)
+            if cp == 0 or cp == 0xfffd or _is_control(char):
+                continue
+            if _is_whitespace(char):
+                output.append(" ")
+            else:
+                output.append(char)
+        return "".join(output)
+
+
+class WordpieceTokenizer(object):
+    """Runs WordPiece tokenization."""
+
+    def __init__(self, vocab, unk_token="[UNK]", max_input_chars_per_word=200):
+        self.vocab = vocab
+        self.unk_token = unk_token
+        self.max_input_chars_per_word = max_input_chars_per_word
+
+    def tokenize(self, text):
+        """Tokenizes a piece of text into its word pieces.
+
+        This uses a greedy longest-match-first algorithm to perform tokenization
+        using the given vocabulary.
+
+        For example:
+          input = "unaffable"
+          output = ["un", "##aff", "##able"]
+
+        Args:
+          text: A single token or whitespace separated tokens. This should have
+            already been passed through `BasicTokenizer.
+
+        Returns:
+          A list of wordpiece tokens.
+        """
+
+        text = convert_to_unicode(text)
+
+        output_tokens = []
+        for token in whitespace_tokenize(text):
+            chars = list(token)
+            if len(chars) > self.max_input_chars_per_word:
+                output_tokens.append(self.unk_token)
+                continue
+
+            is_bad = False
+            start = 0
+            sub_tokens = []
+            while start < len(chars):
+                end = len(chars)
+                cur_substr = None
+                while start < end:
+                    substr = "".join(chars[start:end])
+                    if start > 0:
+                        substr = "##" + substr
+                    if substr in self.vocab:
+                        cur_substr = substr
+                        break
+                    end -= 1
+                if cur_substr is None:
+                    is_bad = True
+                    break
+                sub_tokens.append(cur_substr)
+                start = end
+
+            if is_bad:
+                output_tokens.append(self.unk_token)
+            else:
+                output_tokens.extend(sub_tokens)
+        return output_tokens
+
+
+def _is_whitespace(char):
+    """Checks whether `chars` is a whitespace character."""
+    # \t, \n, and \r are technically control characters but we treat them
+    # as whitespace since they are generally considered as such.
+    if char == " " or char == "\t" or char == "\n" or char == "\r":
+        return True
+    cat = unicodedata.category(char)
+    if cat == "Zs":
+        return True
+    return False
+
+
+def _is_control(char):
+    """Checks whether `chars` is a control character."""
+    # These are technically control characters but we count them as whitespace
+    # characters.
+    if char == "\t" or char == "\n" or char == "\r":
+        return False
+    cat = unicodedata.category(char)
+    if cat in ("Cc", "Cf"):
+        return True
+    return False
+
+
+def _is_punctuation(char):
+    """Checks whether `chars` is a punctuation character."""
+    cp = ord(char)
+    # We treat all non-letter/number ASCII as punctuation.
+    # Characters such as "^", "$", and "`" are not in the Unicode
+    # Punctuation class but we treat them as punctuation anyways, for
+    # consistency.
+    if ((cp >= 33 and cp <= 47) or (cp >= 58 and cp <= 64) or
+            (cp >= 91 and cp <= 96) or (cp >= 123 and cp <= 126)):
+        return True
+    cat = unicodedata.category(char)
+    if cat.startswith("P"):
+        return True
+    return False

examples/tutorial/sequence_parallel/data/tokenizer/tokenizer.py

@@ -0,0 +1,256 @@
+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Megatron tokenizers."""
+
+from abc import ABC
+from abc import abstractmethod
+from colossalai.core import global_context as gpc
+from colossalai.context import ParallelMode
+
+from .bert_tokenization import FullTokenizer as FullBertTokenizer
+
+
+def build_tokenizer(vocab_file, tokenizer_type, vocab_extra_ids=0):
+    """Initialize tokenizer."""
+    if not gpc.is_initialized(ParallelMode.GLOBAL) or gpc.get_global_rank() == 0:
+        print('> building {} tokenizer ...'.format(tokenizer_type),
+              flush=True)
+
+    # Select and instantiate the tokenizer.
+    if tokenizer_type == 'BertWordPieceLowerCase':
+        tokenizer = _BertWordPieceTokenizer(vocab_file=vocab_file,
+                                            lower_case=True,
+                                            vocab_extra_ids=vocab_extra_ids)
+    elif tokenizer_type == 'BertWordPieceCase':
+        tokenizer = _BertWordPieceTokenizer(vocab_file=vocab_file,
+                                            lower_case=False,
+                                            vocab_extra_ids=vocab_extra_ids)
+    else:
+        raise NotImplementedError('{} tokenizer is not '
+                                  'implemented.'.format(tokenizer_type))
+
+    # Add vocab size.
+    padded_vocab_size = _vocab_size_with_padding(tokenizer.vocab_size)
+
+    return tokenizer, padded_vocab_size
+
+
+def _vocab_size_with_padding(orig_vocab_size, make_vocab_size_divisible_by=128):
+    """Pad vocab size so it is divisible by model parallel size and
+    still having GPU friendly size."""
+
+    after = orig_vocab_size
+
+    if gpc.is_initialized(ParallelMode.TENSOR):
+        multiple = make_vocab_size_divisible_by * gpc.get_world_size(ParallelMode.TENSOR)
+    else:
+        multiple = make_vocab_size_divisible_by
+    while (after % multiple) != 0:
+        after += 1
+    if not gpc.is_initialized(ParallelMode.GLOBAL) or gpc.get_global_rank() == 0:
+        print(' > padded vocab (size: {}) with {} dummy tokens '
+              '(new size: {})'.format(
+                  orig_vocab_size, after - orig_vocab_size, after), flush=True)
+    return after
+
+
+class AbstractTokenizer(ABC):
+    """Abstract class for tokenizer."""
+
+    def __init__(self, name):
+        self.name = name
+        super().__init__()
+
+    @property
+    @abstractmethod
+    def vocab_size(self):
+        pass
+
+    @property
+    @abstractmethod
+    def vocab(self):
+        """Dictionary from vocab text token to id token."""
+        pass
+
+    @property
+    @abstractmethod
+    def inv_vocab(self):
+        """Dictionary from vocab id token to text token."""
+        pass
+
+    @abstractmethod
+    def tokenize(self, text):
+        pass
+
+    def detokenize(self, token_ids):
+        raise NotImplementedError('detokenizer is not implemented for {} '
+                                  'tokenizer'.format(self.name))
+
+    @property
+    def cls(self):
+        raise NotImplementedError('CLS is not provided for {} '
+                                  'tokenizer'.format(self.name))
+
+    @property
+    def sep(self):
+        raise NotImplementedError('SEP is not provided for {} '
+                                  'tokenizer'.format(self.name))
+
+    @property
+    def pad(self):
+        raise NotImplementedError('PAD is not provided for {} '
+                                  'tokenizer'.format(self.name))
+
+    @property
+    def eod(self):
+        raise NotImplementedError('EOD is not provided for {} '
+                                  'tokenizer'.format(self.name))
+
+    @property
+    def mask(self):
+        raise NotImplementedError('MASK is not provided for {} '
+                                  'tokenizer'.format(self.name))
+
+
+class _BertWordPieceTokenizer(AbstractTokenizer):
+    """Original BERT wordpiece tokenizer."""
+
+    def __init__(self, vocab_file, lower_case=True, vocab_extra_ids=0):
+        if lower_case:
+            name = 'BERT Lower Case'
+        else:
+            name = 'BERT Upper Case'
+        super().__init__(name)
+        self.tokenizer = FullBertTokenizer(vocab_file, do_lower_case=lower_case)
+        self.cls_id = self.tokenizer.vocab['[CLS]']
+        self.sep_id = self.tokenizer.vocab['[SEP]']
+        self.pad_id = self.tokenizer.vocab['[PAD]']
+        self.mask_id = self.tokenizer.vocab['[MASK]']
+        self._additional_special_tokens = []
+
+        # (dsachan) Add BOS and EOS tokens
+        SPECIAL_TOKENS = {'eos_token': '[EOS]',
+                          'bos_token': '[BOS]'}
+        self._bos_token = '[BOS]'
+        self.add_token(self._bos_token)
+        self._bos_token_id = self.vocab.get(self._bos_token)
+
+        self._eos_token = '[EOS]'
+        self.add_token(self._eos_token)
+        self._eos_token_id = self.vocab.get(self._eos_token)
+
+        # (dsachan) Add additional special tokens
+        # These can be used as sentinel tokens in T5 model inputs
+        additional_special_tokens = []
+        additional_special_tokens.extend(
+            ["<extra_id_{}>".format(i) for i in range(vocab_extra_ids)])
+        self.add_additional_special_tokens(additional_special_tokens)
+
+    def add_token(self, token):
+        if token not in self.vocab:
+            self.inv_vocab[self.vocab_size] = token
+            # self.vocab_size comes from len(vocab)
+            # and it will increase as we add elements
+            self.vocab[token] = self.vocab_size
+
+    def add_additional_special_tokens(self, tokens_list):
+        setattr(self, "additional_special_tokens", tokens_list)
+        for value in tokens_list:
+            self.add_token(value)
+
+    @property
+    def vocab_size(self):
+        return self.tokenizer.vocab_size()
+
+    @property
+    def vocab(self):
+        return self.tokenizer.vocab
+
+    @property
+    def inv_vocab(self):
+        return self.tokenizer.inv_vocab
+
+    def tokenize(self, text):
+        text_tokens = self.tokenizer.tokenize(text)
+        return self.tokenizer.convert_tokens_to_ids(text_tokens)
+
+    def decode(self, ids):
+        tokens = self.tokenizer.convert_ids_to_tokens(ids)
+        return self.tokenizer.convert_tokens_to_string(tokens)
+
+    def decode_token_ids(self, token_ids):
+        tokens = self.tokenizer.convert_ids_to_tokens(token_ids)
+        exclude_list = ['[PAD]', '[CLS]']
+        non_pads = [t for t in tokens if t not in exclude_list]
+
+        result = ""
+        for s in non_pads:
+            if s.startswith("##"):
+                result += s[2:]
+            else:
+                result += " " + s
+
+        return result
+
+    @property
+    def cls(self):
+        return self.cls_id
+
+    @property
+    def sep(self):
+        return self.sep_id
+
+    @property
+    def pad(self):
+        return self.pad_id
+
+    @property
+    def mask(self):
+        return self.mask_id
+
+    @property
+    def bos_token(self):
+        """ Beginning of sentence token id """
+        return self._bos_token
+
+    @property
+    def eos_token(self):
+        """ End of sentence token id """
+        return self._eos_token
+
+    @property
+    def additional_special_tokens(self):
+        """ All the additional special tokens you may want to use (list of strings)."""
+        return self._additional_special_tokens
+
+    @property
+    def bos_token_id(self):
+        """ Id of the beginning of sentence token in the vocabulary."""
+        return self._bos_token_id
+
+    @property
+    def eos_token_id(self):
+        """ Id of the end of sentence token in the vocabulary."""
+        return self._eos_token_id
+
+    @property
+    def additional_special_tokens_ids(self):
+        """ Ids of all the additional special tokens in the vocabulary (list of integers)."""
+        return [self.vocab.get(token) for token in self._additional_special_tokens]
+
+    @additional_special_tokens.setter
+    def additional_special_tokens(self, value):
+        self._additional_special_tokens = value

examples/tutorial/sequence_parallel/loss_func/bert_loss.py

@@ -0,0 +1,41 @@
+import torch
+import torch.nn as nn
+from colossalai.core import global_context as gpc
+from colossalai.context import ParallelMode
+from colossalai.logging import get_dist_logger
+import torch.nn.functional as F
+import torch.distributed as dist
+from .cross_entropy import vocab_cross_entropy
+
+
+class BertLoss(nn.Module):
+
+    def forward(self,
+                lm_loss,
+                sop_logits,
+                loss_mask,
+                sentence_order):
+        lm_loss_ = lm_loss.float()
+        loss_mask = loss_mask.float()
+        loss_mask_sum = loss_mask.sum()
+        lm_loss = torch.sum(
+            lm_loss_.view(-1) * loss_mask.reshape(-1))
+
+        lm_loss /= loss_mask_sum
+
+        torch.distributed.all_reduce(
+            lm_loss,
+            group=gpc.get_group(ParallelMode.SEQUENCE)
+        )
+
+        if sop_logits is not None:
+            sop_loss = F.cross_entropy(sop_logits.view(-1, 2).float(),
+                                       sentence_order.view(-1),
+                                       ignore_index=-1)
+            sop_loss = sop_loss.float()
+            loss = lm_loss + sop_loss * gpc.get_world_size(ParallelMode.SEQUENCE)
+        else:
+            sop_loss = None
+            loss = lm_loss
+
+        return loss

examples/tutorial/sequence_parallel/loss_func/cross_entropy.py

@@ -0,0 +1,75 @@
+from colossalai.context.parallel_mode import ParallelMode
+import torch
+from torch.cuda.amp import custom_bwd, custom_fwd
+
+
+class _VocabCrossEntropy(torch.autograd.Function):
+
+    @staticmethod
+    @custom_fwd
+    def forward(ctx, vocab_parallel_logits, target):
+        # Maximum value along vocab dimension across all GPUs.
+        logits_max = torch.max(vocab_parallel_logits, dim=-1)[0]
+
+        # Subtract the maximum value.
+        vocab_parallel_logits.sub_(logits_max.unsqueeze(dim=-1))
+
+        # Create a mask of valid vocab ids (1 means it needs to be masked).
+        target_mask = target < 0
+        masked_target = target.clone()
+        masked_target[target_mask] = 0
+
+        # Get predicted-logits = logits[target].
+        # For Simplicity, we convert logits to a 2-D tensor with size
+        # [*, partition-vocab-size] and target to a 1-D tensor of size [*].
+        logits_2d = vocab_parallel_logits.view(-1, vocab_parallel_logits.size(-1))
+        masked_target_1d = masked_target.view(-1)
+        arange_1d = torch.arange(start=0, end=logits_2d.size()[0],
+                                 device=logits_2d.device)
+        predicted_logits_1d = logits_2d[arange_1d, masked_target_1d]
+        predicted_logits_1d = predicted_logits_1d.clone().contiguous()
+        predicted_logits = predicted_logits_1d.view_as(target)
+        predicted_logits[target_mask] = 0.0
+
+        # Sum of exponential of logits along vocab dimension across all GPUs.
+        exp_logits = vocab_parallel_logits
+        torch.exp(vocab_parallel_logits, out=exp_logits)
+        sum_exp_logits = exp_logits.sum(dim=-1)
+
+        # Loss = log(sum(exp(logits))) - predicted-logit.
+        loss = torch.log(sum_exp_logits) - predicted_logits
+
+        # Store softmax, target-mask and masked-target for backward pass.
+        exp_logits.div_(sum_exp_logits.unsqueeze(dim=-1))
+        ctx.save_for_backward(exp_logits, target_mask, masked_target_1d)
+
+        return loss
+
+    @staticmethod
+    @custom_bwd
+    def backward(ctx, grad_output):
+        # Retreive tensors from the forward path.
+        softmax, target_mask, masked_target_1d = ctx.saved_tensors
+
+        # All the inputs have softmax as their gradient.
+        grad_input = softmax
+        # For simplicity, work with the 2D gradient.
+        partition_vocab_size = softmax.size()[-1]
+        grad_2d = grad_input.view(-1, partition_vocab_size)
+
+        # Add the gradient from matching classes.
+        arange_1d = torch.arange(start=0, end=grad_2d.size()[0],
+                                 device=grad_2d.device)
+        grad_2d[arange_1d, masked_target_1d] -= (
+            1.0 - target_mask.view(-1).float())
+
+        # Finally elementwise multiplication with the output gradients.
+        grad_input.mul_(grad_output.unsqueeze(dim=-1))
+
+        return grad_input, None
+
+
+def vocab_cross_entropy(vocab_logits, target):
+    """helper function for the cross entropy."""
+
+    return _VocabCrossEntropy.apply(vocab_logits, target)

examples/tutorial/sequence_parallel/loss_func/utils.py

@@ -0,0 +1,55 @@
+
+import torch
+
+
+def ensure_divisibility(numerator, denominator):
+    """Ensure that numerator is divisible by the denominator."""
+    assert numerator % denominator == 0, '{} is not divisible by {}'.format(
+        numerator, denominator)
+
+
+def divide(numerator, denominator):
+    """Ensure that numerator is divisible by the denominator and return
+    the division value."""
+    ensure_divisibility(numerator, denominator)
+    return numerator // denominator
+
+
+def split_tensor_along_last_dim(tensor, num_partitions,
+                                contiguous_split_chunks=False):
+    """Split a tensor along its last dimension.
+    Arguments:
+        tensor: input tensor.
+        num_partitions: number of partitions to split the tensor
+        contiguous_split_chunks: If True, make each chunk contiguous
+                                 in memory.
+    """
+    # Get the size and dimension.
+    last_dim = tensor.dim() - 1
+    last_dim_size = divide(tensor.size()[last_dim], num_partitions)
+    # Split.
+    tensor_list = torch.split(tensor, last_dim_size, dim=last_dim)
+    # Note: torch.split does not create contiguous tensors by default.
+    if contiguous_split_chunks:
+        return tuple(chunk.contiguous() for chunk in tensor_list)
+
+    return tensor_list
+
+
+class VocabUtility:
+    """Split the vocabulary into `world_size` chunks amd return the
+        first and last index of the vocabulary belonging to the `rank`
+        partition: Note that indices in [fist, last)"""
+
+    @staticmethod
+    def vocab_range_from_per_partition_vocab_size(per_partition_vocab_size,
+                                                  rank, world_size):
+        index_f = rank * per_partition_vocab_size
+        index_l = index_f + per_partition_vocab_size
+        return index_f, index_l
+
+    @staticmethod
+    def vocab_range_from_global_vocab_size(global_vocab_size, rank, world_size):
+        per_partition_vocab_size = divide(global_vocab_size, world_size)
+        return VocabUtility.vocab_range_from_per_partition_vocab_size(
+            per_partition_vocab_size, rank, world_size)

examples/tutorial/sequence_parallel/lr_scheduler/__init__.py

@@ -0,0 +1 @@
+from .annealing_lr import AnnealingLR

examples/tutorial/sequence_parallel/lr_scheduler/annealing_lr.py

@@ -0,0 +1,158 @@
+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Learning rate decay functions."""
+
+import math
+
+
+class AnnealingLR(object):
+    """Anneals the learning rate."""
+
+    def __init__(self,
+                 optimizer,
+                 max_lr,
+                 min_lr,
+                 warmup_steps,
+                 decay_steps,
+                 decay_style,
+                 use_checkpoint_lr_scheduler=True,
+                 override_lr_scheduler=False):
+
+        # Class values.
+        self.optimizer = optimizer
+
+        self.max_lr = float(max_lr)
+        self.min_lr = min_lr
+        assert self.min_lr >= 0.0
+        assert self.max_lr >= self.min_lr
+
+        self.warmup_steps = warmup_steps
+        self.num_steps = 0
+        self.decay_steps = decay_steps
+        assert self.decay_steps > 0
+        assert self.warmup_steps < self.decay_steps
+
+        self.decay_style = decay_style
+
+        self.override_lr_scheduler = override_lr_scheduler
+        self.use_checkpoint_lr_scheduler = use_checkpoint_lr_scheduler
+        if self.override_lr_scheduler:
+            assert not self.use_checkpoint_lr_scheduler, 'both override and '\
+                'use-checkpoint are set.'
+
+        # Set the learning rate
+        self.step(0)
+
+    def get_lr(self):
+        """Learning rate decay functions from:
+              https://openreview.net/pdf?id=BJYwwY9ll pg. 4"""
+
+        # Use linear warmup for the initial part.
+        if self.warmup_steps > 0 and self.num_steps <= self.warmup_steps:
+            return self.max_lr * float(self.num_steps) / \
+                float(self.warmup_steps)
+
+        # If the learning rate is constant, just return the initial value.
+        if self.decay_style == 'constant':
+            return self.max_lr
+
+        # For any steps larger than `self.decay_steps`, use `self.min_lr`.
+        if self.num_steps > self.decay_steps:
+            return self.min_lr
+
+        # If we are done with the warmup period, use the decay style.
+        num_steps_ = self.num_steps - self.warmup_steps
+        decay_steps_ = self.decay_steps - self.warmup_steps
+        decay_ratio = float(num_steps_) / float(decay_steps_)
+        assert decay_ratio >= 0.0
+        assert decay_ratio <= 1.0
+        delta_lr = self.max_lr - self.min_lr
+
+        if self.decay_style == 'linear':
+            coeff = (1.0 - decay_ratio)
+        elif self.decay_style == 'cosine':
+            coeff = 0.5 * (math.cos(math.pi * decay_ratio) + 1.0)
+        else:
+            raise Exception('{} decay style is not supported.'.format(
+                self.decay_style))
+
+        return self.min_lr + coeff * delta_lr
+
+    def step(self, increment=1):
+        """Set lr for all parameters groups."""
+        self.num_steps += increment
+        new_lr = self.get_lr()
+        for group in self.optimizer.param_groups:
+            group['lr'] = new_lr
+
+    def state_dict(self):
+        state_dict = {
+            'max_lr': self.max_lr,
+            'warmup_steps': self.warmup_steps,
+            'num_steps': self.num_steps,
+            'decay_style': self.decay_style,
+            'decay_steps': self.decay_steps,
+            'min_lr': self.min_lr
+        }
+        return state_dict
+
+    def _check_and_set(self, cls_value, sd_value, name):
+        """Auxiliary function for checking the values in the checkpoint and
+        setting them."""
+        if self.override_lr_scheduler:
+            return cls_value
+
+        if not self.use_checkpoint_lr_scheduler:
+            assert cls_value == sd_value, \
+                f'AnnealingLR: class input value {cls_value} and checkpoint' \
+                f'value {sd_value} for {name} do not match'
+        return sd_value
+
+    def load_state_dict(self, sd):
+
+        if 'start_lr' in sd:
+            max_lr_ = sd['start_lr']
+        else:
+            max_lr_ = sd['max_lr']
+        self.max_lr = self._check_and_set(self.max_lr, max_lr_,
+                                          'learning rate')
+
+        self.min_lr = self._check_and_set(self.min_lr, sd['min_lr'],
+                                          'minimum learning rate')
+
+        if 'warmup_iter' in sd:
+            warmup_steps_ = sd['warmup_iter']
+        else:
+            warmup_steps_ = sd['warmup_steps']
+        self.warmup_steps = self._check_and_set(self.warmup_steps,
+                                                warmup_steps_,
+                                                'warmup iterations')
+
+        if 'end_iter' in sd:
+            decay_steps_ = sd['end_iter']
+        else:
+            decay_steps_ = sd['decay_steps']
+        self.decay_steps = self._check_and_set(self.decay_steps, decay_steps_,
+                                               'total number of iterations')
+        self.decay_style = self._check_and_set(self.decay_style,
+                                               sd['decay_style'],
+                                               'decay style')
+
+        if 'num_iters' in sd:
+            num_steps = sd['num_iters']
+        else:
+            num_steps = sd['num_steps']
+        self.step(increment=num_steps)

examples/tutorial/sequence_parallel/model/__init__.py

@@ -0,0 +1,2 @@
+
+

examples/tutorial/sequence_parallel/model/bert.py

@@ -0,0 +1,282 @@
+from colossalai.context.parallel_mode import ParallelMode
+import torch
+import torch.nn as nn
+import inspect
+from .layers import Embedding, BertLayer, BertDualHead, PreProcessor, VocabEmbedding
+from .layers.init_method import init_normal, output_init_normal
+from colossalai.core import global_context as gpc
+from colossalai.context import ParallelMode
+from colossalai.kernel import LayerNorm
+from colossalai.nn.layer.wrapper import PipelineSharedModuleWrapper
+from colossalai.logging import get_dist_logger
+from colossalai.pipeline.utils import partition_uniform
+
+
+class BertForPretrain(nn.Module):
+
+    def __init__(self,
+                 vocab_size,
+                 hidden_size,
+                 max_sequence_length,
+                 num_attention_heads,
+                 num_layers,
+                 add_binary_head,
+                 is_naive_fp16,
+                 num_tokentypes=2,
+                 dropout_prob=0.1,
+                 mlp_ratio=4,
+                 init_std=0.02,
+                 convert_fp16_to_fp32_in_softmax=False,
+                 ):
+        super().__init__()
+        self.seq_parallel_size = gpc.get_world_size(ParallelMode.SEQUENCE)
+        assert max_sequence_length % self.seq_parallel_size == 0, 'sequence length is not divisible by the sequence parallel size'
+        self.sub_seq_length = max_sequence_length // self.seq_parallel_size
+        self.init_std = init_std
+        self.num_layers = num_layers
+
+        if not add_binary_head:
+            num_tokentypes = 0
+
+        self.preprocessor = PreProcessor(self.sub_seq_length)
+        self.embedding = Embedding(hidden_size=hidden_size,
+                                   vocab_size=vocab_size,
+                                   max_sequence_length=max_sequence_length,
+                                   embedding_dropout_prob=dropout_prob,
+                                   num_tokentypes=num_tokentypes)
+        self.bert_layers = nn.ModuleList()
+
+        for i in range(num_layers):
+            bert_layer = BertLayer(layer_number=i+1,
+                                   hidden_size=hidden_size,
+                                   num_attention_heads=num_attention_heads,
+                                   attention_dropout=dropout_prob,
+                                   mlp_ratio=mlp_ratio,
+                                   hidden_dropout=dropout_prob,
+                                   convert_fp16_to_fp32_in_softmax=convert_fp16_to_fp32_in_softmax,
+                                   is_naive_fp16=is_naive_fp16
+                                   )
+            self.bert_layers.append(bert_layer)
+
+        self.layer_norm = LayerNorm(hidden_size)
+        self.head = BertDualHead(hidden_size, self.embedding.word_embedding_weight.size(0),
+                                 add_binary_head=add_binary_head)
+        self.reset_parameters()
+
+    def _init_normal(self, tensor):
+        init_normal(tensor, sigma=self.init_std)
+
+    def _output_init_normal(self, tensor):
+        output_init_normal(tensor, sigma=self.init_std, num_layers=self.num_layers)
+
+    def reset_parameters(self):
+        # initialize embedding
+        self._init_normal(self.embedding.word_embedding_weight)
+        self._init_normal(self.embedding.position_embeddings.weight)
+        if self.embedding.tokentype_embeddings:
+            self._init_normal(self.embedding.tokentype_embeddings.weight)
+
+        # initialize bert layer
+        for layer in self.bert_layers:
+            # initialize self attention
+            self._init_normal(layer.self_attention.query_key_value.weight)
+            self._output_init_normal(layer.self_attention.dense.weight)
+            self._init_normal(layer.mlp.dense_h_to_4h.weight)
+            self._output_init_normal(layer.mlp.dense_4h_to_h.weight)
+
+        # initializer head
+        self._init_normal(self.head.lm_head.dense.weight)
+        if self.head.binary_head is not None:
+            self._init_normal(self.head.binary_head.pooler.dense.weight)
+            self._init_normal(self.head.binary_head.dense.weight)
+
+    def forward(self, input_ids, attention_masks, tokentype_ids, lm_labels):
+        # inputs of the forward function
+        # input_ids: [batch_size, sub_seq_len]
+        # attention_mask: [batch_size, seq_len]
+        # tokentype_ids: [batch_size, sub_seq_len]
+        # outputs of preprocessor
+        # pos_ids: [batch_size, sub_seq_len]
+        # attention_masks: [batch_size, 1, sub_seq_len, seq_len]
+        pos_ids, attention_masks = self.preprocessor(input_ids, attention_masks)
+
+        hidden_states = self.embedding(input_ids, pos_ids, tokentype_ids)
+
+        # hidden_states shape change:
+        # [batch_size, sub_seq_len, hidden_size] -> [sub_seq_len, batch_size, hidden_size]
+        hidden_states = hidden_states.transpose(0, 1).contiguous()
+
+        for idx, layer in enumerate(self.bert_layers):
+            hidden_states = layer(hidden_states, attention_masks)
+
+        hidden_states = hidden_states.transpose(0, 1).contiguous()
+        output = self.layer_norm(hidden_states)
+
+        # hidden_states: [sub_seq_len, batch_size, hidden_size]
+        # word_embedding: [vocab_size, hidden_size]
+        return self.head(output, self.embedding.word_embedding_weight, lm_labels)
+
+
+class PipelineBertForPretrain(nn.Module):
+
+    def __init__(self,
+                 vocab_size,
+                 hidden_size,
+                 max_sequence_length,
+                 num_attention_heads,
+                 num_layers,
+                 add_binary_head,
+                 is_naive_fp16,
+                 num_tokentypes=2,
+                 dropout_prob=0.1,
+                 mlp_ratio=4,
+                 init_std=0.02,
+                 convert_fp16_to_fp32_in_softmax=False,
+                 first_stage=True,
+                 last_stage=True,
+                 start_idx=None,
+                 end_idx=None):
+        super().__init__()
+        self.seq_parallel_size = gpc.get_world_size(ParallelMode.SEQUENCE)
+        assert max_sequence_length % self.seq_parallel_size == 0, 'sequence length is not divisible by the sequence parallel size'
+        self.sub_seq_length = max_sequence_length // self.seq_parallel_size
+        self.init_std = init_std
+        self.num_layers = num_layers
+
+        if not add_binary_head:
+            num_tokentypes = 0
+
+        self.first_stage = first_stage
+        self.last_stage = last_stage
+
+        self.preprocessor = PreProcessor(self.sub_seq_length)
+
+        if self.first_stage:
+            self.embedding = Embedding(hidden_size=hidden_size,
+                                       vocab_size=vocab_size,
+                                       max_sequence_length=max_sequence_length,
+                                       embedding_dropout_prob=dropout_prob,
+                                       num_tokentypes=num_tokentypes)
+
+        # transformer layers
+        self.bert_layers = nn.ModuleList()
+
+        if start_idx is None and end_idx is None:
+            start_idx = 0
+            end_idx = num_layers
+
+        for i in range(start_idx, end_idx):
+            bert_layer = BertLayer(layer_number=i+1,
+                                   hidden_size=hidden_size,
+                                   num_attention_heads=num_attention_heads,
+                                   attention_dropout=dropout_prob,
+                                   mlp_ratio=mlp_ratio,
+                                   hidden_dropout=dropout_prob,
+                                   convert_fp16_to_fp32_in_softmax=convert_fp16_to_fp32_in_softmax,
+                                   is_naive_fp16=is_naive_fp16
+                                   )
+            self.bert_layers.append(bert_layer)
+
+        if self.last_stage:
+            self.word_embeddings = VocabEmbedding(vocab_size, hidden_size)
+            self.layer_norm = LayerNorm(hidden_size)
+            self.head = BertDualHead(hidden_size, vocab_size,
+                                     add_binary_head=add_binary_head)
+        self.reset_parameters()
+
+    def _init_normal(self, tensor):
+        init_normal(tensor, sigma=self.init_std)
+
+    def _output_init_normal(self, tensor):
+        output_init_normal(tensor, sigma=self.init_std, num_layers=self.num_layers)
+
+    def reset_parameters(self):
+        # initialize embedding
+        if self.first_stage:
+            self._init_normal(self.embedding.word_embedding_weight)
+            self._init_normal(self.embedding.position_embeddings.weight)
+            if self.embedding.tokentype_embeddings:
+                self._init_normal(self.embedding.tokentype_embeddings.weight)
+
+        # initialize bert layer
+        for layer in self.bert_layers:
+            # initialize self attention
+            self._init_normal(layer.self_attention.query_key_value.weight)
+            self._output_init_normal(layer.self_attention.dense.weight)
+            self._init_normal(layer.mlp.dense_h_to_4h.weight)
+            self._output_init_normal(layer.mlp.dense_4h_to_h.weight)
+
+        # initializer head
+        if self.last_stage:
+            self._init_normal(self.head.lm_head.dense.weight)
+            if self.head.binary_head is not None:
+                self._init_normal(self.head.binary_head.pooler.dense.weight)
+                self._init_normal(self.head.binary_head.dense.weight)
+
+    def forward(self, input_ids, attention_masks, tokentype_ids, lm_labels):
+        # inputs of the forward function
+        # input_ids: [batch_size, sub_seq_len]
+        # attention_mask: [batch_size, seq_len]
+        # tokentype_ids: [batch_size, sub_seq_len]
+        # outputs of preprocessor
+        # pos_ids: [batch_size, sub_seq_len]
+        # attention_masks: [batch_size, 1, sub_seq_len, seq_len]
+        if self.first_stage:
+            pos_ids, attention_masks = self.preprocessor(input_ids, attention_masks)
+        else:
+            _, attention_masks = self.preprocessor(None, attention_masks)
+
+        if self.first_stage:
+            hidden_states = self.embedding(input_ids, pos_ids, tokentype_ids)
+            hidden_states = hidden_states.transpose(0, 1).contiguous()
+        else:
+            hidden_states = input_ids
+
+        # hidden_states shape change:
+        # [batch_size, sub_seq_len, hidden_size] -> [sub_seq_len, batch_size, hidden_size]
+        for idx, layer in enumerate(self.bert_layers):
+            hidden_states = layer(hidden_states, attention_masks)
+
+        if self.last_stage:
+            hidden_states = hidden_states.transpose(0, 1).contiguous()
+            output = self.layer_norm(hidden_states)
+            output = self.head(output, self.word_embeddings.weight, lm_labels)
+        else:
+            output = hidden_states
+
+        # hidden_states: [sub_seq_len, batch_size, hidden_size]
+        # word_embedding: [vocab_size, hidden_size]
+        return output
+
+
+def _filter_kwargs(func, kwargs):
+    sig = inspect.signature(func)
+    return {k: v for k, v in kwargs.items() if k in sig.parameters}
+
+
+def build_pipeline_bert(num_layers, num_chunks, device=torch.device('cuda'), **kwargs):
+    logger = get_dist_logger()
+    pipeline_size = gpc.get_world_size(ParallelMode.PIPELINE)
+    pipeline_rank = gpc.get_local_rank(ParallelMode.PIPELINE)
+    rank = gpc.get_global_rank()
+    wrapper = PipelineSharedModuleWrapper([0, pipeline_size - 1])
+    parts = partition_uniform(num_layers, pipeline_size, num_chunks)[pipeline_rank]
+    models = []
+    for start, end in parts:
+        kwargs['num_layers'] = num_layers
+        kwargs['start_idx'] = start
+        kwargs['end_idx'] = end
+        kwargs['first_stage'] = start == 0
+        kwargs['last_stage'] = end == num_layers
+        logger.info(f'Rank{rank} build layer {start}-{end}, {end-start}/{num_layers} layers')
+        chunk = PipelineBertForPretrain(**_filter_kwargs(PipelineBertForPretrain.__init__, kwargs)).to(device)
+        if start == 0:
+            wrapper.register_module(chunk.embedding.word_embeddings)
+        elif end == num_layers:
+            wrapper.register_module(chunk.word_embeddings)
+        models.append(chunk)
+    if len(models) == 1:
+        model = models[0]
+    else:
+        model = nn.ModuleList(models)
+    return model

