[Pipeline Inference] Merge pp with tp (#4993)

* refactor pipeline into new CaiInferEngine

* updata llama modeling forward

* merge tp with pp

* update docstring

* optimize test workflow and example

* fix typo

* add assert and todo

colossalai/inference/__init__.py

@@ -1,4 +1,4 @@
-from .pipeline import PPInferEngine
+from .hybridengine import CaiInferEngine
+from .hybridengine.polices import LlamaModelInferPolicy
 
-
+__all__ = ["CaiInferEngine", "LlamaModelInferPolicy"]
-__all__ = ['PPInferEngine']

colossalai/inference/hybridengine/__init__.py

@@ -0,0 +1,3 @@
+from .engine import CaiInferEngine
+
+__all__ = ["CaiInferEngine"]

colossalai/inference/hybridengine/engine.py

@@ -1,4 +1,5 @@
 import torch
+import torch.distributed as dist
 import torch.nn as nn
 from transformers.tokenization_utils_base import BatchEncoding
 
@@ -8,23 +9,27 @@ from colossalai.pipeline.stage_manager import PipelineStageManager
 from colossalai.shardformer import ShardConfig, ShardFormer
 from colossalai.shardformer.policies.base_policy import Policy
 
+from ..pipeline.microbatch_manager import MicroBatchManager
 from ..tensor_parallel.kvcache_manager import MemoryManager
-from .microbatch_manager import MicroBatchManager
+
+PP_AXIS, TP_AXIS = 0, 1
+
+_supported_models = [
+    "LlamaForCausalLM",
+]
 
 
-class PPInferEngine:
+class CaiInferEngine:
     """
-    PPInferEngine is a class that handles the pipeline parallel inference.
+    CaiInferEngine is a class that handles the pipeline parallel inference.
 
     Args:
-        pp_size (int): the number of pipeline stages.
+        tp_size (int): the size of tensor parallelism.
-        pp_model (`nn.Module`): the model already in pipeline parallelism style.
+        pp_size (int): the size of pipeline parallelism.
         model (`nn.Module`): the model not in pipeline style, and will be modified with `ShardFormer`.
         model_policy (`colossalai.shardformer.policies.base_policy.Policy`): the policy to shardformer model.
         micro_batch_size (int): the micro batch size.
         micro_batch_buffer_size (int): the buffer size for micro batch. Normally, it should be the same as the number of pipeline stages.
-        new_length (int): the new length of the input sequence.
-        early_stopping (bool): whether to stop early.
         max_batch_size (int): the maximum batch size.
         max_input_len (int): the maximum input length.
         max_output_len (int): the maximum output length.
@@ -32,7 +37,7 @@ class PPInferEngine:
     Example:
 
     ```python
-    from colossalai.inference import PPInferEngine
+    from colossalai.inference import InferEngine
     from colossalai.inference.pipeline.policies import LlamaModelInferPolicy
     import colossalai
     from transformers import LlamaForCausalLM, LlamaTokenizer
@@ -42,7 +47,7 @@ class PPInferEngine:
     model = LlamaForCausalLM.from_pretrained("your_path_to_model")
     tokenizer = LlamaTokenizer.from_pretrained("/home/lczyh/share/models/llama-7b-hf")
     # assume the model is infered with 2 pipeline stages
-    inferengine = PPInferEngine(pp_size=2, model=model, model_policy=LlamaModelInferPolicy(), new_length=8)
+    inferengine = CaiInferEngine(pp_size=2, model=model, model_policy=LlamaModelInferPolicy())
 
     input = ["Introduce a landmark in China ","Introduce a landmark in China "]
     data = tokenizer(input, return_tensors='pt')
@@ -54,12 +59,11 @@ class PPInferEngine:
 
     def __init__(
         self,
-        pp_size: int,
+        tp_size: int = 1,
+        pp_size: int = 1,
         dtype: str = "fp16",
-        pp_model: nn.Module = None,
         model: nn.Module = None,
         model_policy: Policy = None,
-        new_length: int = 32,
         micro_batch_size: int = 1,
         micro_batch_buffer_size: int = None,
         max_batch_size: int = 4,
@@ -71,12 +75,21 @@ class PPInferEngine:
         do_sample: bool = False,
         num_beams: int = 1,
     ) -> None:
-        assert pp_model or (model and model_policy), "Either pp_model or model with model_policy should be provided."
+        assert model.__class__.__name__ in _supported_models, f"Model {model.__class__.__name__} is not supported."
+        assert (
+            tp_size * pp_size == dist.get_world_size()
+        ), f"TP size({tp_size}) * PP size({pp_size}) should be equal to the global world size ({dist.get_world_size()})"
+        assert model and model_policy, "Model with model_policy should be provided."
         assert dtype in ["fp16", "fp32", "bf16"], "dtype should be one of 'fp16', 'fp32', 'bf16'"
 
