[Feature] Support LLaMA-3 CPT and ST (#5619)

* support LLaMA-3

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Run pre-commit

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

applications/Colossal-LLaMA-2/version.txt

@@ -1 +0,0 @@
-0.0.1

applications/Colossal-LLaMA/README.md

@@ -1,6 +1,6 @@
 <div align="center">
 <h1>
-<img src="https://github.com/hpcaitech/public_assets/blob/main/applications/colossal-llama-2/colossalllam2.jpg?raw=true" width=800/>
+Colossal-LLaMA
 </h1>
 </div>
 
@@ -47,6 +47,7 @@
 - [Citations](#citations)
 
 ## News
+* [2024/4] Support continual pre-training and supervised fine-tuning of LLaMA-3.
 * [2024/01] [Construct Refined 13B Private Model With Just $5000 USD, Upgraded Colossal-AI Llama-2 Open Source](https://hpc-ai.com/blog/colossal-llama-2-13b).
 [[code]](https://github.com/hpcaitech/ColossalAI/tree/main/applications/Colossal-LLaMA-2)
 [[blog]](https://hpc-ai.com/blog/colossal-llama-2-13b)
@@ -289,7 +290,7 @@ Here is details about CLI arguments:
 
 #### 1. Install required packages
 ```
-cd Colossal-LLaMA-2
+cd Colossal-LLaMA
 pip install -r requirements.txt
 ```
 #### 2. Install `xentropy`, `layer_norm` and `rotary`
@@ -314,7 +315,7 @@ Initialize new tokenizer with additional Chinese tokens. Additional Chinese toke
 Command to initialize new tokenizer:
 ```bash
 export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION='python'
-python colossal_llama2/tokenizer/init_tokenizer.py \
+python colossal_llama/tokenizer/init_tokenizer.py \
     --source_tokenizer_dir "<SOURCE_TOKENIZER_DIR>" \
     --target_tokenizer_dir "<TARGET_TOKENIZER_DIR>" \
     --expand_tokens_file "<NEW_TOKENS_FILE>.jsonl"
@@ -328,7 +329,7 @@ Here is details about CLI arguments:
 Initialize the new model checkpoint by calculating the mean values from the original model checkpoint.
 Command to initialize new model checkpoint:
 ```bash
-python colossal_llama2/model/init_model.py \
+python colossal_llama/model/init_model.py \
     --source_model_and_tokenizer_path "<SOURCE_MODEL_AND_TOKENIZER_DIR>" \
     --target_tokenizer_path "<TARGET_TOKENIZER_DIR>" \
     --target_model_path "<TARGET_MODEL_DIR>"
@@ -362,18 +363,17 @@ Command to convert jsonl dataset to arrow format:
 python prepare_pretrain_dataset.py \
     --data_input_dirs "<JSONL_DIR_1>,<JSONL_DIR_2>,<JSONL_DIR_3>" \
     --tokenizer_dir "<TOKENIZER_DIR>" \
-    --data_cache_dir "jsonl_to_arrow_cache" \
-    --data_jsonl_output_dir "spliced_tokenized_output_jsonl" \
-    --data_arrow_output_dir "spliced_tokenized_output_arrow" \
+    --data_output_dirs "spliced tokenized output" \
     --max_length 4096 \
     --num_spliced_dataset_bins 10
 ```
 Here is details about CLI arguments:
 * Source data directory: `data_input_dirs`. Each `<JSONL_DIR>` can have multiple file in `jsonl` format.
 * Tokenizer directory: `tokenizer_dir`. Path to the tokenizer in Hugging Face format.
-* Data cache directory: `data_cache_dir`. Directory to store Hugging Face data cache. Default case will create `cache` folder locally.
-* Output directory for jsonl format: `data_jsonl_output_dir`. Output directory to store converted dataset in jsonl format.
-* Output directory for arrow format: `data_arrow_output_dir`. Output directory to store converted dataset in arrow format, which can be used for training directly.
+* Data output directory: `data_output_dirs`. Directory to store preprocessed output, including three sub-directories:
+  * `cache`: Directory to store Hugging Face data cache.
+  * `jsonl`: Output directory to store converted dataset in jsonl format.
+  * `arrow`: Output directory to store converted dataset in arrow format, which can be used for training directly.
 * Max length: `max_length`. Max length of spliced samples. Default value is 4096.
 * Number of bins for each category: `num_spliced_dataset_bins`. Number of bins for each category, used for bucket-based training.
 
@@ -392,13 +392,15 @@ Command to convert jsonl dataset to arrow format is similar to the command in [3
 python prepare_sft_dataset.py.py \
     --data_input_dirs "<JSONL_DIR_1>,<JSONL_DIR_2>,<JSONL_DIR_3>" \
     --tokenizer_dir "<TOKENIZER_DIR>" \
-    --data_cache_dir "jsonl_to_arrow_cache" \
-    --data_jsonl_output_dir "spliced_tokenized_output_jsonl" \
-    --data_arrow_output_dir "spliced_tokenized_output_arrow" \
+    --data_output_dirs "spliced tokenized output" \
     --max_length 4096 \
-    --num_spliced_dataset_bins 10
+    --num_spliced_dataset_bins 10 \
+    --llama_version 3
 ```
 