examples/tutorial/sequence_parallel/model/layers/__init__.py

@@ -0,0 +1,4 @@
+from .embedding import VocabEmbedding, Embedding
+from .bert_layer import BertLayer
+from .head import BertDualHead
+from .preprocess import PreProcessor

examples/tutorial/sequence_parallel/model/layers/bert_layer.py

@@ -0,0 +1,118 @@
+import torch
+import torch.nn as nn
+from colossalai.nn.layer.parallel_sequence import TransformerSelfAttentionRing
+from colossalai.kernel.jit import bias_dropout_add_fused_train, bias_dropout_add_fused_inference
+from colossalai.kernel.cuda_native import LayerNorm
+from .mlp import TransformerMLP
+from .dropout import get_bias_dropout_add
+
+
+def attention_mask_func(attention_scores, attention_mask):
+    attention_scores.masked_fill_(attention_mask, -10000.0)
+    return attention_scores
+
+
+class BertLayer(nn.Module):
+    """A single transformer layer.
+    Transformer layer takes input with size [b, s, h] and returns an
+    output of the same size.
+    """
+
+    def __init__(self,
+                 layer_number,
+                 hidden_size,
+                 num_attention_heads,
+                 attention_dropout,
+                 mlp_ratio,
+                 hidden_dropout,
+                 is_naive_fp16,
+                 apply_residual_connection_post_layernorm=False,
+                 fp32_residual_connection=False,
+                 bias_dropout_fusion: bool = True,
+                 convert_fp16_to_fp32_in_softmax: bool = False):
+        super().__init__()
+        self.layer_number = layer_number
+
+        self.apply_residual_connection_post_layernorm = apply_residual_connection_post_layernorm
+        self.fp32_residual_connection = fp32_residual_connection
+
+        # Layernorm on the input data.
+        self.input_layernorm = LayerNorm(hidden_size)
+
+        # Self attention.
+        self.self_attention = TransformerSelfAttentionRing(
+            hidden_size=hidden_size,
+            num_attention_heads=num_attention_heads,
+            attention_dropout=attention_dropout,
+            attention_mask_func=attention_mask_func,
+            layer_number=layer_number,
+            apply_query_key_layer_scaling=True,
+            convert_fp16_to_fp32_in_softmax=convert_fp16_to_fp32_in_softmax,
+            fp16=is_naive_fp16
+        )
+
+        self.hidden_dropout = hidden_dropout
+        self.bias_dropout_fusion = bias_dropout_fusion
+
+        # Layernorm on the attention output
+        self.post_attention_layernorm = LayerNorm(hidden_size)
+
+        self.mlp = TransformerMLP(hidden_size=hidden_size, mlp_ratio=mlp_ratio)
+
+    def forward(self, hidden_states, attention_mask):
+        # hidden_states: [batch_size, sub_seq_len, hidden_size]
+        # attention_mask: [batch_size, 1, sub_seq_len, seq_len]
+
+        # Layer norm at the beginning of the transformer layer.
+        layernorm_output = self.input_layernorm(hidden_states)
+
+        # Self attention.
+        attention_output, attention_bias = self.self_attention(layernorm_output, attention_mask)
+
+        # Residual connection.
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = hidden_states
+
+        # jit scripting for a nn.module (with dropout) is not
+        # trigerring the fusion kernel. For now, we use two
+        # different nn.functional routines to account for varying
+        # dropout semantics during training and inference phases.
+        if self.bias_dropout_fusion:
+            if self.training:
+                bias_dropout_add_func = bias_dropout_add_fused_train
+            else:
+                bias_dropout_add_func = bias_dropout_add_fused_inference
+        else:
+            bias_dropout_add_func = get_bias_dropout_add(self.training)
+
+        # re-enable torch grad to enable fused optimization.
+        with torch.enable_grad():
+            layernorm_input = bias_dropout_add_func(
+                attention_output,
+                attention_bias.expand_as(residual),
+                residual,
+                self.hidden_dropout)
+
+        # Layer norm post the self attention.
+        layernorm_output = self.post_attention_layernorm(layernorm_input)
+
+        # MLP.
+        mlp_output, mlp_bias = self.mlp(layernorm_output)
+
+        # Second residual connection.
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = layernorm_input
+
+        # re-enable torch grad to enable fused optimization.
+        with torch.enable_grad():
+            output = bias_dropout_add_func(
+                mlp_output,
+                mlp_bias.expand_as(residual),
+                residual,
+                self.hidden_dropout)
+
+        return output

examples/tutorial/sequence_parallel/model/layers/dropout.py

@@ -0,0 +1,13 @@
+import torch
+
+def bias_dropout_add(x, bias, residual, prob, training):
+    # type: (Tensor, Tensor, Tensor, float, bool) -> Tensor
+    out = torch.nn.functional.dropout(x + bias, p=prob, training=training)
+    out = residual + out
+    return out
+
+
+def get_bias_dropout_add(training):
+    def _bias_dropout_add(x, bias, residual, prob):
+        return bias_dropout_add(x, bias, residual, prob, training)
+    return _bias_dropout_add

examples/tutorial/sequence_parallel/model/layers/embedding.py

@@ -0,0 +1,96 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.nn.init as init
+
+
+class VocabEmbedding(torch.nn.Module):
+
+    def __init__(self, num_embeddings, embedding_dim):
+        super(VocabEmbedding, self).__init__()
+        # Keep the input dimensions.
+        self.num_embeddings = num_embeddings
+        self.embedding_dim = embedding_dim
+        self.padding_idx = None
+        self.max_norm = None
+        self.norm_type = 2.
+        self.scale_grad_by_freq = False
+        self.sparse = False
+        self._weight = None
+
+        # Allocate weights and initialize.
+        self.weight = nn.Parameter(torch.empty(
+            self.num_embeddings, self.embedding_dim))
+        init.xavier_uniform_(self.weight)
+
+    def forward(self, hidden_state):
+        output = F.embedding(hidden_state, self.weight,
+                             self.padding_idx, self.max_norm,
+                             self.norm_type, self.scale_grad_by_freq,
+                             self.sparse)
+        return output
+
+    def __repr__(self):
+        return f'VocabEmbedding(num_embeddings={self.num_embeddings}, ' \
+            f'embedding_dim={self.embedding_dim})'
+
+
+class Embedding(nn.Module):
+    """Language model embeddings.
+    Arguments:
+        hidden_size: hidden size
+        vocab_size: vocabulary size
+        max_sequence_length: maximum size of sequence. This
+                             is used for positional embedding
+        embedding_dropout_prob: dropout probability for embeddings
+        init_method: weight initialization method
+        num_tokentypes: size of the token-type embeddings. 0 value
+                        will ignore this embedding
+    """
+
+    def __init__(self,
+                 hidden_size,
+                 vocab_size,
+                 max_sequence_length,
+                 embedding_dropout_prob,
+                 num_tokentypes):
+        super(Embedding, self).__init__()
+
+        self.hidden_size = hidden_size
+        self.num_tokentypes = num_tokentypes
+
+        self.word_embeddings = VocabEmbedding(vocab_size, self.hidden_size)
+
+        # Position embedding (serial).
+        self.position_embeddings = torch.nn.Embedding(
+            max_sequence_length, self.hidden_size)
+
+        # Token type embedding.
+        # Add this as an optional field that can be added through
+        # method call so we can load a pretrain model without
+        # token types and add them as needed.
+        if self.num_tokentypes > 0:
+            self.tokentype_embeddings = torch.nn.Embedding(self.num_tokentypes,
+                                                           self.hidden_size)
+        else:
+            self.tokentype_embeddings = None
+
+        # Embeddings dropout
+        self.embedding_dropout = torch.nn.Dropout(embedding_dropout_prob)
+
+    @property
+    def word_embedding_weight(self):
+        return self.word_embeddings.weight
+
+    def forward(self, input_ids, position_ids, tokentype_ids=None):
+        # Embeddings.
+        words_embeddings = self.word_embeddings(input_ids)
+        position_embeddings = self.position_embeddings(position_ids)
+        embeddings = words_embeddings + position_embeddings
+        if tokentype_ids is not None and self.tokentype_embeddings is not None:
+            embeddings = embeddings + self.tokentype_embeddings(tokentype_ids)
+
+        # Dropout.
+        embeddings = self.embedding_dropout(embeddings)
+
+        return embeddings

examples/tutorial/sequence_parallel/model/layers/head.py

@@ -0,0 +1,78 @@
+import colossalai
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from .pooler import Pooler
+from .linear import Linear
+from .embedding import VocabEmbedding
+from colossalai.core import global_context as gpc
+from colossalai.context import ParallelMode
+from colossalai.kernel import LayerNorm
+from loss_func.cross_entropy import vocab_cross_entropy
+
+
+class BertLMHead(nn.Module):
+    """Masked LM head for Bert
+    Arguments:
+        hidden_size: hidden size
+        init_method: init method for weight initialization
+        layernorm_epsilon: tolerance for layer norm divisions
+    """
+
+    def __init__(self,
+                 vocab_size,
+                 hidden_size,
+                 ):
+
+        super(BertLMHead, self).__init__()
+        self.bias = torch.nn.Parameter(torch.zeros(vocab_size))
+
+        self.dense = Linear(hidden_size, hidden_size)
+        self.layernorm = LayerNorm(hidden_size)
+        self.gelu = torch.nn.functional.gelu
+
+    def forward(self, hidden_states, word_embeddings_weight, lm_labels):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.gelu(hidden_states)
+        hidden_states = self.layernorm(hidden_states)
+
+        output = F.linear(hidden_states, word_embeddings_weight, self.bias)
+        lm_loss = vocab_cross_entropy(output, lm_labels)
+
+        return lm_loss
+
+
+class BertBinaryHead(nn.Module):
+
+    def __init__(self, hidden_size):
+        super().__init__()
+        self.pooler = Pooler(hidden_size)
+        self.dense = Linear(hidden_size, 2)
+
+    def forward(self, hidden_states):
+        if gpc.get_local_rank(ParallelMode.SEQUENCE) == 0:
+            output = self.pooler(hidden_states)
+            output = self.dense(output)
+        else:
+            output = None
+        return output
+
+
+class BertDualHead(nn.Module):
+
+    def __init__(self, hidden_size, vocab_size, add_binary_head):
+        super().__init__()
+        self.lm_head = BertLMHead(vocab_size, hidden_size)
+        self.add_binary_head = add_binary_head
+        if add_binary_head:
+            self.binary_head = BertBinaryHead(hidden_size)
+        else:
+            self.binary_head = None
+
+    def forward(self, hidden_states, word_embeddings_weight, lm_labels):
+        if self.add_binary_head:
+            binary_output = self.binary_head(hidden_states)
+        else:
+            binary_output = None
+        lm_loss = self.lm_head(hidden_states, word_embeddings_weight, lm_labels)
+        return lm_loss, binary_output

examples/tutorial/sequence_parallel/model/layers/init_method.py

@@ -0,0 +1,12 @@
+import torch
+import math
+
+def init_normal(tensor, sigma):
+    """Init method based on N(0, sigma)."""
+    torch.nn.init.normal_(tensor, mean=0.0, std=sigma)
+
+
+def output_init_normal(tensor, sigma, num_layers):
+    """Init method based on N(0, sigma/sqrt(2*num_layers)."""
+    std = sigma / math.sqrt(2.0 * num_layers)
+    torch.nn.init.normal_(tensor, mean=0.0, std=std)

examples/tutorial/sequence_parallel/model/layers/linear.py

@@ -0,0 +1,63 @@
+import torch
+import torch.nn as nn
+from torch.nn import Parameter
+import torch.nn.functional as F
+import torch.nn.init as init
+
+
+class Linear(nn.Module):
+    """Linear layer with column parallelism.
+    The linear layer is defined as Y = XA + b. A is parallelized along
+    its second dimension as A = [A_1, ..., A_p].
+    Arguments:
+        input_size: first dimension of matrix A.
+        output_size: second dimension of matrix A.
+        bias: If true, add bias
+        init_method: method to initialize weights. Note that bias is always set
+                     to zero.
+        stride: For the strided linear layers.
+        keep_master_weight_for_test: This was added for testing and should be
+                                     set to False. It returns the master weights
+                                     used for initialization.
+        skip_bias_add: This was added to enable performance optimations where bias
+                       can be fused with other elementwise operations. we skip
+                       adding bias but instead return it.
+    """
+
+    def __init__(self,
+                 input_size,
+                 output_size,
+                 bias=True,
+                 skip_bias_add=False):
+        super(Linear, self).__init__()
+
+        # Keep input parameters
+        self.input_size = input_size
+        self.output_size = output_size
+        self.skip_bias_add = skip_bias_add
+
+        self.weight = Parameter(torch.empty(self.output_size,
+                                            self.input_size,
+                                            ))
+        init.normal_(self.weight)
+        if bias:
+            self.bias = Parameter(torch.empty(self.output_size))
+            # Always initialize bias to zero.
+            with torch.no_grad():
+                self.bias.zero_()
+        else:
+            self.register_parameter('bias', None)
+
+    def forward(self, input_):
+        # Matrix multiply.
+        bias = self.bias if not self.skip_bias_add else None
+        output = F.linear(input_, self.weight, bias)
+
+        if self.skip_bias_add:
+            return output, self.bias
+        else:
+            return output
+
+    def __repr__(self):
+        return f'Linear(in_features={self.input_size}, out_features={self.output_size}, ' + \
+            f'bias={self.bias is not None}, skip_bias_add={self.skip_bias_add})'

examples/tutorial/sequence_parallel/model/layers/mlp.py

@@ -0,0 +1,50 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from .linear import Linear
+from colossalai.kernel.jit import bias_gelu_impl
+
+
+class TransformerMLP(nn.Module):
+    """MLP.
+    MLP will take the input with h hidden state, project it to 4*h
+    hidden dimension, perform nonlinear transformation, and project the
+    state back into h hidden dimension. At the end, dropout is also
+    applied.
+    """
+
+    def __init__(self, hidden_size, mlp_ratio, fuse_gelu=True):
+        super(TransformerMLP, self).__init__()
+
+        # Project to 4h.
+        self.dense_h_to_4h = Linear(
+            hidden_size,
+            int(hidden_size*mlp_ratio),
+            skip_bias_add=True)
+
+        self.bias_gelu_fusion = fuse_gelu
+        self.activation_func = F.gelu
+
+        # Project back to h.
+        self.dense_4h_to_h = Linear(
+            int(hidden_size*mlp_ratio),
+            hidden_size,
+            skip_bias_add=True)
+
+    def forward(self, hidden_states):
+        # hidden states should be in the shape of [s, b, h]
+        # it will be projects into [s, b, 4h]
+        # and projected back to [s, b, h]
+        intermediate_parallel, bias_parallel = self.dense_h_to_4h(hidden_states)
+
+        if self.bias_gelu_fusion:
+            intermediate_parallel = \
+                bias_gelu_impl(intermediate_parallel, bias_parallel)
+        else:
+            intermediate_parallel = \
+                self.activation_func(intermediate_parallel + bias_parallel)
+
+        # [s, b, h]
+        output, output_bias = self.dense_4h_to_h(intermediate_parallel)
+        return output, output_bias

examples/tutorial/sequence_parallel/model/layers/pooler.py

@@ -0,0 +1,28 @@
+import torch
+import torch.nn as nn
+from .linear import Linear
+
+
+class Pooler(nn.Module):
+    """Pooler layer.
+
+    Pool hidden states of a specific token (for example start of the
+    sequence) and add a linear transformation followed by a tanh.
+
+    Arguments:
+        hidden_size: hidden size
+        init_method: weight initialization method for the linear layer.
+            bias is set to zero.
+    """
+
+    def __init__(self, hidden_size):
+        super(Pooler, self).__init__()
+        self.dense = Linear(hidden_size, hidden_size)
+
+    def forward(self, hidden_states, sequence_index=0):
+        # hidden_states: [b, s, h]
+        # sequence_index: index of the token to pool.
+        pooled = hidden_states[:, sequence_index, :]
+        pooled = self.dense(pooled)
+        pooled = torch.tanh(pooled)
+        return pooled

examples/tutorial/sequence_parallel/model/layers/preprocess.py

@@ -0,0 +1,58 @@
+from colossalai.context.parallel_mode import ParallelMode
+import torch
+import torch.nn as nn
+from colossalai.core import global_context as gpc
+
+
+class PreProcessor(nn.Module):
+
+    def __init__(self, sub_seq_length):
+        super().__init__()
+        self.sub_seq_length = sub_seq_length
+
+    def bert_position_ids(self, token_ids):
+        # Create position ids
+        seq_length = token_ids.size(1)
+        local_rank = gpc.get_local_rank(ParallelMode.SEQUENCE)
+        position_ids = torch.arange(seq_length*local_rank,
+                                    seq_length * (local_rank+1),
+                                    dtype=torch.long,
+                                    device=token_ids.device)
+        position_ids = position_ids.unsqueeze(0).expand_as(token_ids)
+
+        return position_ids
+
+    def bert_extended_attention_mask(self, attention_mask):
+        local_rank = gpc.get_local_rank(ParallelMode.SEQUENCE)
+        start_index = local_rank * self.sub_seq_length
+        end_index = (local_rank + 1) * self.sub_seq_length
+
+        # We create a 3D attention mask from a 2D tensor mask.
+        # [b, 1, s]
+        attention_mask_b1s = attention_mask.unsqueeze(1)
+        # [b, s, 1]
+        attention_mask_bs1 = attention_mask.unsqueeze(2)
+        # [b, s/D, s]
+        attention_mask_bss = attention_mask_b1s * attention_mask_bs1
+
+        attention_mask_bss = attention_mask_bss[:, start_index:end_index, :]
+
+        # [b, 1, s/D, s]
+        extended_attention_mask = attention_mask_bss.unsqueeze(1)
+
+        # Convert attention mask to binary:
+        extended_attention_mask = (extended_attention_mask < 0.5)
+
+        return extended_attention_mask
+
+    def forward(self, input_ids=None, attention_mask=None):
+        if attention_mask is not None:
+            extended_attention_mask = self.bert_extended_attention_mask(attention_mask)
+        else:
+            extended_attention_mask = None
+
+        if input_ids is not None:
+            position_ids = self.bert_position_ids(input_ids)
+        else:
+            position_ids = None
+        return position_ids, extended_attention_mask

examples/tutorial/sequence_parallel/train.py

@@ -0,0 +1,210 @@
+import colossalai
+from colossalai.context.parallel_mode import ParallelMode
+from colossalai.core import global_context as gpc
+from data import build_train_valid_test_data_iterators
+from data.tokenizer import initialize_tokenizer, get_padded_vocab_size
+from data.bert_helper import get_batch_for_sequence_parallel, SequenceParallelDataIterator
+from colossalai.amp import AMP_TYPE
+from colossalai.logging import get_dist_logger
+from colossalai.utils import MultiTimer, is_using_pp
+from model.bert import BertForPretrain
+from lr_scheduler import AnnealingLR
+from loss_func.bert_loss import BertLoss
+import torch
+from colossalai.engine.schedule import PipelineSchedule
+from colossalai.amp import AMP_TYPE
+from colossalai.nn.optimizer import FusedAdam
+from colossalai.kernel import LayerNorm
+from model.bert import build_pipeline_bert
+
+
+def process_batch_data(batch_data):
+    tokens, types, sentence_order, loss_mask, lm_labels, padding_mask = batch_data
+    if gpc.is_first_rank(ParallelMode.PIPELINE):
+        data = dict(input_ids=tokens, attention_masks=padding_mask, tokentype_ids=types, lm_labels=lm_labels)
+    else:
+        data = dict(attention_masks=padding_mask, tokentype_ids=types, lm_labels=lm_labels)
+    label = dict(loss_mask=loss_mask, sentence_order=sentence_order)
+    return data, label
+
+
+def main():
+    # initialize
+    colossalai.launch_from_torch(config='./config.py', seed=1234, backend='nccl')
+
+    logger = get_dist_logger()
+
+    # build dataloader
+    initialize_tokenizer(gpc.config.VOCAB_FILE_PATH, tokenizer_type='BertWordPieceLowerCase')
+    VOCAB_SIZE = get_padded_vocab_size()
+    trainloader, validloader, testloader = build_train_valid_test_data_iterators(
+        train_iters=gpc.config.TRAIN_ITERS,
+        global_batch_size=gpc.config.GLOBAL_BATCH_SIZE,
+        eval_interval=gpc.config.EVAL_INTERVAL,
+        eval_iters=gpc.config.EVAL_ITERS,
+        data_prefix=[gpc.config.DATA_PATH],
+        data_impl='mmap',
+        splits_string='949,50,1',
+        max_seq_length=gpc.config.SEQ_LENGTH,
+        masked_lm_prob=0.15,
+        short_seq_prob=0.1,
+        seed=1234,
+        skip_warmup=True,
+        binary_head=False,
+    )
+
+    logger.info("Dataloaders are built", ranks=[0])
+
+    # build model
+    if hasattr(gpc.config, 'fp16') and gpc.config.fp16.get('mode') == AMP_TYPE.NAIVE:
+        is_naive_fp16 = True
+    else:
+        is_naive_fp16 = False
+
+    use_pipeline = is_using_pp()
+    kwargs = dict(vocab_size=VOCAB_SIZE,
+                  hidden_size=gpc.config.HIDDEN_SIZE,
+                  max_sequence_length=gpc.config.SEQ_LENGTH,
+                  num_attention_heads=gpc.config.NUM_ATTENTION_HEADS,
+                  convert_fp16_to_fp32_in_softmax=True,
+                  is_naive_fp16=is_naive_fp16,
+                  add_binary_head=gpc.config.ADD_BINARY_HEAD)
+
+    if use_pipeline:
+        model = build_pipeline_bert(num_layers=gpc.config.DEPTH, num_chunks=1, **kwargs)
+    else:
+        model = BertForPretrain(num_layers=gpc.config.DEPTH, **kwargs)
+
+    model = model.half()
+    model.reset_parameters()
+    logger.info(f"Model is built with softmax in fp32 = {is_naive_fp16}", ranks=[0])
+
+    total_numel = 0
+    for p in model.parameters():
+        total_numel += p.numel()
+    logger.info(f"This model has {total_numel} parameters")
+
+    # build criterion
+    criterion = BertLoss()
+    logger.info("Criterion is built", ranks=[0])
+
+    # layernorm and bias has no weight decay
+    weight_decay_params = {'params': []}
+    no_weight_decay_params = {'params': [], 'weight_decay': 0.0}
+    for module_ in model.modules():
+        if isinstance(module_, LayerNorm):
+            no_weight_decay_params['params'].extend([p for p in list(module_._parameters.values()) if p is not None])
+        else:
+            weight_decay_params['params'].extend(
+                [p for n, p in list(module_._parameters.items()) if p is not None and n != 'bias'])
+            no_weight_decay_params['params'].extend(
+                [p for n, p in list(module_._parameters.items()) if p is not None and n == 'bias'])
+
+    logger.info(
+        f"without weight decay param: {len(no_weight_decay_params['params'])}, with weight decay param: {len(weight_decay_params['params'])}"
+    )
+    # optimizer
+    optimizer = FusedAdam((weight_decay_params, no_weight_decay_params),
+                          lr=gpc.config.LR,
+                          weight_decay=gpc.config.WEIGHT_DECAY)
+    logger.info("Optimizer is built", ranks=[0])
+
+    # lr scheduler
+    # follow Megatron-LM setting
+    warmup_steps = int(gpc.config.DECAY_ITERS * gpc.config.WARMUP_FRACTION)
+    lr_scheduler = AnnealingLR(optimizer=optimizer,
+                               max_lr=gpc.config.LR,
+                               min_lr=gpc.config.MIN_LR,
+                               warmup_steps=warmup_steps,
+                               decay_steps=gpc.config.DECAY_ITERS,
+                               decay_style='linear')
+    logger.info(f"LR Scheduler is built with {warmup_steps} warmup steps and {gpc.config.DECAY_ITERS} decay steps")
+
+    # # init
+    engine, *dummy = colossalai.initialize(
+        model,
+        optimizer,
+        criterion,
+    )
+
+    # build timer
+    timer = MultiTimer()
+    skip_iters = 0
+
+    # build loss tracker
+    accumulated_train_loss = torch.zeros(1, dtype=torch.float32).cuda()
+    accumulated_eval_loss = torch.zeros(1, dtype=torch.float32).cuda()
+
+    # build data iters for pipeline parallel
+    if use_pipeline:
+        train_data_iter = SequenceParallelDataIterator(trainloader)
+        valid_data_iter = SequenceParallelDataIterator(validloader)
+
+    for step in range(1, gpc.config.TRAIN_ITERS + 1):
+        timer.start('train-iterations')
+        engine.train()
+        if use_pipeline:
+            engine.zero_grad()
+            _, _, train_loss = engine.execute_schedule(train_data_iter, return_output_label=False)
+            engine.step()
+        else:
+            tokens, types, sentence_order, loss_mask, lm_labels, padding_mask = get_batch_for_sequence_parallel(
+                trainloader)
+            engine.zero_grad()
+            lm_loss, sop_output = engine(tokens, padding_mask, types, lm_labels)
+            train_loss = engine.criterion(lm_loss, sop_output, loss_mask, sentence_order)
+            engine.backward(train_loss)
+            engine.step()
+        timer.stop('train-iterations', keep_in_history=True)
+
+        if not gpc.is_initialized(ParallelMode.PIPELINE) or gpc.is_last_rank(ParallelMode.PIPELINE):
+            accumulated_train_loss += train_loss
+
+        lr_scheduler.step()
+
+        if step % gpc.config.EVAL_INTERVAL == 0:
+            engine.eval()
+
+            for j in range(gpc.config.EVAL_ITERS):
+                with torch.no_grad():
+                    if use_pipeline:
+                        _, _, eval_loss = engine.execute_schedule(valid_data_iter,
+                                                                  forward_only=True,
+                                                                  return_output_label=False)
+                    else:
+                        tokens, types, sentence_order, loss_mask, lm_labels, padding_mask = get_batch_for_sequence_parallel(
+                            validloader)
+                        lm_loss, sop_output = engine(tokens, padding_mask, types, lm_labels)
+                        eval_loss = engine.criterion(lm_loss, sop_output, loss_mask, sentence_order)
+
+                    if not gpc.is_initialized(ParallelMode.PIPELINE) or gpc.is_last_rank(ParallelMode.PIPELINE):
+                        accumulated_eval_loss += eval_loss
+
+            if not gpc.is_initialized(ParallelMode.PIPELINE) or gpc.is_last_rank(ParallelMode.PIPELINE):
+                accumulated_eval_loss /= gpc.config.EVAL_ITERS
+                accumulated_train_loss /= gpc.config.EVAL_INTERVAL
+
+            timer_string = []
+            for n, t in timer:
+                timer_string.append(f"{n}: {t.get_history_mean()*1000:.5f}")
+            timer_string = ' | '.join(timer_string)
+            lr = list(engine.optimizer.param_groups)[0]['lr']
+            loss_scale = engine.optimizer.optim.loss_scale.item()
+
+            if gpc.is_initialized(ParallelMode.PIPELINE):
+                ranks = [gpc.get_ranks_in_group(ParallelMode.PIPELINE)[-1]]
+            else:
+                ranks = [0]
+            logger.info(f'Step {step} / {gpc.config.TRAIN_ITERS} | Train Loss: {accumulated_train_loss.item():.5g} ' +
+                        f'| Eval Loss: {accumulated_eval_loss.item():.5g} ' + f'| Loss Scale: {loss_scale}' +
+                        f"| Learning rate: {lr} | " + timer_string,
+                        ranks=ranks)
+
+            for n, t in timer:
+                t.reset()
+            accumulated_eval_loss.zero_()
+            accumulated_train_loss.zero_()
+
+
+if __name__ == '__main__':
+    main()