-        max_output_len = max(max_output_len, max_input_len + new_length)
+        assert max_batch_size <= 64, "Max batch size exceeds the constraint"
+        assert max_input_len + max_output_len <= 4096, "Max length exceeds the constraint"
 
+        # TODO: support only tensor parallel inference
+        assert pp_size > 1, "Not support only tensor parallel inference."
         self.pp_size = pp_size
+        self.tp_size = tp_size
+
         if dtype == "fp16":
             self.dtype = torch.float16
             model.half()
@@ -85,24 +98,29 @@ class PPInferEngine:
             model.to(torch.bfloat16)
         else:
             self.dtype = torch.float32
-        self.pg_mesh = ProcessGroupMesh(pp_size)
+
-        self.stage_manager = PipelineStageManager(self.pg_mesh, 0, True)
+        # Init pg mesh
-        self.model = pp_model or self._shardformer(model, model_policy)
+        pg_mesh = ProcessGroupMesh(pp_size, tp_size)
-        self.cache_manager_list = [
+
-            self._init_manager(max_batch_size, max_input_len, max_output_len)
+        stage_manager = None
-            for _ in range(micro_batch_buffer_size or pp_size)
+        if pp_size > 1:
-        ]
+            stage_manager = PipelineStageManager(pg_mesh, PP_AXIS, True)
-        self.mb_manager = MicroBatchManager(
+            self.cache_manager_list = [
-            self.stage_manager.stage,
+                self._init_manager(model, max_batch_size, max_input_len, max_output_len)
-            new_length,
+                for _ in range(micro_batch_buffer_size or pp_size)
-            micro_batch_size,
+            ]
-            micro_batch_buffer_size or pp_size,
+            self.mb_manager = MicroBatchManager(
-            max_input_len,
+                stage_manager.stage,
-            max_output_len,
+                micro_batch_size,
-            self.cache_manager_list,
+                micro_batch_buffer_size or pp_size,
-        )
+                max_input_len,
-        self.verbose = verbose
+                max_output_len,
-        self.schedule = GenerateSchedule(self.stage_manager, self.mb_manager, verbose)
+                self.cache_manager_list,
+            )
+            self.verbose = verbose
+            self.schedule = GenerateSchedule(stage_manager, self.mb_manager, verbose)
+
+        self.model = self._shardformer(model, model_policy, stage_manager, pg_mesh.get_group_along_axis(TP_AXIS))
 
     def inference(self, input_list):
         """
@@ -124,10 +142,10 @@ class PPInferEngine:
         else:
             return out
 
-    def _shardformer(self, model, model_policy):
+    def _shardformer(self, model, model_policy, stage_manager, tp_group):
         shardconfig = ShardConfig(
-            tensor_parallel_process_group=None,
+            tensor_parallel_process_group=tp_group,
-            pipeline_stage_manager=self.stage_manager,
+            pipeline_stage_manager=stage_manager,
             enable_tensor_parallelism=False,
             enable_fused_normalization=False,
             enable_all_optimization=False,
@@ -139,14 +157,12 @@ class PPInferEngine:
         shard_model, _ = shardformer.optimize(model, model_policy)
         return shard_model.cuda()
 