+Additional CLI arguments:
+* LLaMA verison: `llama_version`. Specify the LLaMA version.
+
 #### 4. Command Line Arguments for Training
 
 ##### 4.1 Arguments for Pretraining

applications/Colossal-LLaMA/colossal_llama/dataset/conversation.py

@@ -83,7 +83,7 @@ class Conversation:
         }
 
 
-conv = Conversation(
+LLaMA2_Conv = Conversation(
     system="A chat between a curious human and an artificial intelligence assistant. "
     "The assistant gives helpful, detailed, and polite answers to the human's questions.\n\n",
     roles=("Human", "Assistant"),
@@ -93,4 +93,14 @@ conv = Conversation(
     seps=["<s>", "</s>"],
 )
 
-default_conversation = conv
+LLaMA3_Conv = Conversation(
+    system="A chat between a curious human and an artificial intelligence assistant. "
+    "The assistant gives helpful, detailed, and polite answers to the human's questions.\n\n",
+    roles=("Human", "Assistant"),
+    messages=[],
+    offset=0,
+    sep_style=SeparatorStyle.ADD_BOS_EOS_TOKEN,
+    seps=["<|begin_of_text|>", "<|end_of_text|>"],
+)
+
+default_conversation = LLaMA3_Conv

applications/Colossal-LLaMA/colossal_llama/dataset/spliced_and_tokenized_dataset.py

@@ -12,6 +12,7 @@ from typing import Any, Callable, Dict, Iterable, List, Tuple, Union
 
 from datasets import dataset_dict
 from torch.utils.data import ConcatDataset, Dataset, IterableDataset
+from transformers import AutoTokenizer
 from transformers.models.llama.tokenization_llama import LlamaTokenizer
 from transformers.tokenization_utils import PreTrainedTokenizer
 
@@ -71,7 +72,7 @@ def supervised_tokenize_pretrain(
 
 def supervised_tokenize_sft(
     data_point: Dict[str, str],
-    tokenizer: LlamaTokenizer,
+    tokenizer: AutoTokenizer,
     conversation_template: Conversation = default_conversation,
     ignore_index: int = None,
     max_length: int = 4096,

applications/Colossal-LLaMA/inference_example.py

@@ -1,7 +1,7 @@
 import argparse
 
 import torch
-from colossal_llama2.dataset.conversation import default_conversation
+from colossal_llama.dataset.conversation import default_conversation
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
 from colossalai.logging import get_dist_logger

applications/Colossal-LLaMA/prepare_pretrain_dataset.py

@@ -11,12 +11,12 @@ import os
 import time
 from multiprocessing import cpu_count
 