examples/tutorial/stable_diffusion/LICENSE

@@ -0,0 +1,82 @@
+Copyright (c) 2022 Robin Rombach and Patrick Esser and contributors
+
+CreativeML Open RAIL-M
+dated August 22, 2022
+
+Section I: PREAMBLE
+
+Multimodal generative models are being widely adopted and used, and have the potential to transform the way artists, among other individuals, conceive and benefit from AI or ML technologies as a tool for content creation.
+
+Notwithstanding the current and potential benefits that these artifacts can bring to society at large, there are also concerns about potential misuses of them, either due to their technical limitations or ethical considerations.
+
+In short, this license strives for both the open and responsible downstream use of the accompanying model. When it comes to the open character, we took inspiration from open source permissive licenses regarding the grant of IP rights. Referring to the downstream responsible use, we added use-based restrictions not permitting the use of the Model in very specific scenarios, in order for the licensor to be able to enforce the license in case potential misuses of the Model may occur. At the same time, we strive to promote open and responsible research on generative models for art and content generation.
+
+Even though downstream derivative versions of the model could be released under different licensing terms, the latter will always have to include - at minimum - the same use-based restrictions as the ones in the original license (this license). We believe in the intersection between open and responsible AI development; thus, this License aims to strike a balance between both in order to enable responsible open-science in the field of AI.
+
+This License governs the use of the model (and its derivatives) and is informed by the model card associated with the model.
+
+NOW THEREFORE, You and Licensor agree as follows:
+
+1. Definitions
+
+- "License" means the terms and conditions for use, reproduction, and Distribution as defined in this document.
+- "Data" means a collection of information and/or content extracted from the dataset used with the Model, including to train, pretrain, or otherwise evaluate the Model. The Data is not licensed under this License.
+- "Output" means the results of operating a Model as embodied in informational content resulting therefrom.
+- "Model" means any accompanying machine-learning based assemblies (including checkpoints), consisting of learnt weights, parameters (including optimizer states), corresponding to the model architecture as embodied in the Complementary Material, that have been trained or tuned, in whole or in part on the Data, using the Complementary Material.
+- "Derivatives of the Model" means all modifications to the Model, works based on the Model, or any other model which is created or initialized by transfer of patterns of the weights, parameters, activations or output of the Model, to the other model, in order to cause the other model to perform similarly to the Model, including - but not limited to - distillation methods entailing the use of intermediate data representations or methods based on the generation of synthetic data by the Model for training the other model.
+- "Complementary Material" means the accompanying source code and scripts used to define, run, load, benchmark or evaluate the Model, and used to prepare data for training or evaluation, if any. This includes any accompanying documentation, tutorials, examples, etc, if any.
+- "Distribution" means any transmission, reproduction, publication or other sharing of the Model or Derivatives of the Model to a third party, including providing the Model as a hosted service made available by electronic or other remote means - e.g. API-based or web access.
+- "Licensor" means the copyright owner or entity authorized by the copyright owner that is granting the License, including the persons or entities that may have rights in the Model and/or distributing the Model.
+- "You" (or "Your") means an individual or Legal Entity exercising permissions granted by this License and/or making use of the Model for whichever purpose and in any field of use, including usage of the Model in an end-use application - e.g. chatbot, translator, image generator.
+- "Third Parties" means individuals or legal entities that are not under common control with Licensor or You.
+- "Contribution" means any work of authorship, including the original version of the Model and any modifications or additions to that Model or Derivatives of the Model thereof, that is intentionally submitted to Licensor for inclusion in the Model by the copyright owner or by an individual or Legal Entity authorized to submit on behalf of the copyright owner. For the purposes of this definition, "submitted" means any form of electronic, verbal, or written communication sent to the Licensor or its representatives, including but not limited to communication on electronic mailing lists, source code control systems, and issue tracking systems that are managed by, or on behalf of, the Licensor for the purpose of discussing and improving the Model, but excluding communication that is conspicuously marked or otherwise designated in writing by the copyright owner as "Not a Contribution."
+- "Contributor" means Licensor and any individual or Legal Entity on behalf of whom a Contribution has been received by Licensor and subsequently incorporated within the Model.
+
+Section II: INTELLECTUAL PROPERTY RIGHTS
+
+Both copyright and patent grants apply to the Model, Derivatives of the Model and Complementary Material. The Model and Derivatives of the Model are subject to additional terms as described in Section III.
+
+2. Grant of Copyright License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare, publicly display, publicly perform, sublicense, and distribute the Complementary Material, the Model, and Derivatives of the Model.
+3. Grant of Patent License. Subject to the terms and conditions of this License and where and as applicable, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this paragraph) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Model and the Complementary Material, where such license applies only to those patent claims licensable by such Contributor that are necessarily infringed by their Contribution(s) alone or by combination of their Contribution(s) with the Model to which such Contribution(s) was submitted. If You institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Model and/or Complementary Material or a Contribution incorporated within the Model and/or Complementary Material constitutes direct or contributory patent infringement, then any patent licenses granted to You under this License for the Model and/or Work shall terminate as of the date such litigation is asserted or filed.
+
+Section III: CONDITIONS OF USAGE, DISTRIBUTION AND REDISTRIBUTION
+
+4. Distribution and Redistribution. You may host for Third Party remote access purposes (e.g. software-as-a-service), reproduce and distribute copies of the Model or Derivatives of the Model thereof in any medium, with or without modifications, provided that You meet the following conditions:
+Use-based restrictions as referenced in paragraph 5 MUST be included as an enforceable provision by You in any type of legal agreement (e.g. a license) governing the use and/or distribution of the Model or Derivatives of the Model, and You shall give notice to subsequent users You Distribute to, that the Model or Derivatives of the Model are subject to paragraph 5. This provision does not apply to the use of Complementary Material.
+You must give any Third Party recipients of the Model or Derivatives of the Model a copy of this License;
+You must cause any modified files to carry prominent notices stating that You changed the files;
+You must retain all copyright, patent, trademark, and attribution notices excluding those notices that do not pertain to any part of the Model, Derivatives of the Model.
+You may add Your own copyright statement to Your modifications and may provide additional or different license terms and conditions - respecting paragraph 4.a. - for use, reproduction, or Distribution of Your modifications, or for any such Derivatives of the Model as a whole, provided Your use, reproduction, and Distribution of the Model otherwise complies with the conditions stated in this License.
+5. Use-based restrictions. The restrictions set forth in Attachment A are considered Use-based restrictions. Therefore You cannot use the Model and the Derivatives of the Model for the specified restricted uses. You may use the Model subject to this License, including only for lawful purposes and in accordance with the License. Use may include creating any content with, finetuning, updating, running, training, evaluating and/or reparametrizing the Model. You shall require all of Your users who use the Model or a Derivative of the Model to comply with the terms of this paragraph (paragraph 5).
+6. The Output You Generate. Except as set forth herein, Licensor claims no rights in the Output You generate using the Model. You are accountable for the Output you generate and its subsequent uses. No use of the output can contravene any provision as stated in the License.
+
+Section IV: OTHER PROVISIONS
+
+7. Updates and Runtime Restrictions. To the maximum extent permitted by law, Licensor reserves the right to restrict (remotely or otherwise) usage of the Model in violation of this License, update the Model through electronic means, or modify the Output of the Model based on updates. You shall undertake reasonable efforts to use the latest version of the Model.
+8. Trademarks and related. Nothing in this License permits You to make use of Licensors’ trademarks, trade names, logos or to otherwise suggest endorsement or misrepresent the relationship between the parties; and any rights not expressly granted herein are reserved by the Licensors.
+9. Disclaimer of Warranty. Unless required by applicable law or agreed to in writing, Licensor provides the Model and the Complementary Material (and each Contributor provides its Contributions) on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. You are solely responsible for determining the appropriateness of using or redistributing the Model, Derivatives of the Model, and the Complementary Material and assume any risks associated with Your exercise of permissions under this License.
+10. Limitation of Liability. In no event and under no legal theory, whether in tort (including negligence), contract, or otherwise, unless required by applicable law (such as deliberate and grossly negligent acts) or agreed to in writing, shall any Contributor be liable to You for damages, including any direct, indirect, special, incidental, or consequential damages of any character arising as a result of this License or out of the use or inability to use the Model and the Complementary Material (including but not limited to damages for loss of goodwill, work stoppage, computer failure or malfunction, or any and all other commercial damages or losses), even if such Contributor has been advised of the possibility of such damages.
+11. Accepting Warranty or Additional Liability. While redistributing the Model, Derivatives of the Model and the Complementary Material thereof, You may choose to offer, and charge a fee for, acceptance of support, warranty, indemnity, or other liability obligations and/or rights consistent with this License. However, in accepting such obligations, You may act only on Your own behalf and on Your sole responsibility, not on behalf of any other Contributor, and only if You agree to indemnify, defend, and hold each Contributor harmless for any liability incurred by, or claims asserted against, such Contributor by reason of your accepting any such warranty or additional liability.
+12. If any provision of this License is held to be invalid, illegal or unenforceable, the remaining provisions shall be unaffected thereby and remain valid as if such provision had not been set forth herein.
+
+END OF TERMS AND CONDITIONS
+
+
+
+
+Attachment A
+
+Use Restrictions
+
+You agree not to use the Model or Derivatives of the Model:
+- In any way that violates any applicable national, federal, state, local or international law or regulation;
+- For the purpose of exploiting, harming or attempting to exploit or harm minors in any way;
+- To generate or disseminate verifiably false information and/or content with the purpose of harming others;
+- To generate or disseminate personal identifiable information that can be used to harm an individual;
+- To defame, disparage or otherwise harass others;
+- For fully automated decision making that adversely impacts an individual’s legal rights or otherwise creates or modifies a binding, enforceable obligation;
+- For any use intended to or which has the effect of discriminating against or harming individuals or groups based on online or offline social behavior or known or predicted personal or personality characteristics;
+- To exploit any of the vulnerabilities of a specific group of persons based on their age, social, physical or mental characteristics, in order to materially distort the behavior of a person pertaining to that group in a manner that causes or is likely to cause that person or another person physical or psychological harm;
+- For any use intended to or which has the effect of discriminating against individuals or groups based on legally protected characteristics or categories;
+- To provide medical advice and medical results interpretation;
+- To generate or disseminate information for the purpose to be used for administration of justice, law enforcement, immigration or asylum processes, such as predicting an individual will commit fraud/crime commitment (e.g. by text profiling, drawing causal relationships between assertions made in documents, indiscriminate and arbitrarily-targeted use).

examples/tutorial/stable_diffusion/README.md

@@ -0,0 +1,115 @@
+# Handson 6: Acceleration of Stable Diffusion
+
+*[Colosssal-AI](https://github.com/hpcaitech/ColossalAI) provides a faster and lower cost solution for pretraining and
+fine-tuning for AIGC (AI-Generated Content) applications such as the model [stable-diffusion](https://github.com/CompVis/stable-diffusion) from [Stability AI](https://stability.ai/).*
+
+We take advantage of [Colosssal-AI](https://github.com/hpcaitech/ColossalAI) to exploit multiple optimization strategies
+, e.g. data parallelism, tensor parallelism, mixed precision & ZeRO, to scale the training to multiple GPUs.
+
+## Stable Diffusion
+[Stable Diffusion](https://huggingface.co/CompVis/stable-diffusion) is a latent text-to-image diffusion
+model.
+Thanks to a generous compute donation from [Stability AI](https://stability.ai/) and support from [LAION](https://laion.ai/), we were able to train a Latent Diffusion Model on 512x512 images from a subset of the [LAION-5B](https://laion.ai/blog/laion-5b/) database.
+Similar to Google's [Imagen](https://arxiv.org/abs/2205.11487),
+this model uses a frozen CLIP ViT-L/14 text encoder to condition the model on text prompts.
+
+<p id="diffusion_train" align="center">
+<img src="https://raw.githubusercontent.com/hpcaitech/public_assets/main/colossalai/img/diffusion_train.png" width=800/>
+</p>
+
+[Stable Diffusion with Colossal-AI](https://github.com/hpcaitech/ColossalAI/tree/main/examples/images/diffusion) provides **6.5x faster training and pretraining cost saving, the hardware cost of fine-tuning can be almost 7X cheaper** (from RTX3090/4090 24GB to RTX3050/2070 8GB).
+
+<p id="diffusion_demo" align="center">
+<img src="https://raw.githubusercontent.com/hpcaitech/public_assets/main/colossalai/img/diffusion_demo.png" width=800/>
+</p>
+
+## Requirements
+A suitable [conda](https://conda.io/) environment named `ldm` can be created
+and activated with:
+
+```
+conda env create -f environment.yaml
+conda activate ldm
+```
+
+You can also update an existing [latent diffusion](https://github.com/CompVis/latent-diffusion) environment by running
+
+```
+conda install pytorch torchvision -c pytorch
+pip install transformers==4.19.2 diffusers invisible-watermark
+pip install -e .
+```
+
+### Install [Colossal-AI v0.1.10](https://colossalai.org/download/) From Our Official Website
+```
+pip install colossalai==0.1.10+torch1.11cu11.3 -f https://release.colossalai.org
+```
+
+### Install [Lightning](https://github.com/Lightning-AI/lightning)
+We use the Sep. 2022 version with commit id as `b04a7aa`.
+```
+git clone https://github.com/Lightning-AI/lightning && cd lightning && git reset --hard b04a7aa
+pip install -r requirements.txt && pip install .
+```
+
+> The specified version is due to the interface incompatibility caused by the latest update of [Lightning](https://github.com/Lightning-AI/lightning), which will be fixed in the near future.
+
+## Dataset
+The DataSet is from [LAION-5B](https://laion.ai/blog/laion-5b/), the subset of [LAION](https://laion.ai/),
+you should the change the `data.file_path` in the `config/train_colossalai.yaml`
+
+## Training
+
+we provide the script `train.sh` to run the training task , and two Stategy in `configs`:`train_colossalai.yaml`, `train_ddp.yaml`
+
+for example, you can run the training from colossalai by
+```
+python main.py --logdir /tmp -t --postfix test -b config/train_colossalai.yaml
+```
+
+- you can change the `--logdir` the save the log information and the last checkpoint
+
+### Training config
+you can change the trainging config in the yaml file
+
+- accelerator: acceleratortype, default 'gpu'
+- devices: device number used for training, default 4
+- max_epochs: max training epochs
+- precision: usefp16 for training or not, default 16, you must use fp16 if you want to apply colossalai
+
+
+## Comments
+
+- Our codebase for the diffusion models builds heavily on [OpenAI's ADM codebase](https://github.com/openai/guided-diffusion)
+, [lucidrains](https://github.com/lucidrains/denoising-diffusion-pytorch),
+[Stable Diffusion](https://github.com/CompVis/stable-diffusion), [Lightning](https://github.com/Lightning-AI/lightning) and [Hugging Face](https://huggingface.co/CompVis/stable-diffusion).
+Thanks for open-sourcing!
+
+- The implementation of the transformer encoder is from [x-transformers](https://github.com/lucidrains/x-transformers) by [lucidrains](https://github.com/lucidrains?tab=repositories).
+
+- The implementation of [flash attention](https://github.com/HazyResearch/flash-attention) is from [HazyResearch](https://github.com/HazyResearch).
+
+## BibTeX
+
+```
+@article{bian2021colossal,
+  title={Colossal-AI: A Unified Deep Learning System For Large-Scale Parallel Training},
+  author={Bian, Zhengda and Liu, Hongxin and Wang, Boxiang and Huang, Haichen and Li, Yongbin and Wang, Chuanrui and Cui, Fan and You, Yang},
+  journal={arXiv preprint arXiv:2110.14883},
+  year={2021}
+}
+@misc{rombach2021highresolution,
+  title={High-Resolution Image Synthesis with Latent Diffusion Models},
+  author={Robin Rombach and Andreas Blattmann and Dominik Lorenz and Patrick Esser and Björn Ommer},
+  year={2021},
+  eprint={2112.10752},
+  archivePrefix={arXiv},
+  primaryClass={cs.CV}
+}
+@article{dao2022flashattention,
+  title={FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness},
+  author={Dao, Tri and Fu, Daniel Y. and Ermon, Stefano and Rudra, Atri and R{\'e}, Christopher},
+  journal={arXiv preprint arXiv:2205.14135},
+  year={2022}
+}
+```

examples/tutorial/stable_diffusion/configs/train_colossalai.yaml

@@ -0,0 +1,116 @@
+model:
+  base_learning_rate: 1.0e-04
+  target: ldm.models.diffusion.ddpm.LatentDiffusion
+  params:
+    linear_start: 0.00085
+    linear_end: 0.0120
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: image
+    cond_stage_key: caption
+    image_size: 64
+    channels: 4
+    cond_stage_trainable: false   # Note: different from the one we trained before
+    conditioning_key: crossattn
+    monitor: val/loss_simple_ema
+    scale_factor: 0.18215
+    use_ema: False
+
+    scheduler_config: # 10000 warmup steps
+      target: ldm.lr_scheduler.LambdaLinearScheduler
+      params:
+        warm_up_steps: [ 1 ] # NOTE for resuming. use 10000 if starting from scratch
+        cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
+        f_start: [ 1.e-6 ]
+        f_max: [ 1.e-4 ]
+        f_min: [ 1.e-10 ]
+
+    unet_config:
+      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        image_size: 32 # unused
+        from_pretrained: '/data/scratch/diffuser/stable-diffusion-v1-4/unet/diffusion_pytorch_model.bin'
+        in_channels: 4
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions: [ 4, 2, 1 ]
+        num_res_blocks: 2
+        channel_mult: [ 1, 2, 4, 4 ]
+        num_heads: 8
+        use_spatial_transformer: True
+        transformer_depth: 1
+        context_dim: 768
+        use_checkpoint: False
+        legacy: False
+
+    first_stage_config:
+      target: ldm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        from_pretrained: '/data/scratch/diffuser/stable-diffusion-v1-4/vae/diffusion_pytorch_model.bin'
+        monitor: val/rec_loss
+        ddconfig:
+          double_z: true
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+          - 1
+          - 2
+          - 4
+          - 4
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+
+    cond_stage_config:
+      target: ldm.modules.encoders.modules.FrozenCLIPEmbedder
+      params:
+        use_fp16: True
+
+data:
+  target: main.DataModuleFromConfig
+  params:
+    batch_size: 64
+    wrap: False
+    train:
+      target: ldm.data.base.Txt2ImgIterableBaseDataset
+      params:
+        file_path: "/data/scratch/diffuser/laion_part0/"
+        world_size: 1
+        rank: 0
+
+lightning:
+  trainer:
+    accelerator: 'gpu' 
+    devices: 4
+    log_gpu_memory: all
+    max_epochs: 2
+    precision: 16
+    auto_select_gpus: False
+    strategy:
+      target: pytorch_lightning.strategies.ColossalAIStrategy
+      params:
+        use_chunk: False
+        enable_distributed_storage: True,
+        placement_policy: cuda
+        force_outputs_fp32: False
+
+    log_every_n_steps: 2
+    logger: True
+    default_root_dir: "/tmp/diff_log/"
+    profiler: pytorch
+
+  logger_config:
+    wandb:
+      target: pytorch_lightning.loggers.WandbLogger
+      params:
+          name: nowname
+          save_dir: "/tmp/diff_log/"
+          offline: opt.debug
+          id: nowname

examples/tutorial/stable_diffusion/configs/train_ddp.yaml

@@ -0,0 +1,113 @@
+model:
+  base_learning_rate: 1.0e-04
+  target: ldm.models.diffusion.ddpm.LatentDiffusion
+  params:
+    linear_start: 0.00085
+    linear_end: 0.0120
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: image
+    cond_stage_key: caption
+    image_size: 32
+    channels: 4
+    cond_stage_trainable: false   # Note: different from the one we trained before
+    conditioning_key: crossattn
+    monitor: val/loss_simple_ema
+    scale_factor: 0.18215
+    use_ema: False
+
+    scheduler_config: # 10000 warmup steps
+      target: ldm.lr_scheduler.LambdaLinearScheduler
+      params:
+        warm_up_steps: [ 100 ]
+        cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
+        f_start: [ 1.e-6 ]
+        f_max: [ 1.e-4 ]
+        f_min: [ 1.e-10  ]
+
+    unet_config:
+      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        image_size: 32 # unused
+        from_pretrained: '/data/scratch/diffuser/stable-diffusion-v1-4/unet/diffusion_pytorch_model.bin'
+        in_channels: 4
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions: [ 4, 2, 1 ]
+        num_res_blocks: 2
+        channel_mult: [ 1, 2, 4, 4 ]
+        num_heads: 8
+        use_spatial_transformer: True
+        transformer_depth: 1
+        context_dim: 768
+        use_checkpoint: False
+        legacy: False
+
+    first_stage_config:
+      target: ldm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        from_pretrained: '/data/scratch/diffuser/stable-diffusion-v1-4/vae/diffusion_pytorch_model.bin'
+        monitor: val/rec_loss
+        ddconfig:
+          double_z: true
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+          - 1
+          - 2
+          - 4
+          - 4
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+
+    cond_stage_config:
+      target: ldm.modules.encoders.modules.FrozenCLIPEmbedder
+      params:
+        use_fp16: True
+
+data:
+  target: main.DataModuleFromConfig
+  params:
+    batch_size: 64
+    wrap: False
+    train:
+      target: ldm.data.base.Txt2ImgIterableBaseDataset
+      params:
+        file_path: "/data/scratch/diffuser/laion_part0/"
+        world_size: 1
+        rank: 0
+
+lightning:
+  trainer:
+    accelerator: 'gpu' 
+    devices: 4
+    log_gpu_memory: all
+    max_epochs: 2
+    precision: 16
+    auto_select_gpus: False
+    strategy:
+      target: pytorch_lightning.strategies.DDPStrategy
+      params:
+        find_unused_parameters: False
+    log_every_n_steps: 2
+#    max_steps: 6o
+    logger: True
+    default_root_dir: "/tmp/diff_log/"
+    # profiler: pytorch
+
+  logger_config:
+    wandb:
+      target: pytorch_lightning.loggers.WandbLogger
+      params:
+          name: nowname
+          save_dir: "/tmp/diff_log/"
+          offline: opt.debug
+          id: nowname

examples/tutorial/stable_diffusion/configs/train_pokemon.yaml

@@ -0,0 +1,121 @@
+model:
+  base_learning_rate: 1.0e-04
+  target: ldm.models.diffusion.ddpm.LatentDiffusion
+  params:
+    linear_start: 0.00085
+    linear_end: 0.0120
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: image
+    cond_stage_key: caption
+    image_size: 32
+    channels: 4
+    cond_stage_trainable: false   # Note: different from the one we trained before
+    conditioning_key: crossattn
+    monitor: val/loss_simple_ema
+    scale_factor: 0.18215
+    use_ema: False
+    check_nan_inf: False
+
+    scheduler_config: # 10000 warmup steps
+      target: ldm.lr_scheduler.LambdaLinearScheduler
+      params:
+        warm_up_steps: [ 10000 ]
+        cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
+        f_start: [ 1.e-6 ]
+        f_max: [ 1.e-4 ]
+        f_min: [ 1.e-10  ]
+
+    unet_config:
+      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        image_size: 32 # unused
+        from_pretrained: '/data/scratch/diffuser/stable-diffusion-v1-4/unet/diffusion_pytorch_model.bin'
+        in_channels: 4
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions: [ 4, 2, 1 ]
+        num_res_blocks: 2
+        channel_mult: [ 1, 2, 4, 4 ]
+        num_heads: 8
+        use_spatial_transformer: True
+        transformer_depth: 1
+        context_dim: 768
+        use_checkpoint: False
+        legacy: False
+
+    first_stage_config:
+      target: ldm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        from_pretrained: '/data/scratch/diffuser/stable-diffusion-v1-4/vae/diffusion_pytorch_model.bin'
+        monitor: val/rec_loss
+        ddconfig:
+          double_z: true
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+          - 1
+          - 2
+          - 4
+          - 4
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+
+    cond_stage_config:
+      target: ldm.modules.encoders.modules.FrozenCLIPEmbedder
+      params:
+        use_fp16: True
+
+data:
+  target: main.DataModuleFromConfig
+  params:
+    batch_size: 32
+    wrap: False
+    train:
+      target: ldm.data.pokemon.PokemonDataset
+      # params:
+        # file_path: "/data/scratch/diffuser/laion_part0/"
+        # world_size: 1
+        # rank: 0
+
+lightning:
+  trainer:
+    accelerator: 'gpu' 
+    devices: 4
+    log_gpu_memory: all
+    max_epochs: 2
+    precision: 16
+    auto_select_gpus: False
+    strategy:
+      target: pytorch_lightning.strategies.ColossalAIStrategy
+      params:
+        use_chunk: False
+        enable_distributed_storage: True,
+        placement_policy: cuda
+        force_outputs_fp32: False
+        initial_scale: 65536
+        min_scale: 1
+        max_scale: 65536
+        # max_scale: 4294967296
+
+    log_every_n_steps: 2
+    logger: True
+    default_root_dir: "/tmp/diff_log/"
+    profiler: pytorch
+
+  logger_config:
+    wandb:
+      target: pytorch_lightning.loggers.WandbLogger
+      params:
+          name: nowname
+          save_dir: "/tmp/diff_log/"
+          offline: opt.debug
+          id: nowname

examples/tutorial/stable_diffusion/environment.yaml

@@ -0,0 +1,32 @@
+name: ldm
+channels:
+  - pytorch
+  - defaults
+dependencies:
+  - python=3.9.12
+  - pip=20.3
+  - cudatoolkit=11.3
+  - pytorch=1.11.0
+  - torchvision=0.12.0
+  - numpy=1.19.2
+  - pip:
+    - albumentations==0.4.3
+    - diffusers
+    - opencv-python==4.6.0.66
+    - pudb==2019.2
+    - invisible-watermark
+    - imageio==2.9.0
+    - imageio-ffmpeg==0.4.2
+    - pytorch-lightning==1.4.2
+    - omegaconf==2.1.1
+    - test-tube>=0.7.5
+    - streamlit>=0.73.1
+    - einops==0.3.0
+    - torch-fidelity==0.3.0
+    - transformers==4.19.2
+    - torchmetrics==0.6.0
+    - kornia==0.6
+    - prefetch_generator
+    - -e git+https://github.com/CompVis/taming-transformers.git@master#egg=taming-transformers
+    - -e git+https://github.com/openai/CLIP.git@main#egg=clip
+    - -e .

examples/tutorial/stable_diffusion/ldm/data/base.py

@@ -0,0 +1,75 @@
+import math
+from abc import abstractmethod
+
+import torch
+from torch.utils.data import Dataset, ConcatDataset, ChainDataset, IterableDataset
+import os
+import numpy as np
+import cv2
+
+class Txt2ImgIterableBaseDataset(IterableDataset):
+    '''
+    Define an interface to make the IterableDatasets for text2img data chainable
+    '''
+    def __init__(self, file_path: str, rank, world_size):
+        super().__init__()
+        self.file_path = file_path
+        self.folder_list = []
+        self.file_list = []
+        self.txt_list = []
+        self.info = self._get_file_info(file_path)
+        self.start = self.info['start']
+        self.end = self.info['end']
+        self.rank = rank
+
+        self.world_size = world_size
+        # self.per_worker = int(math.floor((self.end - self.start) / float(self.world_size)))
+        # self.iter_start = self.start + self.rank * self.per_worker
+        # self.iter_end = min(self.iter_start + self.per_worker, self.end)
+        # self.num_records = self.iter_end - self.iter_start
+        # self.valid_ids = [i for i in range(self.iter_end)]
+        self.num_records = self.end - self.start
+        self.valid_ids = [i for i in range(self.end)]
+
+        print(f'{self.__class__.__name__} dataset contains {self.__len__()} examples.')
+
+    def __len__(self):
+        # return self.iter_end - self.iter_start
+        return self.end - self.start
+
+    def __iter__(self):
+        sample_iterator = self._sample_generator(self.start, self.end)
+        # sample_iterator = self._sample_generator(self.iter_start, self.iter_end)
+        return sample_iterator
+
+    def _sample_generator(self, start, end):
+        for idx in range(start, end):
+            file_name = self.file_list[idx]
+            txt_name = self.txt_list[idx]
+            f_ = open(txt_name, 'r')
+            txt_ = f_.read()
+            f_.close()
+            image = cv2.imdecode(np.fromfile(file_name, dtype=np.uint8), 1)
+            image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
+            image = torch.from_numpy(image) / 255
+            yield {"caption": txt_, "image":image}
+
+
+    def _get_file_info(self, file_path):
+        info = \
+        {
+            "start": 1,
+            "end": 0,
+        }
+        self.folder_list = [file_path + i for i in os.listdir(file_path) if '.' not in i]
+        for folder in self.folder_list:
+            files = [folder + '/' + i for i in os.listdir(folder) if 'jpg' in i]
+            txts = [k.replace('jpg', 'txt') for k in files]
+            self.file_list.extend(files)
+            self.txt_list.extend(txts)
+        info['end'] = len(self.file_list)
+        # with open(file_path, 'r') as fin:
+        #     for _ in enumerate(fin):
+        #         info['end'] += 1
+        # self.txt_list = [k.replace('jpg', 'txt') for k in self.file_list]
+        return info

examples/tutorial/stable_diffusion/ldm/data/imagenet.py

@@ -0,0 +1,394 @@
+import os, yaml, pickle, shutil, tarfile, glob
+import cv2
+import albumentations
+import PIL
+import numpy as np
+import torchvision.transforms.functional as TF
+from omegaconf import OmegaConf
+from functools import partial
+from PIL import Image
+from tqdm import tqdm
+from torch.utils.data import Dataset, Subset
+
+import taming.data.utils as tdu
+from taming.data.imagenet import str_to_indices, give_synsets_from_indices, download, retrieve
+from taming.data.imagenet import ImagePaths
+
+from ldm.modules.image_degradation import degradation_fn_bsr, degradation_fn_bsr_light
+
+
+def synset2idx(path_to_yaml="data/index_synset.yaml"):
+    with open(path_to_yaml) as f:
+        di2s = yaml.load(f)
+    return dict((v,k) for k,v in di2s.items())
+
+
+class ImageNetBase(Dataset):
+    def __init__(self, config=None):
+        self.config = config or OmegaConf.create()
+        if not type(self.config)==dict:
+            self.config = OmegaConf.to_container(self.config)
+        self.keep_orig_class_label = self.config.get("keep_orig_class_label", False)
+        self.process_images = True  # if False we skip loading & processing images and self.data contains filepaths
+        self._prepare()
+        self._prepare_synset_to_human()
+        self._prepare_idx_to_synset()
+        self._prepare_human_to_integer_label()
+        self._load()
+
+    def __len__(self):
+        return len(self.data)
+
+    def __getitem__(self, i):
+        return self.data[i]
+
+    def _prepare(self):
+        raise NotImplementedError()
+
+    def _filter_relpaths(self, relpaths):
+        ignore = set([
+            "n06596364_9591.JPEG",
+        ])
+        relpaths = [rpath for rpath in relpaths if not rpath.split("/")[-1] in ignore]
+        if "sub_indices" in self.config:
+            indices = str_to_indices(self.config["sub_indices"])
+            synsets = give_synsets_from_indices(indices, path_to_yaml=self.idx2syn)  # returns a list of strings
+            self.synset2idx = synset2idx(path_to_yaml=self.idx2syn)
+            files = []
+            for rpath in relpaths:
+                syn = rpath.split("/")[0]
+                if syn in synsets:
+                    files.append(rpath)
+            return files
+        else:
+            return relpaths
+
+    def _prepare_synset_to_human(self):
+        SIZE = 2655750
+        URL = "https://heibox.uni-heidelberg.de/f/9f28e956cd304264bb82/?dl=1"
+        self.human_dict = os.path.join(self.root, "synset_human.txt")
+        if (not os.path.exists(self.human_dict) or
+                not os.path.getsize(self.human_dict)==SIZE):
+            download(URL, self.human_dict)
+
+    def _prepare_idx_to_synset(self):
+        URL = "https://heibox.uni-heidelberg.de/f/d835d5b6ceda4d3aa910/?dl=1"
+        self.idx2syn = os.path.join(self.root, "index_synset.yaml")
+        if (not os.path.exists(self.idx2syn)):
+            download(URL, self.idx2syn)
+
+    def _prepare_human_to_integer_label(self):
+        URL = "https://heibox.uni-heidelberg.de/f/2362b797d5be43b883f6/?dl=1"
+        self.human2integer = os.path.join(self.root, "imagenet1000_clsidx_to_labels.txt")
+        if (not os.path.exists(self.human2integer)):
+            download(URL, self.human2integer)
+        with open(self.human2integer, "r") as f:
+            lines = f.read().splitlines()
+            assert len(lines) == 1000
+            self.human2integer_dict = dict()
+            for line in lines:
+                value, key = line.split(":")
+                self.human2integer_dict[key] = int(value)
+
+    def _load(self):
+        with open(self.txt_filelist, "r") as f:
+            self.relpaths = f.read().splitlines()
+            l1 = len(self.relpaths)
+            self.relpaths = self._filter_relpaths(self.relpaths)
+            print("Removed {} files from filelist during filtering.".format(l1 - len(self.relpaths)))
+
+        self.synsets = [p.split("/")[0] for p in self.relpaths]
+        self.abspaths = [os.path.join(self.datadir, p) for p in self.relpaths]
+
+        unique_synsets = np.unique(self.synsets)
+        class_dict = dict((synset, i) for i, synset in enumerate(unique_synsets))
+        if not self.keep_orig_class_label:
+            self.class_labels = [class_dict[s] for s in self.synsets]
+        else:
+            self.class_labels = [self.synset2idx[s] for s in self.synsets]
+
+        with open(self.human_dict, "r") as f:
+            human_dict = f.read().splitlines()
+            human_dict = dict(line.split(maxsplit=1) for line in human_dict)
+
+        self.human_labels = [human_dict[s] for s in self.synsets]
+
+        labels = {
+            "relpath": np.array(self.relpaths),
+            "synsets": np.array(self.synsets),
+            "class_label": np.array(self.class_labels),
+            "human_label": np.array(self.human_labels),
+        }
+
+        if self.process_images:
+            self.size = retrieve(self.config, "size", default=256)
+            self.data = ImagePaths(self.abspaths,
+                                   labels=labels,
+                                   size=self.size,
+                                   random_crop=self.random_crop,
+                                   )
+        else:
+            self.data = self.abspaths
+
+
+class ImageNetTrain(ImageNetBase):
+    NAME = "ILSVRC2012_train"
+    URL = "http://www.image-net.org/challenges/LSVRC/2012/"
+    AT_HASH = "a306397ccf9c2ead27155983c254227c0fd938e2"
+    FILES = [
+        "ILSVRC2012_img_train.tar",
+    ]
+    SIZES = [
+        147897477120,
+    ]
+
+    def __init__(self, process_images=True, data_root=None, **kwargs):
+        self.process_images = process_images
+        self.data_root = data_root
+        super().__init__(**kwargs)
+
+    def _prepare(self):
+        if self.data_root:
+            self.root = os.path.join(self.data_root, self.NAME)
+        else:
+            cachedir = os.environ.get("XDG_CACHE_HOME", os.path.expanduser("~/.cache"))
+            self.root = os.path.join(cachedir, "autoencoders/data", self.NAME)
+
+        self.datadir = os.path.join(self.root, "data")
+        self.txt_filelist = os.path.join(self.root, "filelist.txt")
+        self.expected_length = 1281167
+        self.random_crop = retrieve(self.config, "ImageNetTrain/random_crop",
+                                    default=True)
+        if not tdu.is_prepared(self.root):
+            # prep
+            print("Preparing dataset {} in {}".format(self.NAME, self.root))
+
+            datadir = self.datadir
+            if not os.path.exists(datadir):
+                path = os.path.join(self.root, self.FILES[0])
+                if not os.path.exists(path) or not os.path.getsize(path)==self.SIZES[0]:
+                    import academictorrents as at
+                    atpath = at.get(self.AT_HASH, datastore=self.root)
+                    assert atpath == path
+
+                print("Extracting {} to {}".format(path, datadir))
+                os.makedirs(datadir, exist_ok=True)
+                with tarfile.open(path, "r:") as tar:
+                    tar.extractall(path=datadir)
+
+                print("Extracting sub-tars.")
+                subpaths = sorted(glob.glob(os.path.join(datadir, "*.tar")))
+                for subpath in tqdm(subpaths):
+                    subdir = subpath[:-len(".tar")]
+                    os.makedirs(subdir, exist_ok=True)
+                    with tarfile.open(subpath, "r:") as tar:
+                        tar.extractall(path=subdir)
+
+            filelist = glob.glob(os.path.join(datadir, "**", "*.JPEG"))
+            filelist = [os.path.relpath(p, start=datadir) for p in filelist]
+            filelist = sorted(filelist)
+            filelist = "\n".join(filelist)+"\n"
+            with open(self.txt_filelist, "w") as f:
+                f.write(filelist)
+
+            tdu.mark_prepared(self.root)
+
+
+class ImageNetValidation(ImageNetBase):
+    NAME = "ILSVRC2012_validation"
+    URL = "http://www.image-net.org/challenges/LSVRC/2012/"
+    AT_HASH = "5d6d0df7ed81efd49ca99ea4737e0ae5e3a5f2e5"
+    VS_URL = "https://heibox.uni-heidelberg.de/f/3e0f6e9c624e45f2bd73/?dl=1"
+    FILES = [
+        "ILSVRC2012_img_val.tar",
+        "validation_synset.txt",
+    ]
+    SIZES = [
+        6744924160,
+        1950000,
+    ]
+
+    def __init__(self, process_images=True, data_root=None, **kwargs):
+        self.data_root = data_root
+        self.process_images = process_images
+        super().__init__(**kwargs)
+
+    def _prepare(self):
+        if self.data_root:
+            self.root = os.path.join(self.data_root, self.NAME)
+        else:
+            cachedir = os.environ.get("XDG_CACHE_HOME", os.path.expanduser("~/.cache"))
+            self.root = os.path.join(cachedir, "autoencoders/data", self.NAME)
+        self.datadir = os.path.join(self.root, "data")
+        self.txt_filelist = os.path.join(self.root, "filelist.txt")
+        self.expected_length = 50000
+        self.random_crop = retrieve(self.config, "ImageNetValidation/random_crop",
+                                    default=False)
+        if not tdu.is_prepared(self.root):
+            # prep
+            print("Preparing dataset {} in {}".format(self.NAME, self.root))
+
+            datadir = self.datadir
+            if not os.path.exists(datadir):
+                path = os.path.join(self.root, self.FILES[0])
+                if not os.path.exists(path) or not os.path.getsize(path)==self.SIZES[0]:
+                    import academictorrents as at
+                    atpath = at.get(self.AT_HASH, datastore=self.root)
+                    assert atpath == path
+
+                print("Extracting {} to {}".format(path, datadir))
+                os.makedirs(datadir, exist_ok=True)
+                with tarfile.open(path, "r:") as tar:
+                    tar.extractall(path=datadir)
+
+                vspath = os.path.join(self.root, self.FILES[1])
+                if not os.path.exists(vspath) or not os.path.getsize(vspath)==self.SIZES[1]:
+                    download(self.VS_URL, vspath)
+
+                with open(vspath, "r") as f:
+                    synset_dict = f.read().splitlines()
+                    synset_dict = dict(line.split() for line in synset_dict)
+
+                print("Reorganizing into synset folders")
+                synsets = np.unique(list(synset_dict.values()))
+                for s in synsets:
+                    os.makedirs(os.path.join(datadir, s), exist_ok=True)
+                for k, v in synset_dict.items():
+                    src = os.path.join(datadir, k)
+                    dst = os.path.join(datadir, v)
+                    shutil.move(src, dst)
+
+            filelist = glob.glob(os.path.join(datadir, "**", "*.JPEG"))
+            filelist = [os.path.relpath(p, start=datadir) for p in filelist]
+            filelist = sorted(filelist)
+            filelist = "\n".join(filelist)+"\n"
+            with open(self.txt_filelist, "w") as f:
+                f.write(filelist)
+
+            tdu.mark_prepared(self.root)
+
+
+
+class ImageNetSR(Dataset):
+    def __init__(self, size=None,
+                 degradation=None, downscale_f=4, min_crop_f=0.5, max_crop_f=1.,
+                 random_crop=True):
+        """
+        Imagenet Superresolution Dataloader
+        Performs following ops in order:
+        1.  crops a crop of size s from image either as random or center crop
+        2.  resizes crop to size with cv2.area_interpolation
+        3.  degrades resized crop with degradation_fn
+
+        :param size: resizing to size after cropping
+        :param degradation: degradation_fn, e.g. cv_bicubic or bsrgan_light
+        :param downscale_f: Low Resolution Downsample factor
+        :param min_crop_f: determines crop size s,
+          where s = c * min_img_side_len with c sampled from interval (min_crop_f, max_crop_f)
+        :param max_crop_f: ""
+        :param data_root:
+        :param random_crop:
+        """
+        self.base = self.get_base()
+        assert size
+        assert (size / downscale_f).is_integer()
+        self.size = size
+        self.LR_size = int(size / downscale_f)
+        self.min_crop_f = min_crop_f
+        self.max_crop_f = max_crop_f
+        assert(max_crop_f <= 1.)
+        self.center_crop = not random_crop
+
+        self.image_rescaler = albumentations.SmallestMaxSize(max_size=size, interpolation=cv2.INTER_AREA)
+
+        self.pil_interpolation = False # gets reset later if incase interp_op is from pillow
+
+        if degradation == "bsrgan":
+            self.degradation_process = partial(degradation_fn_bsr, sf=downscale_f)
+
+        elif degradation == "bsrgan_light":
+            self.degradation_process = partial(degradation_fn_bsr_light, sf=downscale_f)
+
+        else:
+            interpolation_fn = {
+            "cv_nearest": cv2.INTER_NEAREST,
+            "cv_bilinear": cv2.INTER_LINEAR,
+            "cv_bicubic": cv2.INTER_CUBIC,
+            "cv_area": cv2.INTER_AREA,
+            "cv_lanczos": cv2.INTER_LANCZOS4,
+            "pil_nearest": PIL.Image.NEAREST,
+            "pil_bilinear": PIL.Image.BILINEAR,
+            "pil_bicubic": PIL.Image.BICUBIC,
+            "pil_box": PIL.Image.BOX,
+            "pil_hamming": PIL.Image.HAMMING,
+            "pil_lanczos": PIL.Image.LANCZOS,
+            }[degradation]
+
+            self.pil_interpolation = degradation.startswith("pil_")
+
+            if self.pil_interpolation:
+                self.degradation_process = partial(TF.resize, size=self.LR_size, interpolation=interpolation_fn)
+
+            else:
+                self.degradation_process = albumentations.SmallestMaxSize(max_size=self.LR_size,
+                                                                          interpolation=interpolation_fn)
+
+    def __len__(self):
+        return len(self.base)
+
+    def __getitem__(self, i):
+        example = self.base[i]
+        image = Image.open(example["file_path_"])
+
+        if not image.mode == "RGB":
+            image = image.convert("RGB")
+
+        image = np.array(image).astype(np.uint8)
+
+        min_side_len = min(image.shape[:2])
+        crop_side_len = min_side_len * np.random.uniform(self.min_crop_f, self.max_crop_f, size=None)
+        crop_side_len = int(crop_side_len)
+
+        if self.center_crop:
+            self.cropper = albumentations.CenterCrop(height=crop_side_len, width=crop_side_len)
+
+        else:
+            self.cropper = albumentations.RandomCrop(height=crop_side_len, width=crop_side_len)
+
+        image = self.cropper(image=image)["image"]
+        image = self.image_rescaler(image=image)["image"]
+
+        if self.pil_interpolation:
+            image_pil = PIL.Image.fromarray(image)
+            LR_image = self.degradation_process(image_pil)
+            LR_image = np.array(LR_image).astype(np.uint8)
+
+        else:
+            LR_image = self.degradation_process(image=image)["image"]
+
+        example["image"] = (image/127.5 - 1.0).astype(np.float32)
+        example["LR_image"] = (LR_image/127.5 - 1.0).astype(np.float32)
+
+        return example
+
+
+class ImageNetSRTrain(ImageNetSR):
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+
+    def get_base(self):
+        with open("data/imagenet_train_hr_indices.p", "rb") as f:
+            indices = pickle.load(f)
+        dset = ImageNetTrain(process_images=False,)
+        return Subset(dset, indices)
+
+
+class ImageNetSRValidation(ImageNetSR):
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+
+    def get_base(self):
+        with open("data/imagenet_val_hr_indices.p", "rb") as f:
+            indices = pickle.load(f)
+        dset = ImageNetValidation(process_images=False,)
+        return Subset(dset, indices)