-    def _init_manager(self, max_batch_size: int, max_input_len: int, max_output_len: int) -> None:
+    def _init_manager(self, model, max_batch_size: int, max_input_len: int, max_output_len: int) -> None:
         max_total_token_num = max_batch_size * (max_input_len + max_output_len)
-        head_dim = self.model.config.hidden_size // self.model.config.num_attention_heads
+        head_dim = model.config.hidden_size // model.config.num_attention_heads
-        head_num = self.model.config.num_attention_heads
+        head_num = model.config.num_attention_heads
         num_hidden_layers = (
-            self.model.config.num_hidden_layers
+            model.config.num_hidden_layers if hasattr(model.config, "num_hidden_layers") else model.config.num_layers
-            if hasattr(self.model.config, "num_hidden_layers")
-            else self.model.config.num_layers
         )
         layer_num = num_hidden_layers // self.pp_size
 

colossalai/inference/hybridengine/modeling/llama.py

@@ -1,37 +1,25 @@
 # This code is adapted from huggingface transformers: https://github.com/huggingface/transformers/blob/v4.34.1/src/transformers/models/llama/modeling_llama.py
+import math
 from typing import List, Optional, Tuple
 
 import torch
-from transformers.models.llama.modeling_llama import (
+from transformers.models.llama.modeling_llama import LlamaAttention, LlamaDecoderLayer, LlamaForCausalLM, LlamaModel
-    LlamaAttention,
-    LlamaDecoderLayer,
-    LlamaForCausalLM,
-    LlamaModel,
-    LlamaRMSNorm,
-)
 from transformers.utils import logging
 
 from colossalai.inference.tensor_parallel.batch_infer_state import BatchInferState
 from colossalai.kernel.triton import llama_context_attn_fwd, token_attention_fwd
+from colossalai.kernel.triton.token_attention_kernel import Llama2TokenAttentionForwards
 from colossalai.pipeline.stage_manager import PipelineStageManager
 
 from ._utils import copy_kv_to_mem_cache
 
 try:
-    from vllm import layernorm_ops, pos_encoding_ops
+    from lightllm.models.llama2.triton_kernel.context_flashattention_nopad import (
-
+        context_attention_fwd as lightllm_llama2_context_attention_fwd,
-    rms_norm = layernorm_ops.rms_norm
+    )
-    rotary_embedding_neox = pos_encoding_ops.rotary_embedding_neox
+    from lightllm.models.llama.triton_kernel.context_flashattention_nopad import (
-    HAS_VLLM_KERNERL = True
+        context_attention_fwd as lightllm_context_attention_fwd,
-except:
-    print("fall back to original rotary_embedding_neox of huggingface")
-    print("install vllm from https://github.com/vllm-project/vllm to accelerate your inference")
-    print(
-        "if falied to install vllm, please use this branch to install: https://github.com/tiandiao123/vllm/tree/setup_branch"
     )
-    HAS_VLLM_KERNERL = False
-
-try:
     from lightllm.models.llama.triton_kernel.rotary_emb import rotary_emb_fwd as llama_rotary_embedding_fwd
 
     HAS_LIGHTLLM_KERNEL = True
@@ -39,6 +27,14 @@ except:
     print("please install lightllm from source to run inference: https://github.com/ModelTC/lightllm")
     HAS_LIGHTLLM_KERNEL = False
 
+try:
+    from flash_attn import flash_attn_with_kvcache
+
+    HAS_FLASH_KERNEL = True
+except:
+    HAS_FLASH_KERNEL = False
+    print("please install flash attentiom from https://github.com/Dao-AILab/flash-attention")
+
 
 def rotate_half(x):
     """Rotates half the hidden dims of the input."""
@@ -59,6 +55,75 @@ def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
     return q_embed, k_embed
 