-from colossal_llama2.dataset.spliced_and_tokenized_dataset import (
+from colossal_llama.dataset.spliced_and_tokenized_dataset import (
     ClosedToConstantLengthSplicedDataset,
     supervised_tokenize_pretrain,
 )
 from datasets import dataset_dict, load_dataset
-from transformers.models.llama.tokenization_llama import LlamaTokenizer
+from transformers import AutoTokenizer
 
 from colossalai.logging import get_dist_logger
 
@@ -35,35 +35,24 @@ def main():
     parser.add_argument(
         "--tokenizer_dir", type=str, required=True, default=None, help="A directory containing the tokenizer"
     )
-    parser.add_argument("--data_cache_dir", type=str, default="cache", help="Data cache directory")
-    parser.add_argument(
-        "--data_jsonl_output_dir",
-        type=str,
-        default="jsonl_output",
-        help="Output directory of spliced dataset with jsonl format",
-    )
-    parser.add_argument(
-        "--data_arrow_output_dir",
-        type=str,
-        default="arrow_output",
-        help="Output directory of spliced dataset with arrow format",
-    )
-    parser.add_argument("--max_length", type=int, default=4096, help="Max length of each spliced tokenized sequence")
+    parser.add_argument("--data_output_dirs", type=str, default="data_output_dirs", help="Data output directory")
+    parser.add_argument("--max_length", type=int, default=8192, help="Max length of each spliced tokenized sequence")
     parser.add_argument("--num_spliced_dataset_bins", type=int, default=10, help="Number of spliced dataset bins")
     args = parser.parse_args()
 
     if args.num_spliced_dataset_bins >= 100000:
         raise ValueError("Too many spliced divisions, must be smaller than 100000")
 
-    assert not os.path.exists(args.data_cache_dir), f"Find existed data cache dir {args.data_cache_dir}"
-    assert not os.path.exists(
-        args.data_jsonl_output_dir
-    ), f"Find existed jsonl data output dir {args.data_jsonl_output_dir}"
-    assert not os.path.exists(
-        args.data_arrow_output_dir
-    ), f"Find existed arrow data output dir {args.data_arrow_output_dir}"
-    os.makedirs(args.data_jsonl_output_dir)
-    os.makedirs(args.data_arrow_output_dir)
+    args.data_cache_dir = os.path.join(args.data_output_dirs, "cache")
+    args.data_jsonl_output_dir = os.path.join(args.data_output_dirs, "jsonl")
+    args.data_arrow_output_dir = os.path.join(args.data_output_dirs, "arrow")
+
+    if not os.path.exists(args.data_cache_dir):
+        os.makedirs(args.data_cache_dir)
+    if not os.path.exists(args.data_jsonl_output_dir):
+        os.makedirs(args.data_jsonl_output_dir)
+    if not os.path.exists(args.data_arrow_output_dir):
+        os.makedirs(args.data_arrow_output_dir)
 
     # Prepare to all input datasets
     input_data_paths = []
@@ -86,7 +75,7 @@ def main():
         train_splits.append(f"train[{start}%:{end}%]")
 
     # Prepare to the tokenizer.
-    tokenizer = LlamaTokenizer.from_pretrained(args.tokenizer_dir)
+    tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_dir)
     tokenizer.add_bos_token = False
     tokenizer.add_eos_token = False
     if tokenizer.pad_token is None:

applications/Colossal-LLaMA/prepare_sft_dataset.py

@@ -10,10 +10,10 @@ import math
 import os
 from multiprocessing import cpu_count
 