examples/tutorial/stable_diffusion/ldm/data/lsun.py

@@ -0,0 +1,92 @@
+import os
+import numpy as np
+import PIL
+from PIL import Image
+from torch.utils.data import Dataset
+from torchvision import transforms
+
+
+class LSUNBase(Dataset):
+    def __init__(self,
+                 txt_file,
+                 data_root,
+                 size=None,
+                 interpolation="bicubic",
+                 flip_p=0.5
+                 ):
+        self.data_paths = txt_file
+        self.data_root = data_root
+        with open(self.data_paths, "r") as f:
+            self.image_paths = f.read().splitlines()
+        self._length = len(self.image_paths)
+        self.labels = {
+            "relative_file_path_": [l for l in self.image_paths],
+            "file_path_": [os.path.join(self.data_root, l)
+                           for l in self.image_paths],
+        }
+
+        self.size = size
+        self.interpolation = {"linear": PIL.Image.LINEAR,
+                              "bilinear": PIL.Image.BILINEAR,
+                              "bicubic": PIL.Image.BICUBIC,
+                              "lanczos": PIL.Image.LANCZOS,
+                              }[interpolation]
+        self.flip = transforms.RandomHorizontalFlip(p=flip_p)
+
+    def __len__(self):
+        return self._length
+
+    def __getitem__(self, i):
+        example = dict((k, self.labels[k][i]) for k in self.labels)
+        image = Image.open(example["file_path_"])
+        if not image.mode == "RGB":
+            image = image.convert("RGB")
+
+        # default to score-sde preprocessing
+        img = np.array(image).astype(np.uint8)
+        crop = min(img.shape[0], img.shape[1])
+        h, w, = img.shape[0], img.shape[1]
+        img = img[(h - crop) // 2:(h + crop) // 2,
+              (w - crop) // 2:(w + crop) // 2]
+
+        image = Image.fromarray(img)
+        if self.size is not None:
+            image = image.resize((self.size, self.size), resample=self.interpolation)
+
+        image = self.flip(image)
+        image = np.array(image).astype(np.uint8)
+        example["image"] = (image / 127.5 - 1.0).astype(np.float32)
+        return example
+
+
+class LSUNChurchesTrain(LSUNBase):
+    def __init__(self, **kwargs):
+        super().__init__(txt_file="data/lsun/church_outdoor_train.txt", data_root="data/lsun/churches", **kwargs)
+
+
+class LSUNChurchesValidation(LSUNBase):
+    def __init__(self, flip_p=0., **kwargs):
+        super().__init__(txt_file="data/lsun/church_outdoor_val.txt", data_root="data/lsun/churches",
+                         flip_p=flip_p, **kwargs)
+
+
+class LSUNBedroomsTrain(LSUNBase):
+    def __init__(self, **kwargs):
+        super().__init__(txt_file="data/lsun/bedrooms_train.txt", data_root="data/lsun/bedrooms", **kwargs)
+
+
+class LSUNBedroomsValidation(LSUNBase):
+    def __init__(self, flip_p=0.0, **kwargs):
+        super().__init__(txt_file="data/lsun/bedrooms_val.txt", data_root="data/lsun/bedrooms",
+                         flip_p=flip_p, **kwargs)
+
+
+class LSUNCatsTrain(LSUNBase):
+    def __init__(self, **kwargs):
+        super().__init__(txt_file="data/lsun/cat_train.txt", data_root="data/lsun/cats", **kwargs)
+
+
+class LSUNCatsValidation(LSUNBase):
+    def __init__(self, flip_p=0., **kwargs):
+        super().__init__(txt_file="data/lsun/cat_val.txt", data_root="data/lsun/cats",
+                         flip_p=flip_p, **kwargs)

examples/tutorial/stable_diffusion/ldm/lr_scheduler.py

@@ -0,0 +1,98 @@
+import numpy as np
+
+
+class LambdaWarmUpCosineScheduler:
+    """
+    note: use with a base_lr of 1.0
+    """
+    def __init__(self, warm_up_steps, lr_min, lr_max, lr_start, max_decay_steps, verbosity_interval=0):
+        self.lr_warm_up_steps = warm_up_steps
+        self.lr_start = lr_start
+        self.lr_min = lr_min
+        self.lr_max = lr_max
+        self.lr_max_decay_steps = max_decay_steps
+        self.last_lr = 0.
+        self.verbosity_interval = verbosity_interval
+
+    def schedule(self, n, **kwargs):
+        if self.verbosity_interval > 0:
+            if n % self.verbosity_interval == 0: print(f"current step: {n}, recent lr-multiplier: {self.last_lr}")
+        if n < self.lr_warm_up_steps:
+            lr = (self.lr_max - self.lr_start) / self.lr_warm_up_steps * n + self.lr_start
+            self.last_lr = lr
+            return lr
+        else:
+            t = (n - self.lr_warm_up_steps) / (self.lr_max_decay_steps - self.lr_warm_up_steps)
+            t = min(t, 1.0)
+            lr = self.lr_min + 0.5 * (self.lr_max - self.lr_min) * (
+                    1 + np.cos(t * np.pi))
+            self.last_lr = lr
+            return lr
+
+    def __call__(self, n, **kwargs):
+        return self.schedule(n,**kwargs)
+
+
+class LambdaWarmUpCosineScheduler2:
+    """
+    supports repeated iterations, configurable via lists
+    note: use with a base_lr of 1.0.
+    """
+    def __init__(self, warm_up_steps, f_min, f_max, f_start, cycle_lengths, verbosity_interval=0):
+        assert len(warm_up_steps) == len(f_min) == len(f_max) == len(f_start) == len(cycle_lengths)
+        self.lr_warm_up_steps = warm_up_steps
+        self.f_start = f_start
+        self.f_min = f_min
+        self.f_max = f_max
+        self.cycle_lengths = cycle_lengths
+        self.cum_cycles = np.cumsum([0] + list(self.cycle_lengths))
+        self.last_f = 0.
+        self.verbosity_interval = verbosity_interval
+
+    def find_in_interval(self, n):
+        interval = 0
+        for cl in self.cum_cycles[1:]:
+            if n <= cl:
+                return interval
+            interval += 1
+
+    def schedule(self, n, **kwargs):
+        cycle = self.find_in_interval(n)
+        n = n - self.cum_cycles[cycle]
+        if self.verbosity_interval > 0:
+            if n % self.verbosity_interval == 0: print(f"current step: {n}, recent lr-multiplier: {self.last_f}, "
+                                                       f"current cycle {cycle}")
+        if n < self.lr_warm_up_steps[cycle]:
+            f = (self.f_max[cycle] - self.f_start[cycle]) / self.lr_warm_up_steps[cycle] * n + self.f_start[cycle]
+            self.last_f = f
+            return f
+        else:
+            t = (n - self.lr_warm_up_steps[cycle]) / (self.cycle_lengths[cycle] - self.lr_warm_up_steps[cycle])
+            t = min(t, 1.0)
+            f = self.f_min[cycle] + 0.5 * (self.f_max[cycle] - self.f_min[cycle]) * (
+                    1 + np.cos(t * np.pi))
+            self.last_f = f
+            return f
+
+    def __call__(self, n, **kwargs):
+        return self.schedule(n, **kwargs)
+
+
+class LambdaLinearScheduler(LambdaWarmUpCosineScheduler2):
+
+    def schedule(self, n, **kwargs):
+        cycle = self.find_in_interval(n)
+        n = n - self.cum_cycles[cycle]
+        if self.verbosity_interval > 0:
+            if n % self.verbosity_interval == 0: print(f"current step: {n}, recent lr-multiplier: {self.last_f}, "
+                                                       f"current cycle {cycle}")
+
+        if n < self.lr_warm_up_steps[cycle]:
+            f = (self.f_max[cycle] - self.f_start[cycle]) / self.lr_warm_up_steps[cycle] * n + self.f_start[cycle]
+            self.last_f = f
+            return f
+        else:
+            f = self.f_min[cycle] + (self.f_max[cycle] - self.f_min[cycle]) * (self.cycle_lengths[cycle] - n) / (self.cycle_lengths[cycle])
+            self.last_f = f
+            return f
+

examples/tutorial/stable_diffusion/ldm/models/autoencoder.py

@@ -0,0 +1,544 @@
+import torch
+import pytorch_lightning as pl
+import torch.nn.functional as F
+from contextlib import contextmanager
+
+from taming.modules.vqvae.quantize import VectorQuantizer2 as VectorQuantizer
+
+from ldm.modules.diffusionmodules.model import Encoder, Decoder
+from ldm.modules.distributions.distributions import DiagonalGaussianDistribution
+
+from ldm.util import instantiate_from_config
+
+
+class VQModel(pl.LightningModule):
+    def __init__(self,
+                 ddconfig,
+                 lossconfig,
+                 n_embed,
+                 embed_dim,
+                 ckpt_path=None,
+                 ignore_keys=[],
+                 image_key="image",
+                 colorize_nlabels=None,
+                 monitor=None,
+                 batch_resize_range=None,
+                 scheduler_config=None,
+                 lr_g_factor=1.0,
+                 remap=None,
+                 sane_index_shape=False, # tell vector quantizer to return indices as bhw
+                 use_ema=False
+                 ):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.n_embed = n_embed
+        self.image_key = image_key
+        self.encoder = Encoder(**ddconfig)
+        self.decoder = Decoder(**ddconfig)
+        self.loss = instantiate_from_config(lossconfig)
+        self.quantize = VectorQuantizer(n_embed, embed_dim, beta=0.25,
+                                        remap=remap,
+                                        sane_index_shape=sane_index_shape)
+        self.quant_conv = torch.nn.Conv2d(ddconfig["z_channels"], embed_dim, 1)
+        self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
+        if colorize_nlabels is not None:
+            assert type(colorize_nlabels)==int
+            self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1))
+        if monitor is not None:
+            self.monitor = monitor
+        self.batch_resize_range = batch_resize_range
+        if self.batch_resize_range is not None:
+            print(f"{self.__class__.__name__}: Using per-batch resizing in range {batch_resize_range}.")
+
+        self.use_ema = use_ema
+        if self.use_ema:
+            self.model_ema = LitEma(self)
+            print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
+
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
+        self.scheduler_config = scheduler_config
+        self.lr_g_factor = lr_g_factor
+
+    @contextmanager
+    def ema_scope(self, context=None):
+        if self.use_ema:
+            self.model_ema.store(self.parameters())
+            self.model_ema.copy_to(self)
+            if context is not None:
+                print(f"{context}: Switched to EMA weights")
+        try:
+            yield None
+        finally:
+            if self.use_ema:
+                self.model_ema.restore(self.parameters())
+                if context is not None:
+                    print(f"{context}: Restored training weights")
+
+    def init_from_ckpt(self, path, ignore_keys=list()):
+        sd = torch.load(path, map_location="cpu")["state_dict"]
+        keys = list(sd.keys())
+        for k in keys:
+            for ik in ignore_keys:
+                if k.startswith(ik):
+                    print("Deleting key {} from state_dict.".format(k))
+                    del sd[k]
+        missing, unexpected = self.load_state_dict(sd, strict=False)
+        print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
+        if len(missing) > 0:
+            print(f"Missing Keys: {missing}")
+            print(f"Unexpected Keys: {unexpected}")
+
+    def on_train_batch_end(self, *args, **kwargs):
+        if self.use_ema:
+            self.model_ema(self)
+
+    def encode(self, x):
+        h = self.encoder(x)
+        h = self.quant_conv(h)
+        quant, emb_loss, info = self.quantize(h)
+        return quant, emb_loss, info
+
+    def encode_to_prequant(self, x):
+        h = self.encoder(x)
+        h = self.quant_conv(h)
+        return h
+
+    def decode(self, quant):
+        quant = self.post_quant_conv(quant)
+        dec = self.decoder(quant)
+        return dec
+
+    def decode_code(self, code_b):
+        quant_b = self.quantize.embed_code(code_b)
+        dec = self.decode(quant_b)
+        return dec
+
+    def forward(self, input, return_pred_indices=False):
+        quant, diff, (_,_,ind) = self.encode(input)
+        dec = self.decode(quant)
+        if return_pred_indices:
+            return dec, diff, ind
+        return dec, diff
+
+    def get_input(self, batch, k):
+        x = batch[k]
+        if len(x.shape) == 3:
+            x = x[..., None]
+        x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format).float()
+        if self.batch_resize_range is not None:
+            lower_size = self.batch_resize_range[0]
+            upper_size = self.batch_resize_range[1]
+            if self.global_step <= 4:
+                # do the first few batches with max size to avoid later oom
+                new_resize = upper_size
+            else:
+                new_resize = np.random.choice(np.arange(lower_size, upper_size+16, 16))
+            if new_resize != x.shape[2]:
+                x = F.interpolate(x, size=new_resize, mode="bicubic")
+            x = x.detach()
+        return x
+
+    def training_step(self, batch, batch_idx, optimizer_idx):
+        # https://github.com/pytorch/pytorch/issues/37142
+        # try not to fool the heuristics
+        x = self.get_input(batch, self.image_key)
+        xrec, qloss, ind = self(x, return_pred_indices=True)
+
+        if optimizer_idx == 0:
+            # autoencode
+            aeloss, log_dict_ae = self.loss(qloss, x, xrec, optimizer_idx, self.global_step,
+                                            last_layer=self.get_last_layer(), split="train",
+                                            predicted_indices=ind)
+
+            self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=True)
+            return aeloss
+
+        if optimizer_idx == 1:
+            # discriminator
+            discloss, log_dict_disc = self.loss(qloss, x, xrec, optimizer_idx, self.global_step,
+                                            last_layer=self.get_last_layer(), split="train")
+            self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=True)
+            return discloss
+
+    def validation_step(self, batch, batch_idx):
+        log_dict = self._validation_step(batch, batch_idx)
+        with self.ema_scope():
+            log_dict_ema = self._validation_step(batch, batch_idx, suffix="_ema")
+        return log_dict
+
+    def _validation_step(self, batch, batch_idx, suffix=""):
+        x = self.get_input(batch, self.image_key)
+        xrec, qloss, ind = self(x, return_pred_indices=True)
+        aeloss, log_dict_ae = self.loss(qloss, x, xrec, 0,
+                                        self.global_step,
+                                        last_layer=self.get_last_layer(),
+                                        split="val"+suffix,
+                                        predicted_indices=ind
+                                        )
+
+        discloss, log_dict_disc = self.loss(qloss, x, xrec, 1,
+                                            self.global_step,
+                                            last_layer=self.get_last_layer(),
+                                            split="val"+suffix,
+                                            predicted_indices=ind
+                                            )
+        rec_loss = log_dict_ae[f"val{suffix}/rec_loss"]
+        self.log(f"val{suffix}/rec_loss", rec_loss,
+                   prog_bar=True, logger=True, on_step=False, on_epoch=True, sync_dist=True)
+        self.log(f"val{suffix}/aeloss", aeloss,
+                   prog_bar=True, logger=True, on_step=False, on_epoch=True, sync_dist=True)
+        if version.parse(pl.__version__) >= version.parse('1.4.0'):
+            del log_dict_ae[f"val{suffix}/rec_loss"]
+        self.log_dict(log_dict_ae)
+        self.log_dict(log_dict_disc)
+        return self.log_dict
+
+    def configure_optimizers(self):
+        lr_d = self.learning_rate
+        lr_g = self.lr_g_factor*self.learning_rate
+        print("lr_d", lr_d)
+        print("lr_g", lr_g)
+        opt_ae = torch.optim.Adam(list(self.encoder.parameters())+
+                                  list(self.decoder.parameters())+
+                                  list(self.quantize.parameters())+
+                                  list(self.quant_conv.parameters())+
+                                  list(self.post_quant_conv.parameters()),
+                                  lr=lr_g, betas=(0.5, 0.9))
+        opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(),
+                                    lr=lr_d, betas=(0.5, 0.9))
+
+        if self.scheduler_config is not None:
+            scheduler = instantiate_from_config(self.scheduler_config)
+
+            print("Setting up LambdaLR scheduler...")
+            scheduler = [
+                {
+                    'scheduler': LambdaLR(opt_ae, lr_lambda=scheduler.schedule),
+                    'interval': 'step',
+                    'frequency': 1
+                },
+                {
+                    'scheduler': LambdaLR(opt_disc, lr_lambda=scheduler.schedule),
+                    'interval': 'step',
+                    'frequency': 1
+                },
+            ]
+            return [opt_ae, opt_disc], scheduler
+        return [opt_ae, opt_disc], []
+
+    def get_last_layer(self):
+        return self.decoder.conv_out.weight
+
+    def log_images(self, batch, only_inputs=False, plot_ema=False, **kwargs):
+        log = dict()
+        x = self.get_input(batch, self.image_key)
+        x = x.to(self.device)
+        if only_inputs:
+            log["inputs"] = x
+            return log
+        xrec, _ = self(x)
+        if x.shape[1] > 3:
+            # colorize with random projection
+            assert xrec.shape[1] > 3
+            x = self.to_rgb(x)
+            xrec = self.to_rgb(xrec)
+        log["inputs"] = x
+        log["reconstructions"] = xrec
+        if plot_ema:
+            with self.ema_scope():
+                xrec_ema, _ = self(x)
+                if x.shape[1] > 3: xrec_ema = self.to_rgb(xrec_ema)
+                log["reconstructions_ema"] = xrec_ema
+        return log
+
+    def to_rgb(self, x):
+        assert self.image_key == "segmentation"
+        if not hasattr(self, "colorize"):
+            self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x))
+        x = F.conv2d(x, weight=self.colorize)
+        x = 2.*(x-x.min())/(x.max()-x.min()) - 1.
+        return x
+
+
+class VQModelInterface(VQModel):
+    def __init__(self, embed_dim, *args, **kwargs):
+        super().__init__(embed_dim=embed_dim, *args, **kwargs)
+        self.embed_dim = embed_dim
+
+    def encode(self, x):
+        h = self.encoder(x)
+        h = self.quant_conv(h)
+        return h
+
+    def decode(self, h, force_not_quantize=False):
+        # also go through quantization layer
+        if not force_not_quantize:
+            quant, emb_loss, info = self.quantize(h)
+        else:
+            quant = h
+        quant = self.post_quant_conv(quant)
+        dec = self.decoder(quant)
+        return dec
+
+
+class AutoencoderKL(pl.LightningModule):
+    def __init__(self,
+                 ddconfig,
+                 lossconfig,
+                 embed_dim,
+                 ckpt_path=None,
+                 ignore_keys=[],
+                 image_key="image",
+                 colorize_nlabels=None,
+                 monitor=None,
+                 from_pretrained: str=None
+                 ):
+        super().__init__()
+        self.image_key = image_key
+        self.encoder = Encoder(**ddconfig)
+        self.decoder = Decoder(**ddconfig)
+        self.loss = instantiate_from_config(lossconfig)
+        assert ddconfig["double_z"]
+        self.quant_conv = torch.nn.Conv2d(2*ddconfig["z_channels"], 2*embed_dim, 1)
+        self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
+        self.embed_dim = embed_dim
+        if colorize_nlabels is not None:
+            assert type(colorize_nlabels)==int
+            self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1))
+        if monitor is not None:
+            self.monitor = monitor
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
+        from diffusers.modeling_utils import load_state_dict
+        if from_pretrained is not None:
+            state_dict = load_state_dict(from_pretrained)
+            self._load_pretrained_model(state_dict)
+
+    def _state_key_mapping(self, state_dict: dict):
+        import re
+        res_dict = {}
+        key_list = state_dict.keys()
+        key_str = " ".join(key_list)
+        up_block_pattern = re.compile('upsamplers')
+        p1 = re.compile('mid.block_[0-9]')
+        p2 = re.compile('decoder.up.[0-9]')
+        up_blocks_count = int(len(re.findall(up_block_pattern, key_str)) / 2 + 1)
+        for key_, val_ in state_dict.items():
+            key_ = key_.replace("up_blocks", "up").replace("down_blocks", "down").replace('resnets', 'block')\
+                .replace('mid_block', 'mid').replace("mid.block.", "mid.block_")\
+                .replace('mid.attentions.0.key', 'mid.attn_1.k')\
+                .replace('mid.attentions.0.query', 'mid.attn_1.q') \
+                .replace('mid.attentions.0.value', 'mid.attn_1.v') \
+                .replace('mid.attentions.0.group_norm', 'mid.attn_1.norm') \
+                .replace('mid.attentions.0.proj_attn', 'mid.attn_1.proj_out')\
+                .replace('upsamplers.0', 'upsample')\
+                .replace('downsamplers.0', 'downsample')\
+                .replace('conv_shortcut', 'nin_shortcut')\
+                .replace('conv_norm_out', 'norm_out')
+
+            mid_list = re.findall(p1, key_)
+            if len(mid_list) != 0:
+                mid_str = mid_list[0]
+                mid_id = int(mid_str[-1]) + 1
+                key_ = key_.replace(mid_str, mid_str[:-1] + str(mid_id))
+
+            up_list = re.findall(p2, key_)
+            if len(up_list) != 0:
+                up_str = up_list[0]
+                up_id = up_blocks_count - 1 -int(up_str[-1])
+                key_ = key_.replace(up_str, up_str[:-1] + str(up_id))
+            res_dict[key_] = val_
+        return res_dict
+
+    def _load_pretrained_model(self, state_dict, ignore_mismatched_sizes=False):
+        state_dict = self._state_key_mapping(state_dict)
+        model_state_dict = self.state_dict()
+        loaded_keys = [k for k in state_dict.keys()]
+        expected_keys = list(model_state_dict.keys())
+        original_loaded_keys = loaded_keys
+        missing_keys = list(set(expected_keys) - set(loaded_keys))
+        unexpected_keys = list(set(loaded_keys) - set(expected_keys))
+
+        def _find_mismatched_keys(
+            state_dict,
+            model_state_dict,
+            loaded_keys,
+            ignore_mismatched_sizes,
+        ):
+            mismatched_keys = []
+            if ignore_mismatched_sizes:
+                for checkpoint_key in loaded_keys:
+                    model_key = checkpoint_key
+
+                    if (
+                        model_key in model_state_dict
+                        and state_dict[checkpoint_key].shape != model_state_dict[model_key].shape
+                    ):
+                        mismatched_keys.append(
+                            (checkpoint_key, state_dict[checkpoint_key].shape, model_state_dict[model_key].shape)
+                        )
+                        del state_dict[checkpoint_key]
+            return mismatched_keys
+        if state_dict is not None:
+            # Whole checkpoint
+            mismatched_keys = _find_mismatched_keys(
+                state_dict,
+                model_state_dict,
+                original_loaded_keys,
+                ignore_mismatched_sizes,
+            )
+            error_msgs = self._load_state_dict_into_model(state_dict)
+        return missing_keys, unexpected_keys, mismatched_keys, error_msgs
+
+    def _load_state_dict_into_model(self, state_dict):
+        # Convert old format to new format if needed from a PyTorch state_dict
+        # copy state_dict so _load_from_state_dict can modify it
+        state_dict = state_dict.copy()
+        error_msgs = []
+
+        # PyTorch's `_load_from_state_dict` does not copy parameters in a module's descendants
+        # so we need to apply the function recursively.
+        def load(module: torch.nn.Module, prefix=""):
+            args = (state_dict, prefix, {}, True, [], [], error_msgs)
+            module._load_from_state_dict(*args)
+
+            for name, child in module._modules.items():
+                if child is not None:
+                    load(child, prefix + name + ".")
+
+        load(self)
+
+        return error_msgs
+
+    def init_from_ckpt(self, path, ignore_keys=list()):
+        sd = torch.load(path, map_location="cpu")["state_dict"]
+        keys = list(sd.keys())
+        for k in keys:
+            for ik in ignore_keys:
+                if k.startswith(ik):
+                    print("Deleting key {} from state_dict.".format(k))
+                    del sd[k]
+        self.load_state_dict(sd, strict=False)
+        print(f"Restored from {path}")
+
+    def encode(self, x):
+        h = self.encoder(x)
+        moments = self.quant_conv(h)
+        posterior = DiagonalGaussianDistribution(moments)
+        return posterior
+
+    def decode(self, z):
+        z = self.post_quant_conv(z)
+        dec = self.decoder(z)
+        return dec
+
+    def forward(self, input, sample_posterior=True):
+        posterior = self.encode(input)
+        if sample_posterior:
+            z = posterior.sample()
+        else:
+            z = posterior.mode()
+        dec = self.decode(z)
+        return dec, posterior
+
+    def get_input(self, batch, k):
+        x = batch[k]
+        if len(x.shape) == 3:
+            x = x[..., None]
+        x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format).float()
+        return x
+
+    def training_step(self, batch, batch_idx, optimizer_idx):
+        inputs = self.get_input(batch, self.image_key)
+        reconstructions, posterior = self(inputs)
+
+        if optimizer_idx == 0:
+            # train encoder+decoder+logvar
+            aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step,
+                                            last_layer=self.get_last_layer(), split="train")
+            self.log("aeloss", aeloss, prog_bar=True, logger=True, on_step=True, on_epoch=True)
+            self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=False)
+            return aeloss
+
+        if optimizer_idx == 1:
+            # train the discriminator
+            discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step,
+                                                last_layer=self.get_last_layer(), split="train")
+
+            self.log("discloss", discloss, prog_bar=True, logger=True, on_step=True, on_epoch=True)
+            self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=False)
+            return discloss
+
+    def validation_step(self, batch, batch_idx):
+        inputs = self.get_input(batch, self.image_key)
+        reconstructions, posterior = self(inputs)
+        aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, 0, self.global_step,
+                                        last_layer=self.get_last_layer(), split="val")
+
+        discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, 1, self.global_step,
+                                            last_layer=self.get_last_layer(), split="val")
+
+        self.log("val/rec_loss", log_dict_ae["val/rec_loss"])
+        self.log_dict(log_dict_ae)
+        self.log_dict(log_dict_disc)
+        return self.log_dict
+
+    def configure_optimizers(self):
+        lr = self.learning_rate
+        opt_ae = torch.optim.Adam(list(self.encoder.parameters())+
+                                  list(self.decoder.parameters())+
+                                  list(self.quant_conv.parameters())+
+                                  list(self.post_quant_conv.parameters()),
+                                  lr=lr, betas=(0.5, 0.9))
+        opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(),
+                                    lr=lr, betas=(0.5, 0.9))
+        return [opt_ae, opt_disc], []
+
+    def get_last_layer(self):
+        return self.decoder.conv_out.weight
+
+    @torch.no_grad()
+    def log_images(self, batch, only_inputs=False, **kwargs):
+        log = dict()
+        x = self.get_input(batch, self.image_key)
+        x = x.to(self.device)
+        if not only_inputs:
+            xrec, posterior = self(x)
+            if x.shape[1] > 3:
+                # colorize with random projection
+                assert xrec.shape[1] > 3
+                x = self.to_rgb(x)
+                xrec = self.to_rgb(xrec)
+            log["samples"] = self.decode(torch.randn_like(posterior.sample()))
+            log["reconstructions"] = xrec
+        log["inputs"] = x
+        return log
+
+    def to_rgb(self, x):
+        assert self.image_key == "segmentation"
+        if not hasattr(self, "colorize"):
+            self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x))
+        x = F.conv2d(x, weight=self.colorize)
+        x = 2.*(x-x.min())/(x.max()-x.min()) - 1.
+        return x
+
+
+class IdentityFirstStage(torch.nn.Module):
+    def __init__(self, *args, vq_interface=False, **kwargs):
+        self.vq_interface = vq_interface  # TODO: Should be true by default but check to not break older stuff
+        super().__init__()
+
+    def encode(self, x, *args, **kwargs):
+        return x
+
+    def decode(self, x, *args, **kwargs):
+        return x
+
+    def quantize(self, x, *args, **kwargs):
+        if self.vq_interface:
+            return x, None, [None, None, None]
+        return x
+
+    def forward(self, x, *args, **kwargs):
+        return x