 
+def llama_triton_context_attention(
+    query_states, key_states, value_states, attn_output, infer_state, num_key_value_groups=1
+):
+    if num_key_value_groups == 1:
+        if HAS_LIGHTLLM_KERNEL is False:
+            llama_context_attn_fwd(
+                query_states,
+                key_states,
+                value_states,
+                attn_output,
+                infer_state.start_loc,
+                infer_state.seq_len,
+                # infer_state.cache_manager.past_key_values_length,
+                infer_state.max_len_in_batch,
+            )
+        else:
+            lightllm_context_attention_fwd(
+                query_states,
+                key_states,
+                value_states,
+                attn_output,
+                infer_state.start_loc,
+                infer_state.seq_len,
+                # infer_state.cache_manager.past_key_values_length,
+                infer_state.max_len_in_batch,
+            )
+    else:
+        assert HAS_LIGHTLLM_KERNEL is True, "You have to install lightllm kernels to run llama2 model"
+        lightllm_llama2_context_attention_fwd(
+            query_states,
+            key_states,
+            value_states,
+            attn_output,
+            infer_state.start_loc,
+            infer_state.seq_len,
+            # infer_state.cache_manager.past_key_values_length,
+            infer_state.max_len_in_batch,
+        )
+
+
+def llama_triton_token_attention(query_states, attn_output, infer_state, num_key_value_groups=1):
+    assert HAS_LIGHTLLM_KERNEL is True, "You have to install lightllm kernel to run token attention for llama models"
+    if num_key_value_groups == 1:
+        token_attention_fwd(
+            query_states,
+            infer_state.cache_manager.key_buffer[infer_state.decode_layer_id],
+            infer_state.cache_manager.value_buffer[infer_state.decode_layer_id],
+            attn_output,
+            infer_state.block_loc,
+            infer_state.start_loc,
+            infer_state.seq_len,
+            # infer_state.cache_manager.past_key_values_length,
+            infer_state.max_len_in_batch,
+        )
+    else:
+        Llama2TokenAttentionForwards.token_attn(
+            query_states,
+            infer_state.cache_manager.key_buffer[infer_state.decode_layer_id],
+            infer_state.cache_manager.value_buffer[infer_state.decode_layer_id],
+            attn_output,
+            infer_state.block_loc,
+            infer_state.start_loc,
+            infer_state.seq_len,
+            # infer_state.cache_manager.past_key_values_length,
+            infer_state.max_len_in_batch,
+            infer_state.other_kv_index,
+        )
+
+
 class LlamaInferenceForwards:
     """
     This class holds forwards for llama inference.
@@ -144,13 +209,9 @@ class LlamaInferenceForwards:
         hidden_states: Optional[torch.FloatTensor] = None,
         stage_index: Optional[List[int]] = None,
     ):
-        # batch_size = input_ids.shape[0]  # input_ids.shape[0]
-        # print(f"[Before] rank:{torch.distributed.get_rank()}\n->{infer_state}")
-
-        # infer_state = self.infer_state
-        use_cache = use_cache if use_cache is not None else self.config.use_cache
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
         # retrieve input_ids and inputs_embeds
         if stage_manager is None or stage_manager.is_first_stage():
             if input_ids is not None and inputs_embeds is not None:
@@ -172,12 +233,10 @@ class LlamaInferenceForwards:
             batch_size, seq_length = input_shape
             device = hidden_states.device
 
-        seq_length_with_past = seq_length
+        if infer_state.is_context_stage:
-        past_key_values_length = 0
+            past_key_values_length = 0
-
+        else:
-        if infer_state.is_context_stage is False:
+            past_key_values_length = infer_state.max_len_in_batch - 1
-            past_key_values_length = infer_state.cache_manager.past_key_values_length
-            seq_length_with_past = seq_length_with_past + past_key_values_length
 
         # NOTE: differentiate with prefill stage
         #       block_loc require different value-assigning method for two different stage
@@ -197,26 +256,19 @@ class LlamaInferenceForwards:
                 infer_state.decode_mem_index = alloc_mem[0]
                 infer_state.decode_mem_start = alloc_mem[1]
                 infer_state.decode_mem_end = alloc_mem[2]
-                infer_state.block_loc[:, seq_length_with_past - 1] = infer_state.decode_mem_index
+                infer_state.block_loc[:, infer_state.max_len_in_batch - 1] = infer_state.decode_mem_index
             else:
-                print(f" *** Encountered allocation non-contiguous")
-                print(
-                    f"    infer_state.cache_manager.past_key_values_length: {infer_state.cache_manager.past_key_values_length}"
-                )
                 infer_state.decode_is_contiguous = False
                 alloc_mem = infer_state.cache_manager.alloc(batch_size)
                 infer_state.decode_mem_index = alloc_mem
-                # infer_state.decode_key_buffer = torch.empty((batch_size, self.tp_head_num_, self.head_dim_), dtype=torch.float16, device="cuda")
+                infer_state.block_loc[:, infer_state.max_len_in_batch - 1] = infer_state.decode_mem_index
-                # infer_state.decode_value_buffer = torch.empty((batch_size, self.tp_head_num_, self.head_dim_), dtype=torch.float16, device="cuda")
+
-                infer_state.block_loc[:, seq_length_with_past - 1] = infer_state.decode_mem_index
         if position_ids is None:
             position_ids = torch.arange(
                 past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
             )
-            position_ids = position_ids.unsqueeze(0)
+            position_ids = position_ids.repeat(batch_size, 1)
-            new_shape = [1] * position_ids.dim()
+            position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
-            new_shape[0] = batch_size
-            position_ids = position_ids.repeat(*new_shape).view(-1, seq_length)
         else:
             position_ids = position_ids.view(-1, seq_length).long()
 