-from colossal_llama2.dataset.conversation import default_conversation
-from colossal_llama2.dataset.spliced_and_tokenized_dataset import supervised_tokenize_sft
+from colossal_llama.dataset.conversation import default_conversation
+from colossal_llama.dataset.spliced_and_tokenized_dataset import supervised_tokenize_sft
 from datasets import dataset_dict, load_dataset
-from transformers.models.llama.tokenization_llama import LlamaTokenizer
+from transformers import AddedToken, AutoTokenizer
 
 from colossalai.logging import get_dist_logger
 
@@ -32,35 +32,25 @@ def main():
     parser.add_argument(
         "--tokenizer_dir", type=str, required=True, default=None, help="A directory containing the tokenizer"
     )
-    parser.add_argument("--data_cache_dir", type=str, default="cache", help="Data cache directory")
-    parser.add_argument(
-        "--data_jsonl_output_dir",
-        type=str,
-        default="jsonl_output",
-        help="Output directory of spliced dataset with jsonl format",
-    )
-    parser.add_argument(
-        "--data_arrow_output_dir",
-        type=str,
-        default="arrow_output",
-        help="Output directory of spliced dataset with arrow format",
-    )
-    parser.add_argument("--max_length", type=int, default=4096, help="Max length of each spliced tokenized sequence")
+    parser.add_argument("--data_output_dirs", type=str, default="data_output_dirs", help="Data output directory")
+    parser.add_argument("--max_length", type=int, default=8192, help="Max length of each spliced tokenized sequence")
     parser.add_argument("--num_spliced_dataset_bins", type=int, default=10, help="Number of spliced dataset bins")
+    parser.add_argument("--llama_version", type=int, default=3, help="LLaMA version")
     args = parser.parse_args()
 
     if args.num_spliced_dataset_bins >= 100000:
         raise ValueError("Too many spliced divisions, must be smaller than 100000")
 
-    assert not os.path.exists(args.data_cache_dir), f"Find existed data cache dir {args.data_cache_dir}"
-    assert not os.path.exists(
-        args.data_jsonl_output_dir
-    ), f"Find existed jsonl data output dir {args.data_jsonl_output_dir}"
-    assert not os.path.exists(
-        args.data_arrow_output_dir
-    ), f"Find existed arrow data output dir {args.data_arrow_output_dir}"
-    os.makedirs(args.data_jsonl_output_dir)
-    os.makedirs(args.data_arrow_output_dir)
+    args.data_cache_dir = os.path.join(args.data_output_dirs, "cache")
+    args.data_jsonl_output_dir = os.path.join(args.data_output_dirs, "jsonl")
+    args.data_arrow_output_dir = os.path.join(args.data_output_dirs, "arrow")
+
+    if not os.path.exists(args.data_cache_dir):
+        os.makedirs(args.data_cache_dir)
+    if not os.path.exists(args.data_jsonl_output_dir):
+        os.makedirs(args.data_jsonl_output_dir)
+    if not os.path.exists(args.data_arrow_output_dir):
+        os.makedirs(args.data_arrow_output_dir)
 
     # Prepare to all input datasets
     input_data_paths = []
@@ -83,11 +73,20 @@ def main():
         train_splits.append(f"train[{start}%:{end}%]")
 
     # Prepare to the tokenizer.
-    tokenizer = LlamaTokenizer.from_pretrained(args.tokenizer_dir)
+    tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_dir)
+
+    # Fix </s> split issue: https://github.com/huggingface/transformers/issues/23833
+    if args.llama_version == 2:
+        tokenizer.add_tokens(AddedToken("</s>", normalized=False, special=True), special_tokens=True)
+
     tokenizer.add_bos_token = False
     tokenizer.add_eos_token = False
     if tokenizer.pad_token is None:
-        tokenizer.pad_token = tokenizer.unk_token
+        if tokenizer.unk_token is not None:
+            tokenizer.pad_token = tokenizer.unk_token
+        else:
+            tokenizer.pad_token = tokenizer.eos_token
+            tokenizer.unk_token = tokenizer.eos_token
 
     list_dataset = load_dataset(
         path="json",

applications/Colossal-LLaMA/requirements.txt

@@ -1,9 +1,10 @@
-torch<2.0.0, >=1.12.1
-packaging==23.1
-colossalai==0.3.5
+torch==2.1.2
+huggingface-hub
+packaging==24.0
+colossalai==0.3.6
 autoflake==2.2.1
 black==23.9.1
-transformers==4.33.3
+transformers==4.34.1
 tensorboard==2.14.0
 six==1.16.0
 datasets

applications/Colossal-LLaMA/stream_chat_example.py

@@ -1,6 +1,6 @@
 import argparse
 
-from colossal_llama2.utils.stream_chat_patch import streaming_chat
+from colossal_llama.utils.stream_chat_patch import streaming_chat
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
 SYSTEM = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions."