examples/tutorial/stable_diffusion/ldm/models/diffusion/classifier.py

@@ -0,0 +1,267 @@
+import os
+import torch
+import pytorch_lightning as pl
+from omegaconf import OmegaConf
+from torch.nn import functional as F
+from torch.optim import AdamW
+from torch.optim.lr_scheduler import LambdaLR
+from copy import deepcopy
+from einops import rearrange
+from glob import glob
+from natsort import natsorted
+
+from ldm.modules.diffusionmodules.openaimodel import EncoderUNetModel, UNetModel
+from ldm.util import log_txt_as_img, default, ismap, instantiate_from_config
+
+__models__ = {
+    'class_label': EncoderUNetModel,
+    'segmentation': UNetModel
+}
+
+
+def disabled_train(self, mode=True):
+    """Overwrite model.train with this function to make sure train/eval mode
+    does not change anymore."""
+    return self
+
+
+class NoisyLatentImageClassifier(pl.LightningModule):
+
+    def __init__(self,
+                 diffusion_path,
+                 num_classes,
+                 ckpt_path=None,
+                 pool='attention',
+                 label_key=None,
+                 diffusion_ckpt_path=None,
+                 scheduler_config=None,
+                 weight_decay=1.e-2,
+                 log_steps=10,
+                 monitor='val/loss',
+                 *args,
+                 **kwargs):
+        super().__init__(*args, **kwargs)
+        self.num_classes = num_classes
+        # get latest config of diffusion model
+        diffusion_config = natsorted(glob(os.path.join(diffusion_path, 'configs', '*-project.yaml')))[-1]
+        self.diffusion_config = OmegaConf.load(diffusion_config).model
+        self.diffusion_config.params.ckpt_path = diffusion_ckpt_path
+        self.load_diffusion()
+
+        self.monitor = monitor
+        self.numd = self.diffusion_model.first_stage_model.encoder.num_resolutions - 1
+        self.log_time_interval = self.diffusion_model.num_timesteps // log_steps
+        self.log_steps = log_steps
+
+        self.label_key = label_key if not hasattr(self.diffusion_model, 'cond_stage_key') \
+            else self.diffusion_model.cond_stage_key
+
+        assert self.label_key is not None, 'label_key neither in diffusion model nor in model.params'
+
+        if self.label_key not in __models__:
+            raise NotImplementedError()
+
+        self.load_classifier(ckpt_path, pool)
+
+        self.scheduler_config = scheduler_config
+        self.use_scheduler = self.scheduler_config is not None
+        self.weight_decay = weight_decay
+
+    def init_from_ckpt(self, path, ignore_keys=list(), only_model=False):
+        sd = torch.load(path, map_location="cpu")
+        if "state_dict" in list(sd.keys()):
+            sd = sd["state_dict"]
+        keys = list(sd.keys())
+        for k in keys:
+            for ik in ignore_keys:
+                if k.startswith(ik):
+                    print("Deleting key {} from state_dict.".format(k))
+                    del sd[k]
+        missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict(
+            sd, strict=False)
+        print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
+        if len(missing) > 0:
+            print(f"Missing Keys: {missing}")
+        if len(unexpected) > 0:
+            print(f"Unexpected Keys: {unexpected}")
+
+    def load_diffusion(self):
+        model = instantiate_from_config(self.diffusion_config)
+        self.diffusion_model = model.eval()
+        self.diffusion_model.train = disabled_train
+        for param in self.diffusion_model.parameters():
+            param.requires_grad = False
+
+    def load_classifier(self, ckpt_path, pool):
+        model_config = deepcopy(self.diffusion_config.params.unet_config.params)
+        model_config.in_channels = self.diffusion_config.params.unet_config.params.out_channels
+        model_config.out_channels = self.num_classes
+        if self.label_key == 'class_label':
+            model_config.pool = pool
+
+        self.model = __models__[self.label_key](**model_config)
+        if ckpt_path is not None:
+            print('#####################################################################')
+            print(f'load from ckpt "{ckpt_path}"')
+            print('#####################################################################')
+            self.init_from_ckpt(ckpt_path)
+
+    @torch.no_grad()
+    def get_x_noisy(self, x, t, noise=None):
+        noise = default(noise, lambda: torch.randn_like(x))
+        continuous_sqrt_alpha_cumprod = None
+        if self.diffusion_model.use_continuous_noise:
+            continuous_sqrt_alpha_cumprod = self.diffusion_model.sample_continuous_noise_level(x.shape[0], t + 1)
+            # todo: make sure t+1 is correct here
+
+        return self.diffusion_model.q_sample(x_start=x, t=t, noise=noise,
+                                             continuous_sqrt_alpha_cumprod=continuous_sqrt_alpha_cumprod)
+
+    def forward(self, x_noisy, t, *args, **kwargs):
+        return self.model(x_noisy, t)
+
+    @torch.no_grad()
+    def get_input(self, batch, k):
+        x = batch[k]
+        if len(x.shape) == 3:
+            x = x[..., None]
+        x = rearrange(x, 'b h w c -> b c h w')
+        x = x.to(memory_format=torch.contiguous_format).float()
+        return x
+
+    @torch.no_grad()
+    def get_conditioning(self, batch, k=None):
+        if k is None:
+            k = self.label_key
+        assert k is not None, 'Needs to provide label key'
+
+        targets = batch[k].to(self.device)
+
+        if self.label_key == 'segmentation':
+            targets = rearrange(targets, 'b h w c -> b c h w')
+            for down in range(self.numd):
+                h, w = targets.shape[-2:]
+                targets = F.interpolate(targets, size=(h // 2, w // 2), mode='nearest')
+
+            # targets = rearrange(targets,'b c h w -> b h w c')
+
+        return targets
+
+    def compute_top_k(self, logits, labels, k, reduction="mean"):
+        _, top_ks = torch.topk(logits, k, dim=1)
+        if reduction == "mean":
+            return (top_ks == labels[:, None]).float().sum(dim=-1).mean().item()
+        elif reduction == "none":
+            return (top_ks == labels[:, None]).float().sum(dim=-1)
+
+    def on_train_epoch_start(self):
+        # save some memory
+        self.diffusion_model.model.to('cpu')
+
+    @torch.no_grad()
+    def write_logs(self, loss, logits, targets):
+        log_prefix = 'train' if self.training else 'val'
+        log = {}
+        log[f"{log_prefix}/loss"] = loss.mean()
+        log[f"{log_prefix}/acc@1"] = self.compute_top_k(
+            logits, targets, k=1, reduction="mean"
+        )
+        log[f"{log_prefix}/acc@5"] = self.compute_top_k(
+            logits, targets, k=5, reduction="mean"
+        )
+
+        self.log_dict(log, prog_bar=False, logger=True, on_step=self.training, on_epoch=True)
+        self.log('loss', log[f"{log_prefix}/loss"], prog_bar=True, logger=False)
+        self.log('global_step', self.global_step, logger=False, on_epoch=False, prog_bar=True)
+        lr = self.optimizers().param_groups[0]['lr']
+        self.log('lr_abs', lr, on_step=True, logger=True, on_epoch=False, prog_bar=True)
+
+    def shared_step(self, batch, t=None):
+        x, *_ = self.diffusion_model.get_input(batch, k=self.diffusion_model.first_stage_key)
+        targets = self.get_conditioning(batch)
+        if targets.dim() == 4:
+            targets = targets.argmax(dim=1)
+        if t is None:
+            t = torch.randint(0, self.diffusion_model.num_timesteps, (x.shape[0],), device=self.device).long()
+        else:
+            t = torch.full(size=(x.shape[0],), fill_value=t, device=self.device).long()
+        x_noisy = self.get_x_noisy(x, t)
+        logits = self(x_noisy, t)
+
+        loss = F.cross_entropy(logits, targets, reduction='none')
+
+        self.write_logs(loss.detach(), logits.detach(), targets.detach())
+
+        loss = loss.mean()
+        return loss, logits, x_noisy, targets
+
+    def training_step(self, batch, batch_idx):
+        loss, *_ = self.shared_step(batch)
+        return loss
+
+    def reset_noise_accs(self):
+        self.noisy_acc = {t: {'acc@1': [], 'acc@5': []} for t in
+                          range(0, self.diffusion_model.num_timesteps, self.diffusion_model.log_every_t)}
+
+    def on_validation_start(self):
+        self.reset_noise_accs()
+
+    @torch.no_grad()
+    def validation_step(self, batch, batch_idx):
+        loss, *_ = self.shared_step(batch)
+
+        for t in self.noisy_acc:
+            _, logits, _, targets = self.shared_step(batch, t)
+            self.noisy_acc[t]['acc@1'].append(self.compute_top_k(logits, targets, k=1, reduction='mean'))
+            self.noisy_acc[t]['acc@5'].append(self.compute_top_k(logits, targets, k=5, reduction='mean'))
+
+        return loss
+
+    def configure_optimizers(self):
+        optimizer = AdamW(self.model.parameters(), lr=self.learning_rate, weight_decay=self.weight_decay)
+
+        if self.use_scheduler:
+            scheduler = instantiate_from_config(self.scheduler_config)
+
+            print("Setting up LambdaLR scheduler...")
+            scheduler = [
+                {
+                    'scheduler': LambdaLR(optimizer, lr_lambda=scheduler.schedule),
+                    'interval': 'step',
+                    'frequency': 1
+                }]
+            return [optimizer], scheduler
+
+        return optimizer
+
+    @torch.no_grad()
+    def log_images(self, batch, N=8, *args, **kwargs):
+        log = dict()
+        x = self.get_input(batch, self.diffusion_model.first_stage_key)
+        log['inputs'] = x
+
+        y = self.get_conditioning(batch)
+
+        if self.label_key == 'class_label':
+            y = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"])
+            log['labels'] = y
+
+        if ismap(y):
+            log['labels'] = self.diffusion_model.to_rgb(y)
+
+            for step in range(self.log_steps):
+                current_time = step * self.log_time_interval
+
+                _, logits, x_noisy, _ = self.shared_step(batch, t=current_time)
+
+                log[f'inputs@t{current_time}'] = x_noisy
+
+                pred = F.one_hot(logits.argmax(dim=1), num_classes=self.num_classes)
+                pred = rearrange(pred, 'b h w c -> b c h w')
+
+                log[f'pred@t{current_time}'] = self.diffusion_model.to_rgb(pred)
+
+        for key in log:
+            log[key] = log[key][:N]
+
+        return log

examples/tutorial/stable_diffusion/ldm/models/diffusion/ddim.py

@@ -0,0 +1,240 @@
+"""SAMPLING ONLY."""
+
+import torch
+import numpy as np
+from tqdm import tqdm
+from functools import partial
+
+from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like, \
+    extract_into_tensor
+
+
+class DDIMSampler(object):
+    def __init__(self, model, schedule="linear", **kwargs):
+        super().__init__()
+        self.model = model
+        self.ddpm_num_timesteps = model.num_timesteps
+        self.schedule = schedule
+
+    def register_buffer(self, name, attr):
+        if type(attr) == torch.Tensor:
+            if attr.device != torch.device("cuda"):
+                attr = attr.to(torch.device("cuda"))
+        setattr(self, name, attr)
+
+    def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True):
+        self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps,
+                                                  num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose)
+        alphas_cumprod = self.model.alphas_cumprod
+        assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep'
+        to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)
+
+        self.register_buffer('betas', to_torch(self.model.betas))
+        self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
+        self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev))
+
+        # calculations for diffusion q(x_t | x_{t-1}) and others
+        self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu())))
+        self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1)))
+
+        # ddim sampling parameters
+        ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(),
+                                                                                   ddim_timesteps=self.ddim_timesteps,
+                                                                                   eta=ddim_eta,verbose=verbose)
+        self.register_buffer('ddim_sigmas', ddim_sigmas)
+        self.register_buffer('ddim_alphas', ddim_alphas)
+        self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)
+        self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas))
+        sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
+            (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * (
+                        1 - self.alphas_cumprod / self.alphas_cumprod_prev))
+        self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps)
+
+    @torch.no_grad()
+    def sample(self,
+               S,
+               batch_size,
+               shape,
+               conditioning=None,
+               callback=None,
+               normals_sequence=None,
+               img_callback=None,
+               quantize_x0=False,
+               eta=0.,
+               mask=None,
+               x0=None,
+               temperature=1.,
+               noise_dropout=0.,
+               score_corrector=None,
+               corrector_kwargs=None,
+               verbose=True,
+               x_T=None,
+               log_every_t=100,
+               unconditional_guidance_scale=1.,
+               unconditional_conditioning=None,
+               # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
+               **kwargs
+               ):
+        if conditioning is not None:
+            if isinstance(conditioning, dict):
+                cbs = conditioning[list(conditioning.keys())[0]].shape[0]
+                if cbs != batch_size:
+                    print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
+            else:
+                if conditioning.shape[0] != batch_size:
+                    print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
+
+        self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose)
+        # sampling
+        C, H, W = shape
+        size = (batch_size, C, H, W)
+        print(f'Data shape for DDIM sampling is {size}, eta {eta}')
+
+        samples, intermediates = self.ddim_sampling(conditioning, size,
+                                                    callback=callback,
+                                                    img_callback=img_callback,
+                                                    quantize_denoised=quantize_x0,
+                                                    mask=mask, x0=x0,
+                                                    ddim_use_original_steps=False,
+                                                    noise_dropout=noise_dropout,
+                                                    temperature=temperature,
+                                                    score_corrector=score_corrector,
+                                                    corrector_kwargs=corrector_kwargs,
+                                                    x_T=x_T,
+                                                    log_every_t=log_every_t,
+                                                    unconditional_guidance_scale=unconditional_guidance_scale,
+                                                    unconditional_conditioning=unconditional_conditioning,
+                                                    )
+        return samples, intermediates
+
+    @torch.no_grad()
+    def ddim_sampling(self, cond, shape,
+                      x_T=None, ddim_use_original_steps=False,
+                      callback=None, timesteps=None, quantize_denoised=False,
+                      mask=None, x0=None, img_callback=None, log_every_t=100,
+                      temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
+                      unconditional_guidance_scale=1., unconditional_conditioning=None,):
+        device = self.model.betas.device
+        b = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=device)
+        else:
+            img = x_T
+
+        if timesteps is None:
+            timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps
+        elif timesteps is not None and not ddim_use_original_steps:
+            subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1
+            timesteps = self.ddim_timesteps[:subset_end]
+
+        intermediates = {'x_inter': [img], 'pred_x0': [img]}
+        time_range = reversed(range(0,timesteps)) if ddim_use_original_steps else np.flip(timesteps)
+        total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
+        print(f"Running DDIM Sampling with {total_steps} timesteps")
+
+        iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps)
+
+        for i, step in enumerate(iterator):
+            index = total_steps - i - 1
+            ts = torch.full((b,), step, device=device, dtype=torch.long)
+
+            if mask is not None:
+                assert x0 is not None
+                img_orig = self.model.q_sample(x0, ts)  # TODO: deterministic forward pass?
+                img = img_orig * mask + (1. - mask) * img
+            outs = self.p_sample_ddim(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps,
+                                      quantize_denoised=quantize_denoised, temperature=temperature,
+                                      noise_dropout=noise_dropout, score_corrector=score_corrector,
+                                      corrector_kwargs=corrector_kwargs,
+                                      unconditional_guidance_scale=unconditional_guidance_scale,
+                                      unconditional_conditioning=unconditional_conditioning)
+            img, pred_x0 = outs
+            if callback: callback(i)
+            if img_callback: img_callback(pred_x0, i)
+
+            if index % log_every_t == 0 or index == total_steps - 1:
+                intermediates['x_inter'].append(img)
+                intermediates['pred_x0'].append(pred_x0)
+
+        return img, intermediates
+
+    @torch.no_grad()
+    def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,
+                      temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
+                      unconditional_guidance_scale=1., unconditional_conditioning=None):
+        b, *_, device = *x.shape, x.device
+
+        if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
+            e_t = self.model.apply_model(x, t, c)
+        else:
+            x_in = torch.cat([x] * 2)
+            t_in = torch.cat([t] * 2)
+            c_in = torch.cat([unconditional_conditioning, c])
+            e_t_uncond, e_t = self.model.apply_model(x_in, t_in, c_in).chunk(2)
+            e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond)
+
+        if score_corrector is not None:
+            assert self.model.parameterization == "eps"
+            e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs)
+
+        alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
+        alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev
+        sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
+        sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
+        # select parameters corresponding to the currently considered timestep
+        a_t = torch.full((b, 1, 1, 1), alphas[index], device=device)
+        a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device)
+        sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device)
+        sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device)
+
+        # current prediction for x_0
+        pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
+        if quantize_denoised:
+            pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
+        # direction pointing to x_t
+        dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
+        noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
+        if noise_dropout > 0.:
+            noise = torch.nn.functional.dropout(noise, p=noise_dropout)
+        x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
+        return x_prev, pred_x0
+
+    @torch.no_grad()
+    def stochastic_encode(self, x0, t, use_original_steps=False, noise=None):
+        # fast, but does not allow for exact reconstruction
+        # t serves as an index to gather the correct alphas
+        if use_original_steps:
+            sqrt_alphas_cumprod = self.sqrt_alphas_cumprod
+            sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod
+        else:
+            sqrt_alphas_cumprod = torch.sqrt(self.ddim_alphas)
+            sqrt_one_minus_alphas_cumprod = self.ddim_sqrt_one_minus_alphas
+
+        if noise is None:
+            noise = torch.randn_like(x0)
+        return (extract_into_tensor(sqrt_alphas_cumprod, t, x0.shape) * x0 +
+                extract_into_tensor(sqrt_one_minus_alphas_cumprod, t, x0.shape) * noise)
+
+    @torch.no_grad()
+    def decode(self, x_latent, cond, t_start, unconditional_guidance_scale=1.0, unconditional_conditioning=None,
+               use_original_steps=False):
+
+        timesteps = np.arange(self.ddpm_num_timesteps) if use_original_steps else self.ddim_timesteps
+        timesteps = timesteps[:t_start]
+
+        time_range = np.flip(timesteps)
+        total_steps = timesteps.shape[0]
+        print(f"Running DDIM Sampling with {total_steps} timesteps")
+
+        iterator = tqdm(time_range, desc='Decoding image', total=total_steps)
+        x_dec = x_latent
+        for i, step in enumerate(iterator):
+            index = total_steps - i - 1
+            ts = torch.full((x_latent.shape[0],), step, device=x_latent.device, dtype=torch.long)
+            x_dec, _ = self.p_sample_ddim(x_dec, cond, ts, index=index, use_original_steps=use_original_steps,
+                                          unconditional_guidance_scale=unconditional_guidance_scale,
+                                          unconditional_conditioning=unconditional_conditioning)
+        return x_dec

examples/tutorial/stable_diffusion/ldm/models/diffusion/plms.py

@@ -0,0 +1,236 @@
+"""SAMPLING ONLY."""
+
+import torch
+import numpy as np
+from tqdm import tqdm
+from functools import partial
+
+from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like
+
+
+class PLMSSampler(object):
+    def __init__(self, model, schedule="linear", **kwargs):
+        super().__init__()
+        self.model = model
+        self.ddpm_num_timesteps = model.num_timesteps
+        self.schedule = schedule
+
+    def register_buffer(self, name, attr):
+        if type(attr) == torch.Tensor:
+            if attr.device != torch.device("cuda"):
+                attr = attr.to(torch.device("cuda"))
+        setattr(self, name, attr)
+
+    def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True):
+        if ddim_eta != 0:
+            raise ValueError('ddim_eta must be 0 for PLMS')
+        self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps,
+                                                  num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose)
+        alphas_cumprod = self.model.alphas_cumprod
+        assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep'
+        to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)
+
+        self.register_buffer('betas', to_torch(self.model.betas))
+        self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
+        self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev))
+
+        # calculations for diffusion q(x_t | x_{t-1}) and others
+        self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu())))
+        self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1)))
+
+        # ddim sampling parameters
+        ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(),
+                                                                                   ddim_timesteps=self.ddim_timesteps,
+                                                                                   eta=ddim_eta,verbose=verbose)
+        self.register_buffer('ddim_sigmas', ddim_sigmas)
+        self.register_buffer('ddim_alphas', ddim_alphas)
+        self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)
+        self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas))
+        sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
+            (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * (
+                        1 - self.alphas_cumprod / self.alphas_cumprod_prev))
+        self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps)
+
+    @torch.no_grad()
+    def sample(self,
+               S,
+               batch_size,
+               shape,
+               conditioning=None,
+               callback=None,
+               normals_sequence=None,
+               img_callback=None,
+               quantize_x0=False,
+               eta=0.,
+               mask=None,
+               x0=None,
+               temperature=1.,
+               noise_dropout=0.,
+               score_corrector=None,
+               corrector_kwargs=None,
+               verbose=True,
+               x_T=None,
+               log_every_t=100,
+               unconditional_guidance_scale=1.,
+               unconditional_conditioning=None,
+               # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
+               **kwargs
+               ):
+        if conditioning is not None:
+            if isinstance(conditioning, dict):
+                cbs = conditioning[list(conditioning.keys())[0]].shape[0]
+                if cbs != batch_size:
+                    print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
+            else:
+                if conditioning.shape[0] != batch_size:
+                    print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
+
+        self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose)
+        # sampling
+        C, H, W = shape
+        size = (batch_size, C, H, W)
+        print(f'Data shape for PLMS sampling is {size}')
+
+        samples, intermediates = self.plms_sampling(conditioning, size,
+                                                    callback=callback,
+                                                    img_callback=img_callback,
+                                                    quantize_denoised=quantize_x0,
+                                                    mask=mask, x0=x0,
+                                                    ddim_use_original_steps=False,
+                                                    noise_dropout=noise_dropout,
+                                                    temperature=temperature,
+                                                    score_corrector=score_corrector,
+                                                    corrector_kwargs=corrector_kwargs,
+                                                    x_T=x_T,
+                                                    log_every_t=log_every_t,
+                                                    unconditional_guidance_scale=unconditional_guidance_scale,
+                                                    unconditional_conditioning=unconditional_conditioning,
+                                                    )
+        return samples, intermediates
+
+    @torch.no_grad()
+    def plms_sampling(self, cond, shape,
+                      x_T=None, ddim_use_original_steps=False,
+                      callback=None, timesteps=None, quantize_denoised=False,
+                      mask=None, x0=None, img_callback=None, log_every_t=100,
+                      temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
+                      unconditional_guidance_scale=1., unconditional_conditioning=None,):
+        device = self.model.betas.device
+        b = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=device)
+        else:
+            img = x_T
+
+        if timesteps is None:
+            timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps
+        elif timesteps is not None and not ddim_use_original_steps:
+            subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1
+            timesteps = self.ddim_timesteps[:subset_end]
+
+        intermediates = {'x_inter': [img], 'pred_x0': [img]}
+        time_range = list(reversed(range(0,timesteps))) if ddim_use_original_steps else np.flip(timesteps)
+        total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
+        print(f"Running PLMS Sampling with {total_steps} timesteps")
+
+        iterator = tqdm(time_range, desc='PLMS Sampler', total=total_steps)
+        old_eps = []
+
+        for i, step in enumerate(iterator):
+            index = total_steps - i - 1
+            ts = torch.full((b,), step, device=device, dtype=torch.long)
+            ts_next = torch.full((b,), time_range[min(i + 1, len(time_range) - 1)], device=device, dtype=torch.long)
+
+            if mask is not None:
+                assert x0 is not None
+                img_orig = self.model.q_sample(x0, ts)  # TODO: deterministic forward pass?
+                img = img_orig * mask + (1. - mask) * img
+
+            outs = self.p_sample_plms(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps,
+                                      quantize_denoised=quantize_denoised, temperature=temperature,
+                                      noise_dropout=noise_dropout, score_corrector=score_corrector,
+                                      corrector_kwargs=corrector_kwargs,
+                                      unconditional_guidance_scale=unconditional_guidance_scale,
+                                      unconditional_conditioning=unconditional_conditioning,
+                                      old_eps=old_eps, t_next=ts_next)
+            img, pred_x0, e_t = outs
+            old_eps.append(e_t)
+            if len(old_eps) >= 4:
+                old_eps.pop(0)
+            if callback: callback(i)
+            if img_callback: img_callback(pred_x0, i)
+
+            if index % log_every_t == 0 or index == total_steps - 1:
+                intermediates['x_inter'].append(img)
+                intermediates['pred_x0'].append(pred_x0)
+
+        return img, intermediates
+
+    @torch.no_grad()
+    def p_sample_plms(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,
+                      temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
+                      unconditional_guidance_scale=1., unconditional_conditioning=None, old_eps=None, t_next=None):
+        b, *_, device = *x.shape, x.device
+
+        def get_model_output(x, t):
+            if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
+                e_t = self.model.apply_model(x, t, c)
+            else:
+                x_in = torch.cat([x] * 2)
+                t_in = torch.cat([t] * 2)
+                c_in = torch.cat([unconditional_conditioning, c])
+                e_t_uncond, e_t = self.model.apply_model(x_in, t_in, c_in).chunk(2)
+                e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond)
+
+            if score_corrector is not None:
+                assert self.model.parameterization == "eps"
+                e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs)
+
+            return e_t
+
+        alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
+        alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev
+        sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
+        sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
+
+        def get_x_prev_and_pred_x0(e_t, index):
+            # select parameters corresponding to the currently considered timestep
+            a_t = torch.full((b, 1, 1, 1), alphas[index], device=device)
+            a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device)
+            sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device)
+            sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device)
+
+            # current prediction for x_0
+            pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
+            if quantize_denoised:
+                pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
+            # direction pointing to x_t
+            dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
+            noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
+            if noise_dropout > 0.:
+                noise = torch.nn.functional.dropout(noise, p=noise_dropout)
+            x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
+            return x_prev, pred_x0
+
+        e_t = get_model_output(x, t)
+        if len(old_eps) == 0:
+            # Pseudo Improved Euler (2nd order)
+            x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t, index)
+            e_t_next = get_model_output(x_prev, t_next)
+            e_t_prime = (e_t + e_t_next) / 2
+        elif len(old_eps) == 1:
+            # 2nd order Pseudo Linear Multistep (Adams-Bashforth)
+            e_t_prime = (3 * e_t - old_eps[-1]) / 2
+        elif len(old_eps) == 2:
+            # 3nd order Pseudo Linear Multistep (Adams-Bashforth)
+            e_t_prime = (23 * e_t - 16 * old_eps[-1] + 5 * old_eps[-2]) / 12
+        elif len(old_eps) >= 3:
+            # 4nd order Pseudo Linear Multistep (Adams-Bashforth)
+            e_t_prime = (55 * e_t - 59 * old_eps[-1] + 37 * old_eps[-2] - 9 * old_eps[-3]) / 24
+
+        x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t_prime, index)
+
+        return x_prev, pred_x0, e_t

examples/tutorial/stable_diffusion/ldm/modules/attention.py

@@ -0,0 +1,314 @@
+from inspect import isfunction
+import math
+import torch
+import torch.nn.functional as F
+from torch import nn, einsum
+from einops import rearrange, repeat
+
+from torch.utils import checkpoint
+
+try:
+    from ldm.modules.flash_attention import flash_attention_qkv, flash_attention_q_kv
+    FlASH_AVAILABLE = True
+except:
+    FlASH_AVAILABLE = False
+
+USE_FLASH = False
+
+
+def enable_flash_attention():
+    global USE_FLASH
+    USE_FLASH = True
+    if FlASH_AVAILABLE is False:
+        print("Please install flash attention to activate new attention kernel.\n" + 
+              "Use \'pip install git+https://github.com/HazyResearch/flash-attention.git@c422fee3776eb3ea24e011ef641fd5fbeb212623#egg=flash_attn\'")
+
+
+def exists(val):
+    return val is not None
+
+
+def uniq(arr):
+    return{el: True for el in arr}.keys()
+
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+
+def max_neg_value(t):
+    return -torch.finfo(t.dtype).max
+
+
+def init_(tensor):
+    dim = tensor.shape[-1]
+    std = 1 / math.sqrt(dim)
+    tensor.uniform_(-std, std)
+    return tensor
+
+
+# feedforward
+class GEGLU(nn.Module):
+    def __init__(self, dim_in, dim_out):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out * 2)
+
+    def forward(self, x):
+        x, gate = self.proj(x).chunk(2, dim=-1)
+        return x * F.gelu(gate)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        dim_out = default(dim_out, dim)
+        project_in = nn.Sequential(
+            nn.Linear(dim, inner_dim),
+            nn.GELU()
+        ) if not glu else GEGLU(dim, inner_dim)
+
+        self.net = nn.Sequential(
+            project_in,
+            nn.Dropout(dropout),
+            nn.Linear(inner_dim, dim_out)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+def zero_module(module):
+    """
+    Zero out the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+
+def Normalize(in_channels):
+    return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
+
+
+class LinearAttention(nn.Module):
+    def __init__(self, dim, heads=4, dim_head=32):
+        super().__init__()
+        self.heads = heads
+        hidden_dim = dim_head * heads
+        self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias = False)
+        self.to_out = nn.Conv2d(hidden_dim, dim, 1)
+
+    def forward(self, x):
+        b, c, h, w = x.shape
+        qkv = self.to_qkv(x)
+        q, k, v = rearrange(qkv, 'b (qkv heads c) h w -> qkv b heads c (h w)', heads = self.heads, qkv=3)
+        k = k.softmax(dim=-1)  
+        context = torch.einsum('bhdn,bhen->bhde', k, v)
+        out = torch.einsum('bhde,bhdn->bhen', context, q)
+        out = rearrange(out, 'b heads c (h w) -> b (heads c) h w', heads=self.heads, h=h, w=w)
+        return self.to_out(out)
+
+
+class SpatialSelfAttention(nn.Module):
+    def __init__(self, in_channels):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.k = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.v = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.proj_out = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=1,
+                                        stride=1,
+                                        padding=0)
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b,c,h,w = q.shape
+        q = rearrange(q, 'b c h w -> b (h w) c')
+        k = rearrange(k, 'b c h w -> b c (h w)')
+        w_ = torch.einsum('bij,bjk->bik', q, k)
+
+        w_ = w_ * (int(c)**(-0.5))
+        w_ = torch.nn.functional.softmax(w_, dim=2)
+
+        # attend to values
+        v = rearrange(v, 'b c h w -> b c (h w)')
+        w_ = rearrange(w_, 'b i j -> b j i')
+        h_ = torch.einsum('bij,bjk->bik', v, w_)
+        h_ = rearrange(h_, 'b c (h w) -> b c h w', h=h)
+        h_ = self.proj_out(h_)
+
+        return x+h_
+
+
+class CrossAttention(nn.Module):
+    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.):
+        super().__init__()
+        inner_dim = dim_head * heads
+        context_dim = default(context_dim, query_dim)
+
+        self.scale = dim_head ** -0.5
+        self.heads = heads
+
+        self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
+        self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
+        self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
+
+        self.to_out = nn.Sequential(
+            nn.Linear(inner_dim, query_dim),
+            nn.Dropout(dropout)
+        )
+
+    def forward(self, x, context=None, mask=None):
+        q = self.to_q(x)
+        context = default(context, x)
+        k = self.to_k(context)
+        v = self.to_v(context)
+        dim_head = q.shape[-1] / self.heads
+
+        if USE_FLASH and FlASH_AVAILABLE and q.dtype in (torch.float16, torch.bfloat16) and \
+            dim_head <= 128 and (dim_head % 8) == 0:
+            # print("in flash")
+            if q.shape[1] == k.shape[1]:
+                out = self._flash_attention_qkv(q, k, v)
+            else:
+                out = self._flash_attention_q_kv(q, k, v)
+        else:
+            out = self._native_attention(q, k, v, self.heads, mask)
+
+        return self.to_out(out)
+
+    def _native_attention(self, q, k, v, h, mask):
+        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
+        sim = einsum('b i d, b j d -> b i j', q, k) * self.scale
+        if exists(mask):
+            mask = rearrange(mask, 'b ... -> b (...)')
+            max_neg_value = -torch.finfo(sim.dtype).max
+            mask = repeat(mask, 'b j -> (b h) () j', h=h)
+            sim.masked_fill_(~mask, max_neg_value)
+        # attention, what we cannot get enough of
+        out = sim.softmax(dim=-1)
+        out = einsum('b i j, b j d -> b i d', out, v)
+        out = rearrange(out, '(b h) n d -> b n (h d)', h=h)
+        return out
+
+    def _flash_attention_qkv(self, q, k, v):
+        qkv = torch.stack([q, k, v], dim=2)
+        b = qkv.shape[0]
+        n = qkv.shape[1]
+        qkv = rearrange(qkv, 'b n t (h d) -> (b n) t h d', h=self.heads)
+        out = flash_attention_qkv(qkv, self.scale, b, n)
+        out = rearrange(out, '(b n) h d -> b n (h d)', b=b, h=self.heads)
+        return out
+    
+    def _flash_attention_q_kv(self, q, k, v):
+        kv = torch.stack([k, v], dim=2)
+        b = q.shape[0]
+        q_seqlen = q.shape[1]
+        kv_seqlen = kv.shape[1]
+        q = rearrange(q, 'b n (h d) -> (b n) h d', h=self.heads)
+        kv = rearrange(kv, 'b n t (h d) -> (b n) t h d', h=self.heads)
+        out = flash_attention_q_kv(q, kv, self.scale, b, q_seqlen, kv_seqlen)
+        out = rearrange(out, '(b n) h d -> b n (h d)', b=b, h=self.heads)
+        return out
+
+
+class BasicTransformerBlock(nn.Module):
+    def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, use_checkpoint=False):
+        super().__init__()
+        self.attn1 = CrossAttention(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout)  # is a self-attention
+        self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
+        self.attn2 = CrossAttention(query_dim=dim, context_dim=context_dim,
+                                    heads=n_heads, dim_head=d_head, dropout=dropout)  # is self-attn if context is none
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+        self.norm3 = nn.LayerNorm(dim)
+        self.use_checkpoint = use_checkpoint
+
+    def forward(self, x, context=None):
+
+ 
+        if self.use_checkpoint:
+            return checkpoint(self._forward, x, context)
+        else:
+            return self._forward(x, context)
+
+    def _forward(self, x, context=None):
+        x = self.attn1(self.norm1(x)) + x
+        x = self.attn2(self.norm2(x), context=context) + x
+        x = self.ff(self.norm3(x)) + x
+        return x
+        
+
+
+class SpatialTransformer(nn.Module):
+    """
+    Transformer block for image-like data.
+    First, project the input (aka embedding)
+    and reshape to b, t, d.
+    Then apply standard transformer action.
+    Finally, reshape to image
+    """
+    def __init__(self, in_channels, n_heads, d_head,
+                 depth=1, dropout=0., context_dim=None, use_checkpoint=False):
+        super().__init__()
+        self.in_channels = in_channels
+        inner_dim = n_heads * d_head
+        self.norm = Normalize(in_channels)
+
+        self.proj_in = nn.Conv2d(in_channels,
+                                 inner_dim,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+
+        self.transformer_blocks = nn.ModuleList(
+            [BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim, use_checkpoint=use_checkpoint)
+                for d in range(depth)]
+        )
+
+        self.proj_out = zero_module(nn.Conv2d(inner_dim,
+                                              in_channels,
+                                              kernel_size=1,
+                                              stride=1,
+                                              padding=0))
+
+
+    def forward(self, x, context=None):
+        # note: if no context is given, cross-attention defaults to self-attention
+        b, c, h, w = x.shape
+        x_in = x
+        x = self.norm(x)
+        x = self.proj_in(x)
+        x = rearrange(x, 'b c h w -> b (h w) c')
+        x = x.contiguous()
+        for block in self.transformer_blocks:
+            x = block(x, context=context)
+        x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w)
+        x = x.contiguous()
+        x = self.proj_out(x)
+        return x + x_in