@@ -227,15 +279,17 @@ class LlamaInferenceForwards:
             infer_state.position_sin = torch.index_select(self._sin_cached, 0, position_ids.view(-1)).view(
                 position_ids.view(-1).shape[0], -1
             )
+
         else:
             seq_len = infer_state.seq_len
             infer_state.position_cos = torch.index_select(self._cos_cached, 0, seq_len - 1).view(seq_len.shape[0], -1)
             infer_state.position_sin = torch.index_select(self._sin_cached, 0, seq_len - 1).view(seq_len.shape[0], -1)
+            infer_state.other_kv_index = infer_state.block_loc[0, infer_state.max_len_in_batch - 1].item()
 
         # embed positions
         if attention_mask is None:
             attention_mask = torch.ones(
-                (batch_size, seq_length_with_past), dtype=torch.bool, device=hidden_states.device
+                (batch_size, infer_state.max_len_in_batch), dtype=torch.bool, device=hidden_states.device
             )
 
         attention_mask = self._prepare_decoder_attention_mask(
@@ -243,10 +297,6 @@ class LlamaInferenceForwards:
         )
 
         # decoder layers
-        () if output_hidden_states else None
-        () if output_attentions else None
-        next_decoder_cache = () if use_cache else None
-
         infer_state.decode_layer_id = 0
 
         start_idx, end_idx = stage_index[0], stage_index[1]
@@ -268,19 +318,15 @@ class LlamaInferenceForwards:
             infer_state.decode_layer_id += 1
             hidden_states = layer_outputs[0]
 
-            if use_cache:
+        if stage_manager.is_last_stage() or stage_manager.num_stages == 1:
-                next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
-
-        if stage_manager.is_last_stage():
             hidden_states = self.norm(hidden_states)
-        next_cache = next_decoder_cache if use_cache else None
 
         # update indices
         # infer_state.block_loc[:, infer_state.max_len_in_batch-1] = infer_state.total_token_num + torch.arange(0, batch_size, dtype=torch.int32, device="cuda")
-        infer_state.start_loc = infer_state.start_loc + torch.arange(0, batch_size, dtype=torch.int32, device="cuda")
+        infer_state.start_loc += torch.arange(0, batch_size, dtype=torch.int32, device="cuda")
         infer_state.seq_len += 1
+        infer_state.max_len_in_batch += 1
 
-        # TODO: fix this to necessary return
         # if not return_dict:
         #     return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
 
@@ -290,8 +336,7 @@ class LlamaInferenceForwards:
         #     hidden_states=all_hidden_states,
         #     attentions=all_self_attns,
         # )
-        # print(f"[After] rank:{torch.distributed.get_rank()}\n->{infer_state}")
+        return {"hidden_states": hidden_states}
-        return {"hidden_states": hidden_states, "past_key_values": next_cache}
 
     @staticmethod
     def llama_decoder_layer_forward(
@@ -307,7 +352,6 @@ class LlamaInferenceForwards:
         residual = hidden_states
 
         hidden_states = self.input_layernorm(hidden_states)
-
         # Self Attention
         hidden_states, self_attn_weights, present_key_value = self.self_attn(
             hidden_states=hidden_states,
@@ -357,28 +401,24 @@ class LlamaInferenceForwards:
         # key_states_transposed [bs, num_heads, seq_len, head_dim/embed_size_per_head]
 
         query_states = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim)
-        key_states = self.k_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim)
+        key_states = self.k_proj(hidden_states).view(bsz, q_len, self.num_key_value_heads, self.head_dim)
-        value_states = self.v_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim)
+        value_states = self.v_proj(hidden_states).view(bsz, q_len, self.num_key_value_heads, self.head_dim)
 