applications/Colossal-LLaMA/train.py

@@ -12,18 +12,18 @@ from contextlib import nullcontext
 
 import torch
 import torch.distributed as dist
-from colossal_llama2.dataset.loader import (
+from colossal_llama.dataset.loader import (
     DataCollatorForSupervisedDataset,
     StatefulDistributedSampler,
     load_tokenized_dataset,
 )
-from colossal_llama2.utils.ckpt_io import load_checkpoint, save_checkpoint
-from colossal_llama2.utils.flash_attention_patch import replace_with_flash_attention
-from colossal_llama2.utils.froze import freeze_non_embeds_parameters
-from colossal_llama2.utils.neftune_patch import activate_neftune, deactivate_neftune
+from colossal_llama.utils.ckpt_io import load_checkpoint, save_checkpoint
+from colossal_llama.utils.flash_attention_patch import replace_with_flash_attention
+from colossal_llama.utils.froze import freeze_non_embeds_parameters
+from colossal_llama.utils.neftune_patch import activate_neftune, deactivate_neftune
 from torch.utils.tensorboard import SummaryWriter
 from tqdm import tqdm
-from transformers import LlamaForCausalLM, LlamaTokenizer
+from transformers import AutoTokenizer, LlamaForCausalLM
 
 import colossalai
 from colossalai.accelerator import get_accelerator
@@ -89,7 +89,7 @@ def main() -> None:
     parser.add_argument("--accumulation_steps", type=int, default=1, help="Number of accumulation steps")
     parser.add_argument("--micro_batch_size", type=int, default=2, help="Batch size of each process")
     parser.add_argument("--lr", type=float, default=3e-4, help="Learning rate")
-    parser.add_argument("--max_length", type=int, default=4096, help="Model max length")
+    parser.add_argument("--max_length", type=int, default=8192, help="Model max length")
     parser.add_argument(
         "--mixed_precision",
         type=str,
@@ -196,7 +196,7 @@ def main() -> None:
     # ======================================================
     # Initialize Tokenizer, Dataset, Collator and Dataloader
     # ======================================================
-    tokenizer = LlamaTokenizer.from_pretrained(args.pretrained)
+    tokenizer = AutoTokenizer.from_pretrained(args.pretrained)
     if args.pad_token == "eos":
         tokenizer.pad_token = tokenizer.eos_token
     elif args.pad_token == "unk":

applications/Colossal-LLaMA/version.txt

@@ -0,0 +1 @@
+1.0.0

applications/README.md

@@ -5,7 +5,7 @@ This directory contains the applications that are powered by Colossal-AI.
 The list of applications include:
 
 - [X] [Open-Sora](https://github.com/hpcaitech/Open-Sora): Revealing Complete Model Parameters, Training Details, and Everything for Sora-like Video Generation Models
-- [X] [Colossal-LLaMA-2](./Colossal-LLaMA-2/): Continual Pre-training of LLaMA-2.
+- [X] [Colossal-LLaMA](./Colossal-LLaMA/): Continual Pre-training and Supervisied Fine-tuning of LLaMA2 / LLaMA3.
 - [X] [ColossalEval](./ColossalEval): Evaluation Pipeline for LLMs.
 - [X] [ColossalChat](./Chat/README.md): Replication of ChatGPT with RLHF.
 - [X] [FastFold](https://github.com/hpcaitech/FastFold): Optimizing AlphaFold (Biomedicine) Training and Inference on GPU Clusters.