examples/tutorial/stable_diffusion/ldm/modules/diffusionmodules/model.py

@@ -0,0 +1,862 @@
+# pytorch_diffusion + derived encoder decoder
+import math
+import torch
+import torch.nn as nn
+import numpy as np
+from einops import rearrange
+
+from ldm.util import instantiate_from_config
+from ldm.modules.attention import LinearAttention
+
+
+def get_timestep_embedding(timesteps, embedding_dim):
+    """
+    This matches the implementation in Denoising Diffusion Probabilistic Models:
+    From Fairseq.
+    Build sinusoidal embeddings.
+    This matches the implementation in tensor2tensor, but differs slightly
+    from the description in Section 3.5 of "Attention Is All You Need".
+    """
+    assert len(timesteps.shape) == 1
+
+    half_dim = embedding_dim // 2
+    emb = math.log(10000) / (half_dim - 1)
+    emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb)
+    emb = emb.to(device=timesteps.device)
+    emb = timesteps.float()[:, None] * emb[None, :]
+    emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+    if embedding_dim % 2 == 1:  # zero pad
+        emb = torch.nn.functional.pad(emb, (0,1,0,0))
+    return emb
+
+
+def nonlinearity(x):
+    # swish
+    return x*torch.sigmoid(x)
+
+
+def Normalize(in_channels, num_groups=32):
+    return torch.nn.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+
+
+class Upsample(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            self.conv = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
+        if self.with_conv:
+            x = self.conv(x)
+        return x
+
+
+class Downsample(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=3,
+                                        stride=2,
+                                        padding=0)
+
+    def forward(self, x):
+        if self.with_conv:
+            pad = (0,1,0,1)
+            x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
+            x = self.conv(x)
+        else:
+            x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
+        return x
+
+
+class ResnetBlock(nn.Module):
+    def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False,
+                 dropout, temb_channels=512):
+        super().__init__()
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+
+        self.norm1 = Normalize(in_channels)
+        self.conv1 = torch.nn.Conv2d(in_channels,
+                                     out_channels,
+                                     kernel_size=3,
+                                     stride=1,
+                                     padding=1)
+        if temb_channels > 0:
+            self.temb_proj = torch.nn.Linear(temb_channels,
+                                             out_channels)
+        self.norm2 = Normalize(out_channels)
+        self.dropout = torch.nn.Dropout(dropout)
+        self.conv2 = torch.nn.Conv2d(out_channels,
+                                     out_channels,
+                                     kernel_size=3,
+                                     stride=1,
+                                     padding=1)
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                self.conv_shortcut = torch.nn.Conv2d(in_channels,
+                                                     out_channels,
+                                                     kernel_size=3,
+                                                     stride=1,
+                                                     padding=1)
+            else:
+                self.nin_shortcut = torch.nn.Conv2d(in_channels,
+                                                    out_channels,
+                                                    kernel_size=1,
+                                                    stride=1,
+                                                    padding=0)
+
+    def forward(self, x, temb):
+        h = x
+        h = self.norm1(h)
+        h = nonlinearity(h)
+        h = self.conv1(h)
+
+        if temb is not None:
+            h = h + self.temb_proj(nonlinearity(temb))[:,:,None,None]
+
+        h = self.norm2(h)
+        h = nonlinearity(h)
+        h = self.dropout(h)
+        h = self.conv2(h)
+
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                x = self.conv_shortcut(x)
+            else:
+                x = self.nin_shortcut(x)
+
+        return x+h
+
+
+class LinAttnBlock(LinearAttention):
+    """to match AttnBlock usage"""
+    def __init__(self, in_channels):
+        super().__init__(dim=in_channels, heads=1, dim_head=in_channels)
+
+
+class AttnBlock(nn.Module):
+    def __init__(self, in_channels):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.k = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.v = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.proj_out = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=1,
+                                        stride=1,
+                                        padding=0)
+
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b,c,h,w = q.shape
+        q = q.reshape(b,c,h*w)
+        q = q.permute(0,2,1)   # b,hw,c
+        k = k.reshape(b,c,h*w) # b,c,hw
+        w_ = torch.bmm(q,k)     # b,hw,hw    w[b,i,j]=sum_c q[b,i,c]k[b,c,j]
+        w_ = w_ * (int(c)**(-0.5))
+        w_ = torch.nn.functional.softmax(w_, dim=2)
+
+        # attend to values
+        v = v.reshape(b,c,h*w)
+        w_ = w_.permute(0,2,1)   # b,hw,hw (first hw of k, second of q)
+        h_ = torch.bmm(v,w_)     # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j]
+        h_ = h_.reshape(b,c,h,w)
+
+        h_ = self.proj_out(h_)
+
+        return x+h_
+
+
+def make_attn(in_channels, attn_type="vanilla"):
+    assert attn_type in ["vanilla", "linear", "none"], f'attn_type {attn_type} unknown'
+    print(f"making attention of type '{attn_type}' with {in_channels} in_channels")
+    if attn_type == "vanilla":
+        return AttnBlock(in_channels)
+    elif attn_type == "none":
+        return nn.Identity(in_channels)
+    else:
+        return LinAttnBlock(in_channels)
+
+class temb_module(nn.Module):
+    def __init__(self):
+        super().__init__()
+        pass
+
+class Model(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, use_timestep=True, use_linear_attn=False, attn_type="vanilla"):
+        super().__init__()
+        if use_linear_attn: attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = self.ch*4
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+
+        self.use_timestep = use_timestep
+        if self.use_timestep:
+            # timestep embedding
+            # self.temb = nn.Module()
+            self.temb = temb_module()
+            self.temb.dense = nn.ModuleList([
+                torch.nn.Linear(self.ch,
+                                self.temb_ch),
+                torch.nn.Linear(self.temb_ch,
+                                self.temb_ch),
+            ])
+
+        # downsampling
+        self.conv_in = torch.nn.Conv2d(in_channels,
+                                       self.ch,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        curr_res = resolution
+        in_ch_mult = (1,)+tuple(ch_mult)
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch*in_ch_mult[i_level]
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            # down = nn.Module()
+            down = Down_module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions-1:
+                down.downsample = Downsample(block_in, resamp_with_conv)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        # self.mid = nn.Module()
+        self.mid = Mid_module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch*ch_mult[i_level]
+            skip_in = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks+1):
+                if i_block == self.num_res_blocks:
+                    skip_in = ch*in_ch_mult[i_level]
+                block.append(ResnetBlock(in_channels=block_in+skip_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            # up = nn.Module()
+            up = Up_module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up) # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        out_ch,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x, t=None, context=None):
+        #assert x.shape[2] == x.shape[3] == self.resolution
+        if context is not None:
+            # assume aligned context, cat along channel axis
+            x = torch.cat((x, context), dim=1)
+        if self.use_timestep:
+            # timestep embedding
+            assert t is not None
+            temb = get_timestep_embedding(t, self.ch)
+            temb = self.temb.dense[0](temb)
+            temb = nonlinearity(temb)
+            temb = self.temb.dense[1](temb)
+        else:
+            temb = None
+
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1], temb)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if i_level != self.num_resolutions-1:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks+1):
+                h = self.up[i_level].block[i_block](
+                    torch.cat([h, hs.pop()], dim=1), temb)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+    def get_last_layer(self):
+        return self.conv_out.weight
+
+class Down_module(nn.Module):
+    def __init__(self):
+        super().__init__()
+        pass
+
+class Up_module(nn.Module):
+    def __init__(self):
+        super().__init__()
+        pass
+
+class Mid_module(nn.Module):
+    def __init__(self):
+        super().__init__()
+        pass
+
+
+class Encoder(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, z_channels, double_z=True, use_linear_attn=False, attn_type="vanilla",
+                 **ignore_kwargs):
+        super().__init__()
+        if use_linear_attn: attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+
+        # downsampling
+        self.conv_in = torch.nn.Conv2d(in_channels,
+                                       self.ch,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        curr_res = resolution
+        in_ch_mult = (1,)+tuple(ch_mult)
+        self.in_ch_mult = in_ch_mult
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch*in_ch_mult[i_level]
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            # down = nn.Module()
+            down = Down_module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions-1:
+                down.downsample = Downsample(block_in, resamp_with_conv)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        # self.mid = nn.Module()
+        self.mid = Mid_module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        2*z_channels if double_z else z_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        # timestep embedding
+        temb = None
+
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1], temb)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if i_level != self.num_resolutions-1:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class Decoder(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False,
+                 attn_type="vanilla", **ignorekwargs):
+        super().__init__()
+        if use_linear_attn: attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.give_pre_end = give_pre_end
+        self.tanh_out = tanh_out
+
+        # compute in_ch_mult, block_in and curr_res at lowest res
+        in_ch_mult = (1,)+tuple(ch_mult)
+        block_in = ch*ch_mult[self.num_resolutions-1]
+        curr_res = resolution // 2**(self.num_resolutions-1)
+        self.z_shape = (1,z_channels,curr_res,curr_res)
+        print("Working with z of shape {} = {} dimensions.".format(
+            self.z_shape, np.prod(self.z_shape)))
+
+        # z to block_in
+        self.conv_in = torch.nn.Conv2d(z_channels,
+                                       block_in,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        # middle
+        # self.mid = nn.Module()
+        self.mid = Mid_module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks+1):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            # up = nn.Module()
+            up = Up_module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up) # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        out_ch,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, z):
+        #assert z.shape[1:] == self.z_shape[1:]
+        self.last_z_shape = z.shape
+
+        # timestep embedding
+        temb = None
+
+        # z to block_in
+        h = self.conv_in(z)
+
+        # middle
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks+1):
+                h = self.up[i_level].block[i_block](h, temb)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        if self.give_pre_end:
+            return h
+
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        if self.tanh_out:
+            h = torch.tanh(h)
+        return h
+
+
+class SimpleDecoder(nn.Module):
+    def __init__(self, in_channels, out_channels, *args, **kwargs):
+        super().__init__()
+        self.model = nn.ModuleList([nn.Conv2d(in_channels, in_channels, 1),
+                                     ResnetBlock(in_channels=in_channels,
+                                                 out_channels=2 * in_channels,
+                                                 temb_channels=0, dropout=0.0),
+                                     ResnetBlock(in_channels=2 * in_channels,
+                                                out_channels=4 * in_channels,
+                                                temb_channels=0, dropout=0.0),
+                                     ResnetBlock(in_channels=4 * in_channels,
+                                                out_channels=2 * in_channels,
+                                                temb_channels=0, dropout=0.0),
+                                     nn.Conv2d(2*in_channels, in_channels, 1),
+                                     Upsample(in_channels, with_conv=True)])
+        # end
+        self.norm_out = Normalize(in_channels)
+        self.conv_out = torch.nn.Conv2d(in_channels,
+                                        out_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        for i, layer in enumerate(self.model):
+            if i in [1,2,3]:
+                x = layer(x, None)
+            else:
+                x = layer(x)
+
+        h = self.norm_out(x)
+        h = nonlinearity(h)
+        x = self.conv_out(h)
+        return x
+
+
+class UpsampleDecoder(nn.Module):
+    def __init__(self, in_channels, out_channels, ch, num_res_blocks, resolution,
+                 ch_mult=(2,2), dropout=0.0):
+        super().__init__()
+        # upsampling
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        block_in = in_channels
+        curr_res = resolution // 2 ** (self.num_resolutions - 1)
+        self.res_blocks = nn.ModuleList()
+        self.upsample_blocks = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            res_block = []
+            block_out = ch * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks + 1):
+                res_block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+            self.res_blocks.append(nn.ModuleList(res_block))
+            if i_level != self.num_resolutions - 1:
+                self.upsample_blocks.append(Upsample(block_in, True))
+                curr_res = curr_res * 2
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        out_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        # upsampling
+        h = x
+        for k, i_level in enumerate(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.res_blocks[i_level][i_block](h, None)
+            if i_level != self.num_resolutions - 1:
+                h = self.upsample_blocks[k](h)
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class LatentRescaler(nn.Module):
+    def __init__(self, factor, in_channels, mid_channels, out_channels, depth=2):
+        super().__init__()
+        # residual block, interpolate, residual block
+        self.factor = factor
+        self.conv_in = nn.Conv2d(in_channels,
+                                 mid_channels,
+                                 kernel_size=3,
+                                 stride=1,
+                                 padding=1)
+        self.res_block1 = nn.ModuleList([ResnetBlock(in_channels=mid_channels,
+                                                     out_channels=mid_channels,
+                                                     temb_channels=0,
+                                                     dropout=0.0) for _ in range(depth)])
+        self.attn = AttnBlock(mid_channels)
+        self.res_block2 = nn.ModuleList([ResnetBlock(in_channels=mid_channels,
+                                                     out_channels=mid_channels,
+                                                     temb_channels=0,
+                                                     dropout=0.0) for _ in range(depth)])
+
+        self.conv_out = nn.Conv2d(mid_channels,
+                                  out_channels,
+                                  kernel_size=1,
+                                  )
+
+    def forward(self, x):
+        x = self.conv_in(x)
+        for block in self.res_block1:
+            x = block(x, None)
+        x = torch.nn.functional.interpolate(x, size=(int(round(x.shape[2]*self.factor)), int(round(x.shape[3]*self.factor))))
+        x = self.attn(x)
+        for block in self.res_block2:
+            x = block(x, None)
+        x = self.conv_out(x)
+        return x
+
+
+class MergedRescaleEncoder(nn.Module):
+    def __init__(self, in_channels, ch, resolution, out_ch, num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True,
+                 ch_mult=(1,2,4,8), rescale_factor=1.0, rescale_module_depth=1):
+        super().__init__()
+        intermediate_chn = ch * ch_mult[-1]
+        self.encoder = Encoder(in_channels=in_channels, num_res_blocks=num_res_blocks, ch=ch, ch_mult=ch_mult,
+                               z_channels=intermediate_chn, double_z=False, resolution=resolution,
+                               attn_resolutions=attn_resolutions, dropout=dropout, resamp_with_conv=resamp_with_conv,
+                               out_ch=None)
+        self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=intermediate_chn,
+                                       mid_channels=intermediate_chn, out_channels=out_ch, depth=rescale_module_depth)
+
+    def forward(self, x):
+        x = self.encoder(x)
+        x = self.rescaler(x)
+        return x
+
+
+class MergedRescaleDecoder(nn.Module):
+    def __init__(self, z_channels, out_ch, resolution, num_res_blocks, attn_resolutions, ch, ch_mult=(1,2,4,8),
+                 dropout=0.0, resamp_with_conv=True, rescale_factor=1.0, rescale_module_depth=1):
+        super().__init__()
+        tmp_chn = z_channels*ch_mult[-1]
+        self.decoder = Decoder(out_ch=out_ch, z_channels=tmp_chn, attn_resolutions=attn_resolutions, dropout=dropout,
+                               resamp_with_conv=resamp_with_conv, in_channels=None, num_res_blocks=num_res_blocks,
+                               ch_mult=ch_mult, resolution=resolution, ch=ch)
+        self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=z_channels, mid_channels=tmp_chn,
+                                       out_channels=tmp_chn, depth=rescale_module_depth)
+
+    def forward(self, x):
+        x = self.rescaler(x)
+        x = self.decoder(x)
+        return x
+
+
+class Upsampler(nn.Module):
+    def __init__(self, in_size, out_size, in_channels, out_channels, ch_mult=2):
+        super().__init__()
+        assert out_size >= in_size
+        num_blocks = int(np.log2(out_size//in_size))+1
+        factor_up = 1.+ (out_size % in_size)
+        print(f"Building {self.__class__.__name__} with in_size: {in_size} --> out_size {out_size} and factor {factor_up}")
+        self.rescaler = LatentRescaler(factor=factor_up, in_channels=in_channels, mid_channels=2*in_channels,
+                                       out_channels=in_channels)
+        self.decoder = Decoder(out_ch=out_channels, resolution=out_size, z_channels=in_channels, num_res_blocks=2,
+                               attn_resolutions=[], in_channels=None, ch=in_channels,
+                               ch_mult=[ch_mult for _ in range(num_blocks)])
+
+    def forward(self, x):
+        x = self.rescaler(x)
+        x = self.decoder(x)
+        return x
+
+
+class Resize(nn.Module):
+    def __init__(self, in_channels=None, learned=False, mode="bilinear"):
+        super().__init__()
+        self.with_conv = learned
+        self.mode = mode
+        if self.with_conv:
+            print(f"Note: {self.__class__.__name} uses learned downsampling and will ignore the fixed {mode} mode")
+            raise NotImplementedError()
+            assert in_channels is not None
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=4,
+                                        stride=2,
+                                        padding=1)
+
+    def forward(self, x, scale_factor=1.0):
+        if scale_factor==1.0:
+            return x
+        else:
+            x = torch.nn.functional.interpolate(x, mode=self.mode, align_corners=False, scale_factor=scale_factor)
+        return x
+
+class FirstStagePostProcessor(nn.Module):
+
+    def __init__(self, ch_mult:list, in_channels,
+                 pretrained_model:nn.Module=None,
+                 reshape=False,
+                 n_channels=None,
+                 dropout=0.,
+                 pretrained_config=None):
+        super().__init__()
+        if pretrained_config is None:
+            assert pretrained_model is not None, 'Either "pretrained_model" or "pretrained_config" must not be None'
+            self.pretrained_model = pretrained_model
+        else:
+            assert pretrained_config is not None, 'Either "pretrained_model" or "pretrained_config" must not be None'
+            self.instantiate_pretrained(pretrained_config)
+
+        self.do_reshape = reshape
+
+        if n_channels is None:
+            n_channels = self.pretrained_model.encoder.ch
+
+        self.proj_norm = Normalize(in_channels,num_groups=in_channels//2)
+        self.proj = nn.Conv2d(in_channels,n_channels,kernel_size=3,
+                            stride=1,padding=1)
+
+        blocks = []
+        downs = []
+        ch_in = n_channels
+        for m in ch_mult:
+            blocks.append(ResnetBlock(in_channels=ch_in,out_channels=m*n_channels,dropout=dropout))
+            ch_in = m * n_channels
+            downs.append(Downsample(ch_in, with_conv=False))
+
+        self.model = nn.ModuleList(blocks)
+        self.downsampler = nn.ModuleList(downs)
+
+
+    def instantiate_pretrained(self, config):
+        model = instantiate_from_config(config)
+        self.pretrained_model = model.eval()
+        # self.pretrained_model.train = False
+        for param in self.pretrained_model.parameters():
+            param.requires_grad = False
+
+
+    @torch.no_grad()
+    def encode_with_pretrained(self,x):
+        c = self.pretrained_model.encode(x)
+        if isinstance(c, DiagonalGaussianDistribution):
+            c = c.mode()
+        return  c
+
+    def forward(self,x):
+        z_fs = self.encode_with_pretrained(x)
+        z = self.proj_norm(z_fs)
+        z = self.proj(z)
+        z = nonlinearity(z)
+
+        for submodel, downmodel in zip(self.model,self.downsampler):
+            z = submodel(z,temb=None)
+            z = downmodel(z)
+
+        if self.do_reshape:
+            z = rearrange(z,'b c h w -> b (h w) c')
+        return z
+

examples/tutorial/stable_diffusion/ldm/modules/diffusionmodules/util.py

@@ -0,0 +1,276 @@
+# adopted from
+# https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py
+# and
+# https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
+# and
+# https://github.com/openai/guided-diffusion/blob/0ba878e517b276c45d1195eb29f6f5f72659a05b/guided_diffusion/nn.py
+#
+# thanks!
+
+
+import os
+import math
+import torch
+import torch.nn as nn
+import numpy as np
+from einops import repeat
+
+from ldm.util import instantiate_from_config
+
+
+def make_beta_schedule(schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+    if schedule == "linear":
+        betas = (
+                torch.linspace(linear_start ** 0.5, linear_end ** 0.5, n_timestep, dtype=torch.float64) ** 2
+        )
+
+    elif schedule == "cosine":
+        timesteps = (
+                torch.arange(n_timestep + 1, dtype=torch.float64) / n_timestep + cosine_s
+        )
+        alphas = timesteps / (1 + cosine_s) * np.pi / 2
+        alphas = torch.cos(alphas).pow(2)
+        alphas = alphas / alphas[0]
+        betas = 1 - alphas[1:] / alphas[:-1]
+        betas = np.clip(betas, a_min=0, a_max=0.999)
+
+    elif schedule == "sqrt_linear":
+        betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64)
+    elif schedule == "sqrt":
+        betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) ** 0.5
+    else:
+        raise ValueError(f"schedule '{schedule}' unknown.")
+    return betas.numpy()
+
+
+def make_ddim_timesteps(ddim_discr_method, num_ddim_timesteps, num_ddpm_timesteps, verbose=True):
+    if ddim_discr_method == 'uniform':
+        c = num_ddpm_timesteps // num_ddim_timesteps
+        ddim_timesteps = np.asarray(list(range(0, num_ddpm_timesteps, c)))
+    elif ddim_discr_method == 'quad':
+        ddim_timesteps = ((np.linspace(0, np.sqrt(num_ddpm_timesteps * .8), num_ddim_timesteps)) ** 2).astype(int)
+    else:
+        raise NotImplementedError(f'There is no ddim discretization method called "{ddim_discr_method}"')
+
+    # assert ddim_timesteps.shape[0] == num_ddim_timesteps
+    # add one to get the final alpha values right (the ones from first scale to data during sampling)
+    steps_out = ddim_timesteps + 1
+    if verbose:
+        print(f'Selected timesteps for ddim sampler: {steps_out}')
+    return steps_out
+
+
+def make_ddim_sampling_parameters(alphacums, ddim_timesteps, eta, verbose=True):
+    # select alphas for computing the variance schedule
+    alphas = alphacums[ddim_timesteps]
+    alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist())
+
+    # according the the formula provided in https://arxiv.org/abs/2010.02502
+    sigmas = eta * np.sqrt((1 - alphas_prev) / (1 - alphas) * (1 - alphas / alphas_prev))
+    if verbose:
+        print(f'Selected alphas for ddim sampler: a_t: {alphas}; a_(t-1): {alphas_prev}')
+        print(f'For the chosen value of eta, which is {eta}, '
+              f'this results in the following sigma_t schedule for ddim sampler {sigmas}')
+    return sigmas, alphas, alphas_prev
+
+
+def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999):
+    """
+    Create a beta schedule that discretizes the given alpha_t_bar function,
+    which defines the cumulative product of (1-beta) over time from t = [0,1].
+    :param num_diffusion_timesteps: the number of betas to produce.
+    :param alpha_bar: a lambda that takes an argument t from 0 to 1 and
+                      produces the cumulative product of (1-beta) up to that
+                      part of the diffusion process.
+    :param max_beta: the maximum beta to use; use values lower than 1 to
+                     prevent singularities.
+    """
+    betas = []
+    for i in range(num_diffusion_timesteps):
+        t1 = i / num_diffusion_timesteps
+        t2 = (i + 1) / num_diffusion_timesteps
+        betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta))
+    return np.array(betas)
+
+
+def extract_into_tensor(a, t, x_shape):
+    b, *_ = t.shape
+    out = a.gather(-1, t)
+    return out.reshape(b, *((1,) * (len(x_shape) - 1)))
+
+
+def checkpoint(func, inputs, params, flag):
+    """
+    Evaluate a function without caching intermediate activations, allowing for
+    reduced memory at the expense of extra compute in the backward pass.
+    :param func: the function to evaluate.
+    :param inputs: the argument sequence to pass to `func`.
+    :param params: a sequence of parameters `func` depends on but does not
+                   explicitly take as arguments.
+    :param flag: if False, disable gradient checkpointing.
+    """
+    if flag:
+        args = tuple(inputs) + tuple(params)
+        return CheckpointFunction.apply(func, len(inputs), *args)
+    else:
+        return func(*inputs)
+
+
+class CheckpointFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, run_function, length, *args):
+        ctx.run_function = run_function
+        ctx.input_tensors = list(args[:length])
+        ctx.input_params = list(args[length:])
+
+        with torch.no_grad():
+            output_tensors = ctx.run_function(*ctx.input_tensors)
+        return output_tensors
+
+    @staticmethod
+    def backward(ctx, *output_grads):
+        ctx.input_tensors = [x.detach().requires_grad_(True) for x in ctx.input_tensors]
+        with torch.enable_grad():
+            # Fixes a bug where the first op in run_function modifies the
+            # Tensor storage in place, which is not allowed for detach()'d
+            # Tensors.
+            shallow_copies = [x.view_as(x) for x in ctx.input_tensors]
+            output_tensors = ctx.run_function(*shallow_copies)
+        input_grads = torch.autograd.grad(
+            output_tensors,
+            ctx.input_tensors + ctx.input_params,
+            output_grads,
+            allow_unused=True,
+        )
+        del ctx.input_tensors
+        del ctx.input_params
+        del output_tensors
+        return (None, None) + input_grads
+
+
+def timestep_embedding(timesteps, dim, max_period=10000, repeat_only=False, use_fp16=True):
+    """
+    Create sinusoidal timestep embeddings.
+    :param timesteps: a 1-D Tensor of N indices, one per batch element.
+                      These may be fractional.
+    :param dim: the dimension of the output.
+    :param max_period: controls the minimum frequency of the embeddings.
+    :return: an [N x dim] Tensor of positional embeddings.
+    """
+    if not repeat_only:
+        half = dim // 2
+        freqs = torch.exp(
+            -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
+        ).to(device=timesteps.device)
+        args = timesteps[:, None].float() * freqs[None]
+        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+        if dim % 2:
+            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    else:
+        embedding = repeat(timesteps, 'b -> b d', d=dim)
+    if use_fp16:
+        return embedding.half()
+    else:
+        return embedding
+
+
+def zero_module(module):
+    """
+    Zero out the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+
+def scale_module(module, scale):
+    """
+    Scale the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().mul_(scale)
+    return module
+
+
+def mean_flat(tensor):
+    """
+    Take the mean over all non-batch dimensions.
+    """
+    return tensor.mean(dim=list(range(1, len(tensor.shape))))
+
+
+def normalization(channels, precision=16):
+    """
+    Make a standard normalization layer.
+    :param channels: number of input channels.
+    :return: an nn.Module for normalization.
+    """
+    if precision == 16:
+        return GroupNorm16(16, channels)
+    else:
+        return GroupNorm32(32, channels)
+
+
+# PyTorch 1.7 has SiLU, but we support PyTorch 1.5.
+class SiLU(nn.Module):
+    def forward(self, x):
+        return x * torch.sigmoid(x)
+
+class GroupNorm16(nn.GroupNorm):
+    def forward(self, x):
+        return super().forward(x.half()).type(x.dtype)
+
+class GroupNorm32(nn.GroupNorm):
+    def forward(self, x):
+        return super().forward(x.float()).type(x.dtype)
+
+def conv_nd(dims, *args, **kwargs):
+    """
+    Create a 1D, 2D, or 3D convolution module.
+    """
+    if dims == 1:
+        return nn.Conv1d(*args, **kwargs)
+    elif dims == 2:
+        return nn.Conv2d(*args, **kwargs)
+    elif dims == 3:
+        return nn.Conv3d(*args, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")
+
+
+def linear(*args, **kwargs):
+    """
+    Create a linear module.
+    """
+    return nn.Linear(*args, **kwargs)
+
+
+def avg_pool_nd(dims, *args, **kwargs):
+    """
+    Create a 1D, 2D, or 3D average pooling module.
+    """
+    if dims == 1:
+        return nn.AvgPool1d(*args, **kwargs)
+    elif dims == 2:
+        return nn.AvgPool2d(*args, **kwargs)
+    elif dims == 3:
+        return nn.AvgPool3d(*args, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")
+
+
+class HybridConditioner(nn.Module):
+
+    def __init__(self, c_concat_config, c_crossattn_config):
+        super().__init__()
+        self.concat_conditioner = instantiate_from_config(c_concat_config)
+        self.crossattn_conditioner = instantiate_from_config(c_crossattn_config)
+
+    def forward(self, c_concat, c_crossattn):
+        c_concat = self.concat_conditioner(c_concat)
+        c_crossattn = self.crossattn_conditioner(c_crossattn)
+        return {'c_concat': [c_concat], 'c_crossattn': [c_crossattn]}
+
+
+def noise_like(shape, device, repeat=False):
+    repeat_noise = lambda: torch.randn((1, *shape[1:]), device=device).repeat(shape[0], *((1,) * (len(shape) - 1)))
+    noise = lambda: torch.randn(shape, device=device)
+    return repeat_noise() if repeat else noise()

examples/tutorial/stable_diffusion/ldm/modules/distributions/distributions.py

@@ -0,0 +1,92 @@
+import torch
+import numpy as np
+
+
+class AbstractDistribution:
+    def sample(self):
+        raise NotImplementedError()
+
+    def mode(self):
+        raise NotImplementedError()
+
+
+class DiracDistribution(AbstractDistribution):
+    def __init__(self, value):
+        self.value = value
+
+    def sample(self):
+        return self.value
+
+    def mode(self):
+        return self.value
+
+
+class DiagonalGaussianDistribution(object):
+    def __init__(self, parameters, deterministic=False):
+        self.parameters = parameters
+        self.mean, self.logvar = torch.chunk(parameters, 2, dim=1)
+        self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
+        self.deterministic = deterministic
+        self.std = torch.exp(0.5 * self.logvar)
+        self.var = torch.exp(self.logvar)
+        if self.deterministic:
+            self.var = self.std = torch.zeros_like(self.mean).to(device=self.parameters.device)
+
+    def sample(self):
+        x = self.mean + self.std * torch.randn(self.mean.shape).to(device=self.parameters.device)
+        return x
+
+    def kl(self, other=None):
+        if self.deterministic:
+            return torch.Tensor([0.])
+        else:
+            if other is None:
+                return 0.5 * torch.sum(torch.pow(self.mean, 2)
+                                       + self.var - 1.0 - self.logvar,
+                                       dim=[1, 2, 3])
+            else:
+                return 0.5 * torch.sum(
+                    torch.pow(self.mean - other.mean, 2) / other.var
+                    + self.var / other.var - 1.0 - self.logvar + other.logvar,
+                    dim=[1, 2, 3])
+
+    def nll(self, sample, dims=[1,2,3]):
+        if self.deterministic:
+            return torch.Tensor([0.])
+        logtwopi = np.log(2.0 * np.pi)
+        return 0.5 * torch.sum(
+            logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var,
+            dim=dims)
+
+    def mode(self):
+        return self.mean
+
+
+def normal_kl(mean1, logvar1, mean2, logvar2):
+    """
+    source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12
+    Compute the KL divergence between two gaussians.
+    Shapes are automatically broadcasted, so batches can be compared to
+    scalars, among other use cases.
+    """
+    tensor = None
+    for obj in (mean1, logvar1, mean2, logvar2):
+        if isinstance(obj, torch.Tensor):
+            tensor = obj
+            break
+    assert tensor is not None, "at least one argument must be a Tensor"
+
+    # Force variances to be Tensors. Broadcasting helps convert scalars to
+    # Tensors, but it does not work for torch.exp().
+    logvar1, logvar2 = [
+        x if isinstance(x, torch.Tensor) else torch.tensor(x).to(tensor)
+        for x in (logvar1, logvar2)
+    ]
+
+    return 0.5 * (
+        -1.0
+        + logvar2
+        - logvar1
+        + torch.exp(logvar1 - logvar2)
+        + ((mean1 - mean2) ** 2) * torch.exp(-logvar2)
+    )

examples/tutorial/stable_diffusion/ldm/modules/ema.py

@@ -0,0 +1,76 @@
+import torch
+from torch import nn
+
+
+class LitEma(nn.Module):
+    def __init__(self, model, decay=0.9999, use_num_upates=True):
+        super().__init__()
+        if decay < 0.0 or decay > 1.0:
+            raise ValueError('Decay must be between 0 and 1')
+
+        self.m_name2s_name = {}
+        self.register_buffer('decay', torch.tensor(decay, dtype=torch.float32))
+        self.register_buffer('num_updates', torch.tensor(0,dtype=torch.int) if use_num_upates
+                             else torch.tensor(-1,dtype=torch.int))
+
+        for name, p in model.named_parameters():
+            if p.requires_grad:
+                #remove as '.'-character is not allowed in buffers
+                s_name = name.replace('.','')
+                self.m_name2s_name.update({name:s_name})
+                self.register_buffer(s_name,p.clone().detach().data)
+
+        self.collected_params = []
+
+    def forward(self,model):
+        decay = self.decay
+
+        if self.num_updates >= 0:
+            self.num_updates += 1
+            decay = min(self.decay,(1 + self.num_updates) / (10 + self.num_updates))
+
+        one_minus_decay = 1.0 - decay
+
+        with torch.no_grad():
+            m_param = dict(model.named_parameters())
+            shadow_params = dict(self.named_buffers())
+
+            for key in m_param:
+                if m_param[key].requires_grad:
+                    sname = self.m_name2s_name[key]
+                    shadow_params[sname] = shadow_params[sname].type_as(m_param[key])
+                    shadow_params[sname].sub_(one_minus_decay * (shadow_params[sname] - m_param[key]))
+                else:
+                    assert not key in self.m_name2s_name
+
+    def copy_to(self, model):
+        m_param = dict(model.named_parameters())
+        shadow_params = dict(self.named_buffers())
+        for key in m_param:
+            if m_param[key].requires_grad:
+                m_param[key].data.copy_(shadow_params[self.m_name2s_name[key]].data)
+            else:
+                assert not key in self.m_name2s_name
+
+    def store(self, parameters):
+        """
+        Save the current parameters for restoring later.
+        Args:
+          parameters: Iterable of `torch.nn.Parameter`; the parameters to be
+            temporarily stored.
+        """
+        self.collected_params = [param.clone() for param in parameters]
+
+    def restore(self, parameters):
+        """
+        Restore the parameters stored with the `store` method.
+        Useful to validate the model with EMA parameters without affecting the
+        original optimization process. Store the parameters before the
+        `copy_to` method. After validation (or model saving), use this to
+        restore the former parameters.
+        Args:
+          parameters: Iterable of `torch.nn.Parameter`; the parameters to be
+            updated with the stored parameters.
+        """
+        for c_param, param in zip(self.collected_params, parameters):
+            param.data.copy_(c_param.data)

examples/tutorial/stable_diffusion/ldm/modules/encoders/modules.py

@@ -0,0 +1,264 @@
+import types 
+
+import torch
+import torch.nn as nn
+from functools import partial
+import clip
+from einops import rearrange, repeat
+from transformers import CLIPTokenizer, CLIPTextModel, CLIPTextConfig
+import kornia
+from transformers.models.clip.modeling_clip import CLIPTextTransformer
+
+from ldm.modules.x_transformer import Encoder, TransformerWrapper  # TODO: can we directly rely on lucidrains code and simply add this as a reuirement? --> test
+
+
+class AbstractEncoder(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def encode(self, *args, **kwargs):
+        raise NotImplementedError
+
+
+
+class ClassEmbedder(nn.Module):
+    def __init__(self, embed_dim, n_classes=1000, key='class'):
+        super().__init__()
+        self.key = key
+        self.embedding = nn.Embedding(n_classes, embed_dim)
+
+    def forward(self, batch, key=None):
+        if key is None:
+            key = self.key
+        # this is for use in crossattn
+        c = batch[key][:, None]
+        c = self.embedding(c)
+        return c
+
+
+class TransformerEmbedder(AbstractEncoder):
+    """Some transformer encoder layers"""
+    def __init__(self, n_embed, n_layer, vocab_size, max_seq_len=77, device="cuda"):
+        super().__init__()
+        self.device = device
+        self.transformer = TransformerWrapper(num_tokens=vocab_size, max_seq_len=max_seq_len,
+                                              attn_layers=Encoder(dim=n_embed, depth=n_layer))
+
+    def forward(self, tokens):
+        tokens = tokens.to(self.device)  # meh
+        z = self.transformer(tokens, return_embeddings=True)
+        return z
+
+    def encode(self, x):
+        return self(x)
+
+
+class BERTTokenizer(AbstractEncoder):
+    """ Uses a pretrained BERT tokenizer by huggingface. Vocab size: 30522 (?)"""
+    def __init__(self, device="cuda", vq_interface=True, max_length=77):
+        super().__init__()
+        from transformers import BertTokenizerFast  # TODO: add to reuquirements
+        self.tokenizer = BertTokenizerFast.from_pretrained("bert-base-uncased")
+        self.device = device
+        self.vq_interface = vq_interface
+        self.max_length = max_length
+
+    def forward(self, text):
+        batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
+                                        return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
+        tokens = batch_encoding["input_ids"].to(self.device)
+        return tokens
+
+    @torch.no_grad()
+    def encode(self, text):
+        tokens = self(text)
+        if not self.vq_interface:
+            return tokens
+        return None, None, [None, None, tokens]
+
+    def decode(self, text):
+        return text
+
+
+class BERTEmbedder(AbstractEncoder):
+    """Uses the BERT tokenizr model and add some transformer encoder layers"""
+    def __init__(self, n_embed, n_layer, vocab_size=30522, max_seq_len=77,
+                 device="cuda",use_tokenizer=True, embedding_dropout=0.0):
+        super().__init__()
+        self.use_tknz_fn = use_tokenizer
+        if self.use_tknz_fn:
+            self.tknz_fn = BERTTokenizer(vq_interface=False, max_length=max_seq_len)
+        self.device = device
+        self.transformer = TransformerWrapper(num_tokens=vocab_size, max_seq_len=max_seq_len,
+                                              attn_layers=Encoder(dim=n_embed, depth=n_layer),
+                                              emb_dropout=embedding_dropout)
+
+    def forward(self, text):
+        if self.use_tknz_fn:
+            tokens = self.tknz_fn(text)#.to(self.device)
+        else:
+            tokens = text
+        z = self.transformer(tokens, return_embeddings=True)
+        return z
+
+    def encode(self, text):
+        # output of length 77
+        return self(text)
+
+
+class SpatialRescaler(nn.Module):
+    def __init__(self,
+                 n_stages=1,
+                 method='bilinear',
+                 multiplier=0.5,
+                 in_channels=3,
+                 out_channels=None,
+                 bias=False):
+        super().__init__()
+        self.n_stages = n_stages
+        assert self.n_stages >= 0
+        assert method in ['nearest','linear','bilinear','trilinear','bicubic','area']
+        self.multiplier = multiplier
+        self.interpolator = partial(torch.nn.functional.interpolate, mode=method)
+        self.remap_output = out_channels is not None
+        if self.remap_output:
+            print(f'Spatial Rescaler mapping from {in_channels} to {out_channels} channels after resizing.')
+            self.channel_mapper = nn.Conv2d(in_channels,out_channels,1,bias=bias)
+
+    def forward(self,x):
+        for stage in range(self.n_stages):
+            x = self.interpolator(x, scale_factor=self.multiplier)
+
+
+        if self.remap_output:
+            x = self.channel_mapper(x)
+        return x
+
+    def encode(self, x):
+        return self(x)
+
+
+class CLIPTextModelZero(CLIPTextModel):
+    config_class = CLIPTextConfig
+
+    def __init__(self, config: CLIPTextConfig):
+        super().__init__(config)
+        self.text_model = CLIPTextTransformerZero(config)
+
+class CLIPTextTransformerZero(CLIPTextTransformer):
+    def _build_causal_attention_mask(self, bsz, seq_len):
+        # lazily create causal attention mask, with full attention between the vision tokens
+        # pytorch uses additive attention mask; fill with -inf
+        mask = torch.empty(bsz, seq_len, seq_len)
+        mask.fill_(float("-inf"))
+        mask.triu_(1)  # zero out the lower diagonal
+        mask = mask.unsqueeze(1)  # expand mask
+        return mask.half()
+
+class FrozenCLIPEmbedder(AbstractEncoder):
+    """Uses the CLIP transformer encoder for text (from Hugging Face)"""
+    def __init__(self, version="openai/clip-vit-large-patch14", device="cuda", max_length=77, use_fp16=True):
+        super().__init__()
+        self.tokenizer = CLIPTokenizer.from_pretrained(version)
+        
+        if use_fp16:
+            self.transformer = CLIPTextModelZero.from_pretrained(version)
+        else:
+            self.transformer = CLIPTextModel.from_pretrained(version)
+
+        # print(self.transformer.modules())
+        # print("check model dtyoe: {}, {}".format(self.tokenizer.dtype, self.transformer.dtype))
+        self.device = device
+        self.max_length = max_length
+        self.freeze() 
+
+    def freeze(self):
+        self.transformer = self.transformer.eval()
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def forward(self, text):
+        batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
+                                        return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
+        # tokens = batch_encoding["input_ids"].to(self.device)
+        tokens = batch_encoding["input_ids"].to(self.device)
+        # print("token type: {}".format(tokens.dtype))
+        outputs = self.transformer(input_ids=tokens)
+
+        z = outputs.last_hidden_state
+        return z
+
+    def encode(self, text):
+        return self(text)
+
+
+class FrozenCLIPTextEmbedder(nn.Module):
+    """
+    Uses the CLIP transformer encoder for text.
+    """
+    def __init__(self, version='ViT-L/14', device="cuda", max_length=77, n_repeat=1, normalize=True):
+        super().__init__()
+        self.model, _ = clip.load(version, jit=False, device="cpu")
+        self.device = device
+        self.max_length = max_length
+        self.n_repeat = n_repeat
+        self.normalize = normalize
+
+    def freeze(self):
+        self.model = self.model.eval()
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def forward(self, text):
+        tokens = clip.tokenize(text).to(self.device)
+        z = self.model.encode_text(tokens)
+        if self.normalize:
+            z = z / torch.linalg.norm(z, dim=1, keepdim=True)
+        return z
+
+    def encode(self, text):
+        z = self(text)
+        if z.ndim==2:
+            z = z[:, None, :]
+        z = repeat(z, 'b 1 d -> b k d', k=self.n_repeat)
+        return z
+
+
+class FrozenClipImageEmbedder(nn.Module):
+    """
+        Uses the CLIP image encoder.
+        """
+    def __init__(
+            self,
+            model,
+            jit=False,
+            device='cuda' if torch.cuda.is_available() else 'cpu',
+            antialias=False,
+        ):
+        super().__init__()
+        self.model, _ = clip.load(name=model, device=device, jit=jit)
+
+        self.antialias = antialias
+
+        self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False)
+        self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False)
+
+    def preprocess(self, x):
+        # normalize to [0,1]
+        x = kornia.geometry.resize(x, (224, 224),
+                                   interpolation='bicubic',align_corners=True,
+                                   antialias=self.antialias)
+        x = (x + 1.) / 2.
+        # renormalize according to clip
+        x = kornia.enhance.normalize(x, self.mean, self.std)
+        return x
+
+    def forward(self, x):
+        # x is assumed to be in range [-1,1]
+        return self.model.encode_image(self.preprocess(x))
+
+
+if __name__ == "__main__":
+    from ldm.util import count_params
+    model = FrozenCLIPEmbedder()
+    count_params(model, verbose=True)

examples/tutorial/stable_diffusion/ldm/modules/flash_attention.py

@@ -0,0 +1,50 @@
+"""
+Fused Attention
+===============
+This is a Triton implementation of the Flash Attention algorithm
+(see: Dao et al., https://arxiv.org/pdf/2205.14135v2.pdf; Rabe and Staats https://arxiv.org/pdf/2112.05682v2.pdf; Triton https://github.com/openai/triton)
+"""
+
+import torch
+try:
+    from flash_attn.flash_attn_interface import flash_attn_unpadded_qkvpacked_func, flash_attn_unpadded_kvpacked_func
+except ImportError:
+    raise ImportError('please install flash_attn from https://github.com/HazyResearch/flash-attention')
+
+
+
+def flash_attention_qkv(qkv, sm_scale, batch_size, seq_len):
+    """
+    Arguments:
+        qkv: (batch*seq, 3, nheads, headdim)
+        batch_size: int.
+        seq_len: int.
+        sm_scale: float. The scaling of QK^T before applying softmax.
+    Return:
+        out: (total, nheads, headdim).
+    """
+    max_s = seq_len
+    cu_seqlens = torch.arange(0, (batch_size + 1) * seq_len, step=seq_len, dtype=torch.int32,
+        device=qkv.device)
+    out = flash_attn_unpadded_qkvpacked_func(
+        qkv, cu_seqlens, max_s, 0.0,
+        softmax_scale=sm_scale, causal=False
+    )
+    return out
+
+
+def flash_attention_q_kv(q, kv, sm_scale, batch_size, q_seqlen, kv_seqlen):
+    """
+    Arguments:
+        q: (batch*seq, nheads, headdim)
+        kv: (batch*seq, 2, nheads, headdim)
+        batch_size: int.
+        seq_len: int.
+        sm_scale: float. The scaling of QK^T before applying softmax.
+    Return:
+        out: (total, nheads, headdim).
+    """
+    cu_seqlens_q = torch.arange(0, (batch_size + 1) * q_seqlen, step=q_seqlen, dtype=torch.int32, device=q.device)
+    cu_seqlens_k = torch.arange(0, (batch_size + 1) * kv_seqlen, step=kv_seqlen, dtype=torch.int32, device=kv.device)
+    out = flash_attn_unpadded_kvpacked_func(q, kv, cu_seqlens_q, cu_seqlens_k, q_seqlen, kv_seqlen, 0.0, sm_scale)
+    return out

examples/tutorial/stable_diffusion/ldm/modules/image_degradation/__init__.py

@@ -0,0 +1,2 @@
+from ldm.modules.image_degradation.bsrgan import degradation_bsrgan_variant as degradation_fn_bsr
+from ldm.modules.image_degradation.bsrgan_light import degradation_bsrgan_variant as degradation_fn_bsr_light

examples/tutorial/stable_diffusion/ldm/modules/image_degradation/bsrgan.py

@@ -0,0 +1,730 @@
+# -*- coding: utf-8 -*-
+"""
+# --------------------------------------------
+# Super-Resolution
+# --------------------------------------------
+#
+# Kai Zhang (cskaizhang@gmail.com)
+# https://github.com/cszn
+# From 2019/03--2021/08
+# --------------------------------------------
+"""
+
+import numpy as np
+import cv2
+import torch
+
+from functools import partial
+import random
+from scipy import ndimage
+import scipy
+import scipy.stats as ss
+from scipy.interpolate import interp2d
+from scipy.linalg import orth
+import albumentations
+
+import ldm.modules.image_degradation.utils_image as util
+
+
+def modcrop_np(img, sf):
+    '''
+    Args:
+        img: numpy image, WxH or WxHxC
+        sf: scale factor
+    Return:
+        cropped image
+    '''
+    w, h = img.shape[:2]
+    im = np.copy(img)
+    return im[:w - w % sf, :h - h % sf, ...]
+
+
+"""
+# --------------------------------------------
+# anisotropic Gaussian kernels
+# --------------------------------------------
+"""
+
+
+def analytic_kernel(k):
+    """Calculate the X4 kernel from the X2 kernel (for proof see appendix in paper)"""
+    k_size = k.shape[0]
+    # Calculate the big kernels size
+    big_k = np.zeros((3 * k_size - 2, 3 * k_size - 2))
+    # Loop over the small kernel to fill the big one
+    for r in range(k_size):
+        for c in range(k_size):
+            big_k[2 * r:2 * r + k_size, 2 * c:2 * c + k_size] += k[r, c] * k
+    # Crop the edges of the big kernel to ignore very small values and increase run time of SR
+    crop = k_size // 2
+    cropped_big_k = big_k[crop:-crop, crop:-crop]
+    # Normalize to 1
+    return cropped_big_k / cropped_big_k.sum()
+
+
+def anisotropic_Gaussian(ksize=15, theta=np.pi, l1=6, l2=6):
+    """ generate an anisotropic Gaussian kernel
+    Args:
+        ksize : e.g., 15, kernel size
+        theta : [0,  pi], rotation angle range
+        l1    : [0.1,50], scaling of eigenvalues
+        l2    : [0.1,l1], scaling of eigenvalues
+        If l1 = l2, will get an isotropic Gaussian kernel.
+    Returns:
+        k     : kernel
+    """
+
+    v = np.dot(np.array([[np.cos(theta), -np.sin(theta)], [np.sin(theta), np.cos(theta)]]), np.array([1., 0.]))
+    V = np.array([[v[0], v[1]], [v[1], -v[0]]])
+    D = np.array([[l1, 0], [0, l2]])
+    Sigma = np.dot(np.dot(V, D), np.linalg.inv(V))
+    k = gm_blur_kernel(mean=[0, 0], cov=Sigma, size=ksize)
+
+    return k
+
+
+def gm_blur_kernel(mean, cov, size=15):
+    center = size / 2.0 + 0.5
+    k = np.zeros([size, size])
+    for y in range(size):
+        for x in range(size):
+            cy = y - center + 1
+            cx = x - center + 1
+            k[y, x] = ss.multivariate_normal.pdf([cx, cy], mean=mean, cov=cov)
+
+    k = k / np.sum(k)
+    return k
+
+
+def shift_pixel(x, sf, upper_left=True):
+    """shift pixel for super-resolution with different scale factors
+    Args:
+        x: WxHxC or WxH
+        sf: scale factor
+        upper_left: shift direction
+    """
+    h, w = x.shape[:2]
+    shift = (sf - 1) * 0.5
+    xv, yv = np.arange(0, w, 1.0), np.arange(0, h, 1.0)
+    if upper_left:
+        x1 = xv + shift
+        y1 = yv + shift
+    else:
+        x1 = xv - shift
+        y1 = yv - shift
+
+    x1 = np.clip(x1, 0, w - 1)
+    y1 = np.clip(y1, 0, h - 1)
+
+    if x.ndim == 2:
+        x = interp2d(xv, yv, x)(x1, y1)
+    if x.ndim == 3:
+        for i in range(x.shape[-1]):
+            x[:, :, i] = interp2d(xv, yv, x[:, :, i])(x1, y1)
+
+    return x
+
+
+def blur(x, k):
+    '''
+    x: image, NxcxHxW
+    k: kernel, Nx1xhxw
+    '''
+    n, c = x.shape[:2]
+    p1, p2 = (k.shape[-2] - 1) // 2, (k.shape[-1] - 1) // 2
+    x = torch.nn.functional.pad(x, pad=(p1, p2, p1, p2), mode='replicate')
+    k = k.repeat(1, c, 1, 1)
+    k = k.view(-1, 1, k.shape[2], k.shape[3])
+    x = x.view(1, -1, x.shape[2], x.shape[3])
+    x = torch.nn.functional.conv2d(x, k, bias=None, stride=1, padding=0, groups=n * c)
+    x = x.view(n, c, x.shape[2], x.shape[3])
+
+    return x
+
+
+def gen_kernel(k_size=np.array([15, 15]), scale_factor=np.array([4, 4]), min_var=0.6, max_var=10., noise_level=0):
+    """"
+    # modified version of https://github.com/assafshocher/BlindSR_dataset_generator
+    # Kai Zhang
+    # min_var = 0.175 * sf  # variance of the gaussian kernel will be sampled between min_var and max_var
+    # max_var = 2.5 * sf
+    """
+    # Set random eigen-vals (lambdas) and angle (theta) for COV matrix
+    lambda_1 = min_var + np.random.rand() * (max_var - min_var)
+    lambda_2 = min_var + np.random.rand() * (max_var - min_var)
+    theta = np.random.rand() * np.pi  # random theta
+    noise = -noise_level + np.random.rand(*k_size) * noise_level * 2
+
+    # Set COV matrix using Lambdas and Theta
+    LAMBDA = np.diag([lambda_1, lambda_2])
+    Q = np.array([[np.cos(theta), -np.sin(theta)],
+                  [np.sin(theta), np.cos(theta)]])
+    SIGMA = Q @ LAMBDA @ Q.T
+    INV_SIGMA = np.linalg.inv(SIGMA)[None, None, :, :]
+
+    # Set expectation position (shifting kernel for aligned image)
+    MU = k_size // 2 - 0.5 * (scale_factor - 1)  # - 0.5 * (scale_factor - k_size % 2)
+    MU = MU[None, None, :, None]
+
+    # Create meshgrid for Gaussian
+    [X, Y] = np.meshgrid(range(k_size[0]), range(k_size[1]))
+    Z = np.stack([X, Y], 2)[:, :, :, None]
+
+    # Calcualte Gaussian for every pixel of the kernel
+    ZZ = Z - MU
+    ZZ_t = ZZ.transpose(0, 1, 3, 2)
+    raw_kernel = np.exp(-0.5 * np.squeeze(ZZ_t @ INV_SIGMA @ ZZ)) * (1 + noise)
+
+    # shift the kernel so it will be centered
+    # raw_kernel_centered = kernel_shift(raw_kernel, scale_factor)
+
+    # Normalize the kernel and return
+    # kernel = raw_kernel_centered / np.sum(raw_kernel_centered)
+    kernel = raw_kernel / np.sum(raw_kernel)
+    return kernel
+
+
+def fspecial_gaussian(hsize, sigma):
+    hsize = [hsize, hsize]
+    siz = [(hsize[0] - 1.0) / 2.0, (hsize[1] - 1.0) / 2.0]
+    std = sigma
+    [x, y] = np.meshgrid(np.arange(-siz[1], siz[1] + 1), np.arange(-siz[0], siz[0] + 1))
+    arg = -(x * x + y * y) / (2 * std * std)
+    h = np.exp(arg)
+    h[h < scipy.finfo(float).eps * h.max()] = 0
+    sumh = h.sum()
+    if sumh != 0:
+        h = h / sumh
+    return h
+
+
+def fspecial_laplacian(alpha):
+    alpha = max([0, min([alpha, 1])])
+    h1 = alpha / (alpha + 1)
+    h2 = (1 - alpha) / (alpha + 1)
+    h = [[h1, h2, h1], [h2, -4 / (alpha + 1), h2], [h1, h2, h1]]
+    h = np.array(h)
+    return h
+
+
+def fspecial(filter_type, *args, **kwargs):
+    '''
+    python code from:
+    https://github.com/ronaldosena/imagens-medicas-2/blob/40171a6c259edec7827a6693a93955de2bd39e76/Aulas/aula_2_-_uniform_filter/matlab_fspecial.py
+    '''
+    if filter_type == 'gaussian':
+        return fspecial_gaussian(*args, **kwargs)
+    if filter_type == 'laplacian':
+        return fspecial_laplacian(*args, **kwargs)
+
+
+"""
+# --------------------------------------------
+# degradation models
+# --------------------------------------------
+"""
+
+
+def bicubic_degradation(x, sf=3):
+    '''
+    Args:
+        x: HxWxC image, [0, 1]
+        sf: down-scale factor
+    Return:
+        bicubicly downsampled LR image
+    '''
+    x = util.imresize_np(x, scale=1 / sf)
+    return x
+
+
+def srmd_degradation(x, k, sf=3):
+    ''' blur + bicubic downsampling
+    Args:
+        x: HxWxC image, [0, 1]
+        k: hxw, double
+        sf: down-scale factor
+    Return:
+        downsampled LR image
+    Reference:
+        @inproceedings{zhang2018learning,
+          title={Learning a single convolutional super-resolution network for multiple degradations},
+          author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei},
+          booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
+          pages={3262--3271},
+          year={2018}
+        }
+    '''
+    x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap')  # 'nearest' | 'mirror'
+    x = bicubic_degradation(x, sf=sf)
+    return x
+
+
+def dpsr_degradation(x, k, sf=3):
+    ''' bicubic downsampling + blur
+    Args:
+        x: HxWxC image, [0, 1]
+        k: hxw, double
+        sf: down-scale factor
+    Return:
+        downsampled LR image
+    Reference:
+        @inproceedings{zhang2019deep,
+          title={Deep Plug-and-Play Super-Resolution for Arbitrary Blur Kernels},
+          author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei},
+          booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
+          pages={1671--1681},
+          year={2019}
+        }
+    '''
+    x = bicubic_degradation(x, sf=sf)
+    x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap')
+    return x
+
+
+def classical_degradation(x, k, sf=3):
+    ''' blur + downsampling
+    Args:
+        x: HxWxC image, [0, 1]/[0, 255]
+        k: hxw, double
+        sf: down-scale factor
+    Return:
+        downsampled LR image
+    '''
+    x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap')
+    # x = filters.correlate(x, np.expand_dims(np.flip(k), axis=2))
+    st = 0
+    return x[st::sf, st::sf, ...]
+
+
+def add_sharpening(img, weight=0.5, radius=50, threshold=10):
+    """USM sharpening. borrowed from real-ESRGAN
+    Input image: I; Blurry image: B.
+    1. K = I + weight * (I - B)
+    2. Mask = 1 if abs(I - B) > threshold, else: 0
+    3. Blur mask:
+    4. Out = Mask * K + (1 - Mask) * I
+    Args:
+        img (Numpy array): Input image, HWC, BGR; float32, [0, 1].
+        weight (float): Sharp weight. Default: 1.
+        radius (float): Kernel size of Gaussian blur. Default: 50.
+        threshold (int):
+    """
+    if radius % 2 == 0:
+        radius += 1
+    blur = cv2.GaussianBlur(img, (radius, radius), 0)
+    residual = img - blur
+    mask = np.abs(residual) * 255 > threshold
+    mask = mask.astype('float32')
+    soft_mask = cv2.GaussianBlur(mask, (radius, radius), 0)
+
+    K = img + weight * residual
+    K = np.clip(K, 0, 1)
+    return soft_mask * K + (1 - soft_mask) * img
+
+
+def add_blur(img, sf=4):
+    wd2 = 4.0 + sf
+    wd = 2.0 + 0.2 * sf
+    if random.random() < 0.5:
+        l1 = wd2 * random.random()
+        l2 = wd2 * random.random()
+        k = anisotropic_Gaussian(ksize=2 * random.randint(2, 11) + 3, theta=random.random() * np.pi, l1=l1, l2=l2)
+    else:
+        k = fspecial('gaussian', 2 * random.randint(2, 11) + 3, wd * random.random())
+    img = ndimage.filters.convolve(img, np.expand_dims(k, axis=2), mode='mirror')
+
+    return img
+
+
+def add_resize(img, sf=4):
+    rnum = np.random.rand()
+    if rnum > 0.8:  # up
+        sf1 = random.uniform(1, 2)
+    elif rnum < 0.7:  # down
+        sf1 = random.uniform(0.5 / sf, 1)
+    else:
+        sf1 = 1.0
+    img = cv2.resize(img, (int(sf1 * img.shape[1]), int(sf1 * img.shape[0])), interpolation=random.choice([1, 2, 3]))
+    img = np.clip(img, 0.0, 1.0)
+
+    return img
+
+
+# def add_Gaussian_noise(img, noise_level1=2, noise_level2=25):
+#     noise_level = random.randint(noise_level1, noise_level2)
+#     rnum = np.random.rand()
+#     if rnum > 0.6:  # add color Gaussian noise
+#         img += np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
+#     elif rnum < 0.4:  # add grayscale Gaussian noise
+#         img += np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
+#     else:  # add  noise
+#         L = noise_level2 / 255.
+#         D = np.diag(np.random.rand(3))
+#         U = orth(np.random.rand(3, 3))
+#         conv = np.dot(np.dot(np.transpose(U), D), U)
+#         img += np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
+#     img = np.clip(img, 0.0, 1.0)
+#     return img
+
+def add_Gaussian_noise(img, noise_level1=2, noise_level2=25):
+    noise_level = random.randint(noise_level1, noise_level2)
+    rnum = np.random.rand()
+    if rnum > 0.6:  # add color Gaussian noise
+        img = img + np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
+    elif rnum < 0.4:  # add grayscale Gaussian noise
+        img = img + np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
+    else:  # add  noise
+        L = noise_level2 / 255.
+        D = np.diag(np.random.rand(3))
+        U = orth(np.random.rand(3, 3))
+        conv = np.dot(np.dot(np.transpose(U), D), U)
+        img = img + np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
+    img = np.clip(img, 0.0, 1.0)
+    return img
+
+
+def add_speckle_noise(img, noise_level1=2, noise_level2=25):
+    noise_level = random.randint(noise_level1, noise_level2)
+    img = np.clip(img, 0.0, 1.0)
+    rnum = random.random()
+    if rnum > 0.6:
+        img += img * np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
+    elif rnum < 0.4:
+        img += img * np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
+    else:
+        L = noise_level2 / 255.
+        D = np.diag(np.random.rand(3))
+        U = orth(np.random.rand(3, 3))
+        conv = np.dot(np.dot(np.transpose(U), D), U)
+        img += img * np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
+    img = np.clip(img, 0.0, 1.0)
+    return img
+
+
+def add_Poisson_noise(img):
+    img = np.clip((img * 255.0).round(), 0, 255) / 255.
+    vals = 10 ** (2 * random.random() + 2.0)  # [2, 4]
+    if random.random() < 0.5:
+        img = np.random.poisson(img * vals).astype(np.float32) / vals
+    else:
+        img_gray = np.dot(img[..., :3], [0.299, 0.587, 0.114])
+        img_gray = np.clip((img_gray * 255.0).round(), 0, 255) / 255.
+        noise_gray = np.random.poisson(img_gray * vals).astype(np.float32) / vals - img_gray
+        img += noise_gray[:, :, np.newaxis]
+    img = np.clip(img, 0.0, 1.0)
+    return img
+
+
+def add_JPEG_noise(img):
+    quality_factor = random.randint(30, 95)
+    img = cv2.cvtColor(util.single2uint(img), cv2.COLOR_RGB2BGR)
+    result, encimg = cv2.imencode('.jpg', img, [int(cv2.IMWRITE_JPEG_QUALITY), quality_factor])
+    img = cv2.imdecode(encimg, 1)
+    img = cv2.cvtColor(util.uint2single(img), cv2.COLOR_BGR2RGB)
+    return img
+
+
+def random_crop(lq, hq, sf=4, lq_patchsize=64):
+    h, w = lq.shape[:2]
+    rnd_h = random.randint(0, h - lq_patchsize)
+    rnd_w = random.randint(0, w - lq_patchsize)
+    lq = lq[rnd_h:rnd_h + lq_patchsize, rnd_w:rnd_w + lq_patchsize, :]
+
+    rnd_h_H, rnd_w_H = int(rnd_h * sf), int(rnd_w * sf)
+    hq = hq[rnd_h_H:rnd_h_H + lq_patchsize * sf, rnd_w_H:rnd_w_H + lq_patchsize * sf, :]
+    return lq, hq
+
+
+def degradation_bsrgan(img, sf=4, lq_patchsize=72, isp_model=None):
+    """
+    This is the degradation model of BSRGAN from the paper
+    "Designing a Practical Degradation Model for Deep Blind Image Super-Resolution"
+    ----------
+    img: HXWXC, [0, 1], its size should be large than (lq_patchsizexsf)x(lq_patchsizexsf)
+    sf: scale factor
+    isp_model: camera ISP model
+    Returns
+    -------
+    img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1]
+    hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1]
+    """
+    isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25
+    sf_ori = sf
+
+    h1, w1 = img.shape[:2]
+    img = img.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...]  # mod crop
+    h, w = img.shape[:2]
+
+    if h < lq_patchsize * sf or w < lq_patchsize * sf:
+        raise ValueError(f'img size ({h1}X{w1}) is too small!')
+
+    hq = img.copy()
+
+    if sf == 4 and random.random() < scale2_prob:  # downsample1
+        if np.random.rand() < 0.5:
+            img = cv2.resize(img, (int(1 / 2 * img.shape[1]), int(1 / 2 * img.shape[0])),
+                             interpolation=random.choice([1, 2, 3]))
+        else:
+            img = util.imresize_np(img, 1 / 2, True)
+        img = np.clip(img, 0.0, 1.0)
+        sf = 2
+
+    shuffle_order = random.sample(range(7), 7)
+    idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3)
+    if idx1 > idx2:  # keep downsample3 last
+        shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1]
+
+    for i in shuffle_order:
+
+        if i == 0:
+            img = add_blur(img, sf=sf)
+
+        elif i == 1:
+            img = add_blur(img, sf=sf)
+
+        elif i == 2:
+            a, b = img.shape[1], img.shape[0]
+            # downsample2
+            if random.random() < 0.75:
+                sf1 = random.uniform(1, 2 * sf)
+                img = cv2.resize(img, (int(1 / sf1 * img.shape[1]), int(1 / sf1 * img.shape[0])),
+                                 interpolation=random.choice([1, 2, 3]))
+            else:
+                k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf))
+                k_shifted = shift_pixel(k, sf)
+                k_shifted = k_shifted / k_shifted.sum()  # blur with shifted kernel
+                img = ndimage.filters.convolve(img, np.expand_dims(k_shifted, axis=2), mode='mirror')
+                img = img[0::sf, 0::sf, ...]  # nearest downsampling
+            img = np.clip(img, 0.0, 1.0)
+
+        elif i == 3:
+            # downsample3
+            img = cv2.resize(img, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3]))
+            img = np.clip(img, 0.0, 1.0)
+
+        elif i == 4:
+            # add Gaussian noise
+            img = add_Gaussian_noise(img, noise_level1=2, noise_level2=25)
+
+        elif i == 5:
+            # add JPEG noise
+            if random.random() < jpeg_prob:
+                img = add_JPEG_noise(img)
+
+        elif i == 6:
+            # add processed camera sensor noise
+            if random.random() < isp_prob and isp_model is not None:
+                with torch.no_grad():
+                    img, hq = isp_model.forward(img.copy(), hq)
+
+    # add final JPEG compression noise
+    img = add_JPEG_noise(img)
+
+    # random crop
+    img, hq = random_crop(img, hq, sf_ori, lq_patchsize)
+
+    return img, hq
+
+
+# todo no isp_model?
+def degradation_bsrgan_variant(image, sf=4, isp_model=None):
+    """
+    This is the degradation model of BSRGAN from the paper
+    "Designing a Practical Degradation Model for Deep Blind Image Super-Resolution"
+    ----------
+    sf: scale factor
+    isp_model: camera ISP model
+    Returns
+    -------
+    img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1]
+    hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1]
+    """
+    image = util.uint2single(image)
+    isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25
+    sf_ori = sf
+
+    h1, w1 = image.shape[:2]
+    image = image.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...]  # mod crop
+    h, w = image.shape[:2]
+
+    hq = image.copy()
+
+    if sf == 4 and random.random() < scale2_prob:  # downsample1
+        if np.random.rand() < 0.5:
+            image = cv2.resize(image, (int(1 / 2 * image.shape[1]), int(1 / 2 * image.shape[0])),
+                               interpolation=random.choice([1, 2, 3]))
+        else:
+            image = util.imresize_np(image, 1 / 2, True)
+        image = np.clip(image, 0.0, 1.0)
+        sf = 2
+
+    shuffle_order = random.sample(range(7), 7)
+    idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3)
+    if idx1 > idx2:  # keep downsample3 last
+        shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1]
+
+    for i in shuffle_order:
+
+        if i == 0:
+            image = add_blur(image, sf=sf)
+
+        elif i == 1:
+            image = add_blur(image, sf=sf)
+
+        elif i == 2:
+            a, b = image.shape[1], image.shape[0]
+            # downsample2
+            if random.random() < 0.75:
+                sf1 = random.uniform(1, 2 * sf)
+                image = cv2.resize(image, (int(1 / sf1 * image.shape[1]), int(1 / sf1 * image.shape[0])),
+                                   interpolation=random.choice([1, 2, 3]))
+            else:
+                k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf))
+                k_shifted = shift_pixel(k, sf)
+                k_shifted = k_shifted / k_shifted.sum()  # blur with shifted kernel
+                image = ndimage.filters.convolve(image, np.expand_dims(k_shifted, axis=2), mode='mirror')
+                image = image[0::sf, 0::sf, ...]  # nearest downsampling
+            image = np.clip(image, 0.0, 1.0)
+
+        elif i == 3:
+            # downsample3
+            image = cv2.resize(image, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3]))
+            image = np.clip(image, 0.0, 1.0)
+
+        elif i == 4:
+            # add Gaussian noise
+            image = add_Gaussian_noise(image, noise_level1=2, noise_level2=25)
+
+        elif i == 5:
+            # add JPEG noise
+            if random.random() < jpeg_prob:
+                image = add_JPEG_noise(image)
+
+        # elif i == 6:
+        #     # add processed camera sensor noise
+        #     if random.random() < isp_prob and isp_model is not None:
+        #         with torch.no_grad():
+        #             img, hq = isp_model.forward(img.copy(), hq)
+
+    # add final JPEG compression noise
+    image = add_JPEG_noise(image)
+    image = util.single2uint(image)
+    example = {"image":image}
+    return example
+
+
+# TODO incase there is a pickle error one needs to replace a += x with a = a + x in add_speckle_noise etc...
+def degradation_bsrgan_plus(img, sf=4, shuffle_prob=0.5, use_sharp=True, lq_patchsize=64, isp_model=None):
+    """
+    This is an extended degradation model by combining
+    the degradation models of BSRGAN and Real-ESRGAN
+    ----------
+    img: HXWXC, [0, 1], its size should be large than (lq_patchsizexsf)x(lq_patchsizexsf)
+    sf: scale factor
+    use_shuffle: the degradation shuffle
+    use_sharp: sharpening the img
+    Returns
+    -------
+    img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1]
+    hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1]
+    """
+
+    h1, w1 = img.shape[:2]
+    img = img.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...]  # mod crop
+    h, w = img.shape[:2]
+
+    if h < lq_patchsize * sf or w < lq_patchsize * sf:
+        raise ValueError(f'img size ({h1}X{w1}) is too small!')
+
+    if use_sharp:
+        img = add_sharpening(img)
+    hq = img.copy()
+
+    if random.random() < shuffle_prob:
+        shuffle_order = random.sample(range(13), 13)
+    else:
+        shuffle_order = list(range(13))
+        # local shuffle for noise, JPEG is always the last one
+        shuffle_order[2:6] = random.sample(shuffle_order[2:6], len(range(2, 6)))
+        shuffle_order[9:13] = random.sample(shuffle_order[9:13], len(range(9, 13)))
+
+    poisson_prob, speckle_prob, isp_prob = 0.1, 0.1, 0.1
+
+    for i in shuffle_order:
+        if i == 0:
+            img = add_blur(img, sf=sf)
+        elif i == 1:
+            img = add_resize(img, sf=sf)
+        elif i == 2:
+            img = add_Gaussian_noise(img, noise_level1=2, noise_level2=25)
+        elif i == 3:
+            if random.random() < poisson_prob:
+                img = add_Poisson_noise(img)
+        elif i == 4:
+            if random.random() < speckle_prob:
+                img = add_speckle_noise(img)
+        elif i == 5:
+            if random.random() < isp_prob and isp_model is not None:
+                with torch.no_grad():
+                    img, hq = isp_model.forward(img.copy(), hq)
+        elif i == 6:
+            img = add_JPEG_noise(img)
+        elif i == 7:
+            img = add_blur(img, sf=sf)
+        elif i == 8:
+            img = add_resize(img, sf=sf)
+        elif i == 9:
+            img = add_Gaussian_noise(img, noise_level1=2, noise_level2=25)
+        elif i == 10:
+            if random.random() < poisson_prob:
+                img = add_Poisson_noise(img)
+        elif i == 11:
+            if random.random() < speckle_prob:
+                img = add_speckle_noise(img)
+        elif i == 12:
+            if random.random() < isp_prob and isp_model is not None:
+                with torch.no_grad():
+                    img, hq = isp_model.forward(img.copy(), hq)
+        else:
+            print('check the shuffle!')
+
+    # resize to desired size
+    img = cv2.resize(img, (int(1 / sf * hq.shape[1]), int(1 / sf * hq.shape[0])),
+                     interpolation=random.choice([1, 2, 3]))
+
+    # add final JPEG compression noise
+    img = add_JPEG_noise(img)
+
+    # random crop
+    img, hq = random_crop(img, hq, sf, lq_patchsize)
+
+    return img, hq
+
+
+if __name__ == '__main__':
+	print("hey")
+	img = util.imread_uint('utils/test.png', 3)
+	print(img)
+	img = util.uint2single(img)
+	print(img)
+	img = img[:448, :448]
+	h = img.shape[0] // 4
+	print("resizing to", h)
+	sf = 4
+	deg_fn = partial(degradation_bsrgan_variant, sf=sf)
+	for i in range(20):
+		print(i)
+		img_lq = deg_fn(img)
+		print(img_lq)
+		img_lq_bicubic = albumentations.SmallestMaxSize(max_size=h, interpolation=cv2.INTER_CUBIC)(image=img)["image"]
+		print(img_lq.shape)
+		print("bicubic", img_lq_bicubic.shape)
+		print(img_hq.shape)
+		lq_nearest = cv2.resize(util.single2uint(img_lq), (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])),
+		                        interpolation=0)
+		lq_bicubic_nearest = cv2.resize(util.single2uint(img_lq_bicubic), (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])),
+		                        interpolation=0)
+		img_concat = np.concatenate([lq_bicubic_nearest, lq_nearest, util.single2uint(img_hq)], axis=1)
+		util.imsave(img_concat, str(i) + '.png')
+
+

examples/tutorial/stable_diffusion/ldm/modules/image_degradation/bsrgan_light.py

@@ -0,0 +1,650 @@
+# -*- coding: utf-8 -*-
+import numpy as np
+import cv2
+import torch
+
+from functools import partial
+import random
+from scipy import ndimage
+import scipy
+import scipy.stats as ss
+from scipy.interpolate import interp2d
+from scipy.linalg import orth
+import albumentations
+
+import ldm.modules.image_degradation.utils_image as util
+
+"""
+# --------------------------------------------
+# Super-Resolution
+# --------------------------------------------
+#
+# Kai Zhang (cskaizhang@gmail.com)
+# https://github.com/cszn
+# From 2019/03--2021/08
+# --------------------------------------------
+"""
+
+
+def modcrop_np(img, sf):
+    '''
+    Args:
+        img: numpy image, WxH or WxHxC
+        sf: scale factor
+    Return:
+        cropped image
+    '''
+    w, h = img.shape[:2]
+    im = np.copy(img)
+    return im[:w - w % sf, :h - h % sf, ...]
+
+
+"""
+# --------------------------------------------
+# anisotropic Gaussian kernels
+# --------------------------------------------
+"""
+
+
+def analytic_kernel(k):
+    """Calculate the X4 kernel from the X2 kernel (for proof see appendix in paper)"""
+    k_size = k.shape[0]
+    # Calculate the big kernels size
+    big_k = np.zeros((3 * k_size - 2, 3 * k_size - 2))
+    # Loop over the small kernel to fill the big one
+    for r in range(k_size):
+        for c in range(k_size):
+            big_k[2 * r:2 * r + k_size, 2 * c:2 * c + k_size] += k[r, c] * k
+    # Crop the edges of the big kernel to ignore very small values and increase run time of SR
+    crop = k_size // 2
+    cropped_big_k = big_k[crop:-crop, crop:-crop]
+    # Normalize to 1
+    return cropped_big_k / cropped_big_k.sum()
+
+
+def anisotropic_Gaussian(ksize=15, theta=np.pi, l1=6, l2=6):
+    """ generate an anisotropic Gaussian kernel
+    Args:
+        ksize : e.g., 15, kernel size
+        theta : [0,  pi], rotation angle range
+        l1    : [0.1,50], scaling of eigenvalues
+        l2    : [0.1,l1], scaling of eigenvalues
+        If l1 = l2, will get an isotropic Gaussian kernel.
+    Returns:
+        k     : kernel
+    """
+
+    v = np.dot(np.array([[np.cos(theta), -np.sin(theta)], [np.sin(theta), np.cos(theta)]]), np.array([1., 0.]))
+    V = np.array([[v[0], v[1]], [v[1], -v[0]]])
+    D = np.array([[l1, 0], [0, l2]])
+    Sigma = np.dot(np.dot(V, D), np.linalg.inv(V))
+    k = gm_blur_kernel(mean=[0, 0], cov=Sigma, size=ksize)
+
+    return k
+
+
+def gm_blur_kernel(mean, cov, size=15):
+    center = size / 2.0 + 0.5
+    k = np.zeros([size, size])
+    for y in range(size):
+        for x in range(size):
+            cy = y - center + 1
+            cx = x - center + 1
+            k[y, x] = ss.multivariate_normal.pdf([cx, cy], mean=mean, cov=cov)
+
+    k = k / np.sum(k)
+    return k
+
+
+def shift_pixel(x, sf, upper_left=True):
+    """shift pixel for super-resolution with different scale factors
+    Args:
+        x: WxHxC or WxH
+        sf: scale factor
+        upper_left: shift direction
+    """
+    h, w = x.shape[:2]
+    shift = (sf - 1) * 0.5
+    xv, yv = np.arange(0, w, 1.0), np.arange(0, h, 1.0)
+    if upper_left:
+        x1 = xv + shift
+        y1 = yv + shift
+    else:
+        x1 = xv - shift
+        y1 = yv - shift
+
+    x1 = np.clip(x1, 0, w - 1)
+    y1 = np.clip(y1, 0, h - 1)
+
+    if x.ndim == 2:
+        x = interp2d(xv, yv, x)(x1, y1)
+    if x.ndim == 3:
+        for i in range(x.shape[-1]):
+            x[:, :, i] = interp2d(xv, yv, x[:, :, i])(x1, y1)
+
+    return x
+
+
+def blur(x, k):
+    '''
+    x: image, NxcxHxW
+    k: kernel, Nx1xhxw
+    '''
+    n, c = x.shape[:2]
+    p1, p2 = (k.shape[-2] - 1) // 2, (k.shape[-1] - 1) // 2
+    x = torch.nn.functional.pad(x, pad=(p1, p2, p1, p2), mode='replicate')
+    k = k.repeat(1, c, 1, 1)
+    k = k.view(-1, 1, k.shape[2], k.shape[3])
+    x = x.view(1, -1, x.shape[2], x.shape[3])
+    x = torch.nn.functional.conv2d(x, k, bias=None, stride=1, padding=0, groups=n * c)
+    x = x.view(n, c, x.shape[2], x.shape[3])
+
+    return x
+
+
+def gen_kernel(k_size=np.array([15, 15]), scale_factor=np.array([4, 4]), min_var=0.6, max_var=10., noise_level=0):
+    """"
+    # modified version of https://github.com/assafshocher/BlindSR_dataset_generator
+    # Kai Zhang
+    # min_var = 0.175 * sf  # variance of the gaussian kernel will be sampled between min_var and max_var
+    # max_var = 2.5 * sf
+    """
+    # Set random eigen-vals (lambdas) and angle (theta) for COV matrix
+    lambda_1 = min_var + np.random.rand() * (max_var - min_var)
+    lambda_2 = min_var + np.random.rand() * (max_var - min_var)
+    theta = np.random.rand() * np.pi  # random theta
+    noise = -noise_level + np.random.rand(*k_size) * noise_level * 2
+
+    # Set COV matrix using Lambdas and Theta
+    LAMBDA = np.diag([lambda_1, lambda_2])
+    Q = np.array([[np.cos(theta), -np.sin(theta)],
+                  [np.sin(theta), np.cos(theta)]])
+    SIGMA = Q @ LAMBDA @ Q.T
+    INV_SIGMA = np.linalg.inv(SIGMA)[None, None, :, :]
+
+    # Set expectation position (shifting kernel for aligned image)
+    MU = k_size // 2 - 0.5 * (scale_factor - 1)  # - 0.5 * (scale_factor - k_size % 2)
+    MU = MU[None, None, :, None]
+
+    # Create meshgrid for Gaussian
+    [X, Y] = np.meshgrid(range(k_size[0]), range(k_size[1]))
+    Z = np.stack([X, Y], 2)[:, :, :, None]
+
+    # Calcualte Gaussian for every pixel of the kernel
+    ZZ = Z - MU
+    ZZ_t = ZZ.transpose(0, 1, 3, 2)
+    raw_kernel = np.exp(-0.5 * np.squeeze(ZZ_t @ INV_SIGMA @ ZZ)) * (1 + noise)
+
+    # shift the kernel so it will be centered
+    # raw_kernel_centered = kernel_shift(raw_kernel, scale_factor)
+
+    # Normalize the kernel and return
+    # kernel = raw_kernel_centered / np.sum(raw_kernel_centered)
+    kernel = raw_kernel / np.sum(raw_kernel)
+    return kernel
+
+
+def fspecial_gaussian(hsize, sigma):
+    hsize = [hsize, hsize]
+    siz = [(hsize[0] - 1.0) / 2.0, (hsize[1] - 1.0) / 2.0]
+    std = sigma
+    [x, y] = np.meshgrid(np.arange(-siz[1], siz[1] + 1), np.arange(-siz[0], siz[0] + 1))
+    arg = -(x * x + y * y) / (2 * std * std)
+    h = np.exp(arg)
+    h[h < scipy.finfo(float).eps * h.max()] = 0
+    sumh = h.sum()
+    if sumh != 0:
+        h = h / sumh
+    return h
+
+
+def fspecial_laplacian(alpha):
+    alpha = max([0, min([alpha, 1])])
+    h1 = alpha / (alpha + 1)
+    h2 = (1 - alpha) / (alpha + 1)
+    h = [[h1, h2, h1], [h2, -4 / (alpha + 1), h2], [h1, h2, h1]]
+    h = np.array(h)
+    return h
+
+
+def fspecial(filter_type, *args, **kwargs):
+    '''
+    python code from:
+    https://github.com/ronaldosena/imagens-medicas-2/blob/40171a6c259edec7827a6693a93955de2bd39e76/Aulas/aula_2_-_uniform_filter/matlab_fspecial.py
+    '''
+    if filter_type == 'gaussian':
+        return fspecial_gaussian(*args, **kwargs)
+    if filter_type == 'laplacian':
+        return fspecial_laplacian(*args, **kwargs)
+
+
+"""
+# --------------------------------------------
+# degradation models
+# --------------------------------------------
+"""
+
+
+def bicubic_degradation(x, sf=3):
+    '''
+    Args:
+        x: HxWxC image, [0, 1]
+        sf: down-scale factor
+    Return:
+        bicubicly downsampled LR image
+    '''
+    x = util.imresize_np(x, scale=1 / sf)
+    return x
+
+
+def srmd_degradation(x, k, sf=3):
+    ''' blur + bicubic downsampling
+    Args:
+        x: HxWxC image, [0, 1]
+        k: hxw, double
+        sf: down-scale factor
+    Return:
+        downsampled LR image
+    Reference:
+        @inproceedings{zhang2018learning,
+          title={Learning a single convolutional super-resolution network for multiple degradations},
+          author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei},
+          booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
+          pages={3262--3271},
+          year={2018}
+        }
+    '''
+    x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap')  # 'nearest' | 'mirror'
+    x = bicubic_degradation(x, sf=sf)
+    return x
+
+
+def dpsr_degradation(x, k, sf=3):
+    ''' bicubic downsampling + blur
+    Args:
+        x: HxWxC image, [0, 1]
+        k: hxw, double
+        sf: down-scale factor
+    Return:
+        downsampled LR image
+    Reference:
+        @inproceedings{zhang2019deep,
+          title={Deep Plug-and-Play Super-Resolution for Arbitrary Blur Kernels},
+          author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei},
+          booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
+          pages={1671--1681},
+          year={2019}
+        }
+    '''
+    x = bicubic_degradation(x, sf=sf)
+    x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap')
+    return x
+
+
+def classical_degradation(x, k, sf=3):
+    ''' blur + downsampling
+    Args:
+        x: HxWxC image, [0, 1]/[0, 255]
+        k: hxw, double
+        sf: down-scale factor
+    Return:
+        downsampled LR image
+    '''
+    x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap')
+    # x = filters.correlate(x, np.expand_dims(np.flip(k), axis=2))
+    st = 0
+    return x[st::sf, st::sf, ...]
+
+
+def add_sharpening(img, weight=0.5, radius=50, threshold=10):
+    """USM sharpening. borrowed from real-ESRGAN
+    Input image: I; Blurry image: B.
+    1. K = I + weight * (I - B)
+    2. Mask = 1 if abs(I - B) > threshold, else: 0
+    3. Blur mask:
+    4. Out = Mask * K + (1 - Mask) * I
+    Args:
+        img (Numpy array): Input image, HWC, BGR; float32, [0, 1].
+        weight (float): Sharp weight. Default: 1.
+        radius (float): Kernel size of Gaussian blur. Default: 50.
+        threshold (int):
+    """
+    if radius % 2 == 0:
+        radius += 1
+    blur = cv2.GaussianBlur(img, (radius, radius), 0)
+    residual = img - blur
+    mask = np.abs(residual) * 255 > threshold
+    mask = mask.astype('float32')
+    soft_mask = cv2.GaussianBlur(mask, (radius, radius), 0)
+
+    K = img + weight * residual
+    K = np.clip(K, 0, 1)
+    return soft_mask * K + (1 - soft_mask) * img
+
+
+def add_blur(img, sf=4):
+    wd2 = 4.0 + sf
+    wd = 2.0 + 0.2 * sf
+
+    wd2 = wd2/4
+    wd = wd/4
+
+    if random.random() < 0.5:
+        l1 = wd2 * random.random()
+        l2 = wd2 * random.random()
+        k = anisotropic_Gaussian(ksize=random.randint(2, 11) + 3, theta=random.random() * np.pi, l1=l1, l2=l2)
+    else:
+        k = fspecial('gaussian', random.randint(2, 4) + 3, wd * random.random())
+    img = ndimage.filters.convolve(img, np.expand_dims(k, axis=2), mode='mirror')
+
+    return img
+
+
+def add_resize(img, sf=4):
+    rnum = np.random.rand()
+    if rnum > 0.8:  # up
+        sf1 = random.uniform(1, 2)
+    elif rnum < 0.7:  # down
+        sf1 = random.uniform(0.5 / sf, 1)
+    else:
+        sf1 = 1.0
+    img = cv2.resize(img, (int(sf1 * img.shape[1]), int(sf1 * img.shape[0])), interpolation=random.choice([1, 2, 3]))
+    img = np.clip(img, 0.0, 1.0)
+
+    return img
+
+
+# def add_Gaussian_noise(img, noise_level1=2, noise_level2=25):
+#     noise_level = random.randint(noise_level1, noise_level2)
+#     rnum = np.random.rand()
+#     if rnum > 0.6:  # add color Gaussian noise
+#         img += np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
+#     elif rnum < 0.4:  # add grayscale Gaussian noise
+#         img += np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
+#     else:  # add  noise
+#         L = noise_level2 / 255.
+#         D = np.diag(np.random.rand(3))
+#         U = orth(np.random.rand(3, 3))
+#         conv = np.dot(np.dot(np.transpose(U), D), U)
+#         img += np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
+#     img = np.clip(img, 0.0, 1.0)
+#     return img
+
+def add_Gaussian_noise(img, noise_level1=2, noise_level2=25):
+    noise_level = random.randint(noise_level1, noise_level2)
+    rnum = np.random.rand()
+    if rnum > 0.6:  # add color Gaussian noise
+        img = img + np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
+    elif rnum < 0.4:  # add grayscale Gaussian noise
+        img = img + np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
+    else:  # add  noise
+        L = noise_level2 / 255.
+        D = np.diag(np.random.rand(3))
+        U = orth(np.random.rand(3, 3))
+        conv = np.dot(np.dot(np.transpose(U), D), U)
+        img = img + np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
+    img = np.clip(img, 0.0, 1.0)
+    return img
+
+
+def add_speckle_noise(img, noise_level1=2, noise_level2=25):
+    noise_level = random.randint(noise_level1, noise_level2)
+    img = np.clip(img, 0.0, 1.0)
+    rnum = random.random()
+    if rnum > 0.6:
+        img += img * np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
+    elif rnum < 0.4:
+        img += img * np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
+    else:
+        L = noise_level2 / 255.
+        D = np.diag(np.random.rand(3))
+        U = orth(np.random.rand(3, 3))
+        conv = np.dot(np.dot(np.transpose(U), D), U)
+        img += img * np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
+    img = np.clip(img, 0.0, 1.0)
+    return img
+
+
+def add_Poisson_noise(img):
+    img = np.clip((img * 255.0).round(), 0, 255) / 255.
+    vals = 10 ** (2 * random.random() + 2.0)  # [2, 4]
+    if random.random() < 0.5:
+        img = np.random.poisson(img * vals).astype(np.float32) / vals
+    else:
+        img_gray = np.dot(img[..., :3], [0.299, 0.587, 0.114])
+        img_gray = np.clip((img_gray * 255.0).round(), 0, 255) / 255.
+        noise_gray = np.random.poisson(img_gray * vals).astype(np.float32) / vals - img_gray
+        img += noise_gray[:, :, np.newaxis]
+    img = np.clip(img, 0.0, 1.0)
+    return img
+
+
+def add_JPEG_noise(img):
+    quality_factor = random.randint(80, 95)
+    img = cv2.cvtColor(util.single2uint(img), cv2.COLOR_RGB2BGR)
+    result, encimg = cv2.imencode('.jpg', img, [int(cv2.IMWRITE_JPEG_QUALITY), quality_factor])
+    img = cv2.imdecode(encimg, 1)
+    img = cv2.cvtColor(util.uint2single(img), cv2.COLOR_BGR2RGB)
+    return img
+
+
+def random_crop(lq, hq, sf=4, lq_patchsize=64):
+    h, w = lq.shape[:2]
+    rnd_h = random.randint(0, h - lq_patchsize)
+    rnd_w = random.randint(0, w - lq_patchsize)
+    lq = lq[rnd_h:rnd_h + lq_patchsize, rnd_w:rnd_w + lq_patchsize, :]
+
+    rnd_h_H, rnd_w_H = int(rnd_h * sf), int(rnd_w * sf)
+    hq = hq[rnd_h_H:rnd_h_H + lq_patchsize * sf, rnd_w_H:rnd_w_H + lq_patchsize * sf, :]
+    return lq, hq
+
+
+def degradation_bsrgan(img, sf=4, lq_patchsize=72, isp_model=None):
+    """
+    This is the degradation model of BSRGAN from the paper
+    "Designing a Practical Degradation Model for Deep Blind Image Super-Resolution"
+    ----------
+    img: HXWXC, [0, 1], its size should be large than (lq_patchsizexsf)x(lq_patchsizexsf)
+    sf: scale factor
+    isp_model: camera ISP model
+    Returns
+    -------
+    img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1]
+    hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1]
+    """
+    isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25
+    sf_ori = sf
+
+    h1, w1 = img.shape[:2]
+    img = img.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...]  # mod crop
+    h, w = img.shape[:2]
+
+    if h < lq_patchsize * sf or w < lq_patchsize * sf:
+        raise ValueError(f'img size ({h1}X{w1}) is too small!')
+
+    hq = img.copy()
+
+    if sf == 4 and random.random() < scale2_prob:  # downsample1
+        if np.random.rand() < 0.5:
+            img = cv2.resize(img, (int(1 / 2 * img.shape[1]), int(1 / 2 * img.shape[0])),
+                             interpolation=random.choice([1, 2, 3]))
+        else:
+            img = util.imresize_np(img, 1 / 2, True)
+        img = np.clip(img, 0.0, 1.0)
+        sf = 2
+
+    shuffle_order = random.sample(range(7), 7)
+    idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3)
+    if idx1 > idx2:  # keep downsample3 last
+        shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1]
+
+    for i in shuffle_order:
+
+        if i == 0:
+            img = add_blur(img, sf=sf)
+
+        elif i == 1:
+            img = add_blur(img, sf=sf)
+
+        elif i == 2:
+            a, b = img.shape[1], img.shape[0]
+            # downsample2
+            if random.random() < 0.75:
+                sf1 = random.uniform(1, 2 * sf)
+                img = cv2.resize(img, (int(1 / sf1 * img.shape[1]), int(1 / sf1 * img.shape[0])),
+                                 interpolation=random.choice([1, 2, 3]))
+            else:
+                k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf))
+                k_shifted = shift_pixel(k, sf)
+                k_shifted = k_shifted / k_shifted.sum()  # blur with shifted kernel
+                img = ndimage.filters.convolve(img, np.expand_dims(k_shifted, axis=2), mode='mirror')
+                img = img[0::sf, 0::sf, ...]  # nearest downsampling
+            img = np.clip(img, 0.0, 1.0)
+
+        elif i == 3:
+            # downsample3
+            img = cv2.resize(img, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3]))
+            img = np.clip(img, 0.0, 1.0)
+
+        elif i == 4:
+            # add Gaussian noise
+            img = add_Gaussian_noise(img, noise_level1=2, noise_level2=8)
+
+        elif i == 5:
+            # add JPEG noise
+            if random.random() < jpeg_prob:
+                img = add_JPEG_noise(img)
+
+        elif i == 6:
+            # add processed camera sensor noise
+            if random.random() < isp_prob and isp_model is not None:
+                with torch.no_grad():
+                    img, hq = isp_model.forward(img.copy(), hq)
+
+    # add final JPEG compression noise
+    img = add_JPEG_noise(img)
+
+    # random crop
+    img, hq = random_crop(img, hq, sf_ori, lq_patchsize)
+
+    return img, hq
+
+
+# todo no isp_model?
+def degradation_bsrgan_variant(image, sf=4, isp_model=None):
+    """
+    This is the degradation model of BSRGAN from the paper
+    "Designing a Practical Degradation Model for Deep Blind Image Super-Resolution"
+    ----------
+    sf: scale factor
+    isp_model: camera ISP model
+    Returns
+    -------
+    img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1]
+    hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1]
+    """
+    image = util.uint2single(image)
+    isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25
+    sf_ori = sf
+
+    h1, w1 = image.shape[:2]
+    image = image.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...]  # mod crop
+    h, w = image.shape[:2]
+
+    hq = image.copy()
+
+    if sf == 4 and random.random() < scale2_prob:  # downsample1
+        if np.random.rand() < 0.5:
+            image = cv2.resize(image, (int(1 / 2 * image.shape[1]), int(1 / 2 * image.shape[0])),
+                               interpolation=random.choice([1, 2, 3]))
+        else:
+            image = util.imresize_np(image, 1 / 2, True)
+        image = np.clip(image, 0.0, 1.0)
+        sf = 2
+
+    shuffle_order = random.sample(range(7), 7)
+    idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3)
+    if idx1 > idx2:  # keep downsample3 last
+        shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1]
+
+    for i in shuffle_order:
+
+        if i == 0:
+            image = add_blur(image, sf=sf)
+
+        # elif i == 1:
+        #     image = add_blur(image, sf=sf)
+
+        if i == 0:
+            pass
+
+        elif i == 2:
+            a, b = image.shape[1], image.shape[0]
+            # downsample2
+            if random.random() < 0.8:
+                sf1 = random.uniform(1, 2 * sf)
+                image = cv2.resize(image, (int(1 / sf1 * image.shape[1]), int(1 / sf1 * image.shape[0])),
+                                   interpolation=random.choice([1, 2, 3]))
+            else:
+                k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf))
+                k_shifted = shift_pixel(k, sf)
+                k_shifted = k_shifted / k_shifted.sum()  # blur with shifted kernel
+                image = ndimage.filters.convolve(image, np.expand_dims(k_shifted, axis=2), mode='mirror')
+                image = image[0::sf, 0::sf, ...]  # nearest downsampling
+
+            image = np.clip(image, 0.0, 1.0)
+
+        elif i == 3:
+            # downsample3
+            image = cv2.resize(image, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3]))
+            image = np.clip(image, 0.0, 1.0)
+
+        elif i == 4:
+            # add Gaussian noise
+            image = add_Gaussian_noise(image, noise_level1=1, noise_level2=2)
+
+        elif i == 5:
+            # add JPEG noise
+            if random.random() < jpeg_prob:
+                image = add_JPEG_noise(image)
+        #
+        # elif i == 6:
+        #     # add processed camera sensor noise
+        #     if random.random() < isp_prob and isp_model is not None:
+        #         with torch.no_grad():
+        #             img, hq = isp_model.forward(img.copy(), hq)
+
+    # add final JPEG compression noise
+    image = add_JPEG_noise(image)
+    image = util.single2uint(image)
+    example = {"image": image}
+    return example
+
+
+
+
+if __name__ == '__main__':
+    print("hey")
+    img = util.imread_uint('utils/test.png', 3)
+    img = img[:448, :448]
+    h = img.shape[0] // 4
+    print("resizing to", h)
+    sf = 4
+    deg_fn = partial(degradation_bsrgan_variant, sf=sf)
+    for i in range(20):
+        print(i)
+        img_hq = img
+        img_lq = deg_fn(img)["image"]
+        img_hq, img_lq = util.uint2single(img_hq), util.uint2single(img_lq)
+        print(img_lq)
+        img_lq_bicubic = albumentations.SmallestMaxSize(max_size=h, interpolation=cv2.INTER_CUBIC)(image=img_hq)["image"]
+        print(img_lq.shape)
+        print("bicubic", img_lq_bicubic.shape)
+        print(img_hq.shape)
+        lq_nearest = cv2.resize(util.single2uint(img_lq), (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])),
+                                interpolation=0)
+        lq_bicubic_nearest = cv2.resize(util.single2uint(img_lq_bicubic),
+                                        (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])),
+                                        interpolation=0)
+        img_concat = np.concatenate([lq_bicubic_nearest, lq_nearest, util.single2uint(img_hq)], axis=1)
+        util.imsave(img_concat, str(i) + '.png')