         # NOTE might want to revise
         #   need some way to record the length of past key values cache
         #   since we won't return past_key_value_cache right now
-        if infer_state.decode_layer_id == 0:  # once per model.forward
-            infer_state.cache_manager.past_key_values_length += q_len  # seq_len
 
         cos, sin = infer_state.position_cos, infer_state.position_sin
 
         llama_rotary_embedding_fwd(query_states.view(-1, self.num_heads, self.head_dim), cos, sin)
-        llama_rotary_embedding_fwd(key_states.view(-1, self.num_heads, self.head_dim), cos, sin)
+        llama_rotary_embedding_fwd(key_states.view(-1, self.num_key_value_heads, self.head_dim), cos, sin)
 
         query_states = query_states.reshape(-1, self.num_heads, self.head_dim)
-        key_states = key_states.reshape(-1, self.num_heads, self.head_dim)
+        key_states = key_states.reshape(-1, self.num_key_value_heads, self.head_dim)
-        value_states = value_states.reshape(-1, self.num_heads, self.head_dim)
+        value_states = value_states.reshape(-1, self.num_key_value_heads, self.head_dim)
 
         if infer_state.is_context_stage:
-            # print(f"rank:{torch.distributed.get_rank()}, {infer_state}")
             # first token generation
-
             # copy key and value calculated in current step to memory manager
             copy_kv_to_mem_cache(
                 infer_state.decode_layer_id,
@@ -387,19 +427,16 @@ class LlamaInferenceForwards:
                 infer_state.context_mem_index,
                 infer_state.cache_manager,
             )
-
             attn_output = torch.empty_like(query_states)
 
-            llama_context_attn_fwd(
+            llama_triton_context_attention(
                 query_states,
                 key_states,
                 value_states,
                 attn_output,
-                infer_state.start_loc,
+                infer_state,
-                infer_state.seq_len,
+                num_key_value_groups=self.num_key_value_groups,
-                infer_state.cache_manager.past_key_values_length,
             )
-
         else:
             if infer_state.decode_is_contiguous:
                 # if decode is contiguous, then we copy to key cache and value cache in cache manager directly
@@ -422,45 +459,31 @@ class LlamaInferenceForwards:
                     infer_state.cache_manager,
                 )
 
-            # second token and follows
+            if HAS_LIGHTLLM_KERNEL:
-            # kv = torch.stack((key_states, value_states), dim=2)
+                attn_output = torch.empty_like(query_states)
-            # (batch_size, seqlen, nheads, headdim)
+                llama_triton_token_attention(
-            attn_output = torch.empty_like(query_states)
+                    query_states, attn_output, infer_state, num_key_value_groups=self.num_key_value_groups
+                )
+            else:
+                self.num_heads // self.num_key_value_heads
+                cache_k = infer_state.cache_manager.key_buffer[infer_state.decode_layer_id]
+                cache_v = infer_state.cache_manager.value_buffer[infer_state.decode_layer_id]
 
-            token_attention_fwd(
+                query_states = query_states.view(bsz, -1, self.num_heads, self.head_dim)
-                query_states,
+                copy_cache_k = cache_k.view(bsz, -1, self.num_key_value_heads, self.head_dim)
-                infer_state.cache_manager.key_buffer[infer_state.decode_layer_id],
+                copy_cache_v = cache_v.view(bsz, -1, self.num_key_value_heads, self.head_dim)
-                infer_state.cache_manager.value_buffer[infer_state.decode_layer_id],
+
-                attn_output,
+                attn_output = flash_attn_with_kvcache(
-                infer_state.block_loc,
+                    q=query_states,
-                infer_state.start_loc,
+                    k_cache=copy_cache_k,
-                infer_state.seq_len,
+                    v_cache=copy_cache_v,
-                infer_state.cache_manager.past_key_values_length,
+                    softmax_scale=1 / math.sqrt(self.head_dim),
-            )
+                    causal=True,
+                )
 
         attn_output = attn_output.view(bsz, q_len, self.hidden_size)
-        # print(f"rank:{torch.distributed.get_rank()}, {attn_output}")
+
         attn_output = self.o_proj(attn_output)
 
         # return past_key_value as None
         return attn_output, None, None
-
-
-def get_llama_vllm_rmsnorm_forward():
-    if HAS_VLLM_KERNERL:
-
-        def _vllm_rmsnorm_forward(self: LlamaRMSNorm, hidden_states: torch.Tensor):
-            x = hidden_states
-            out = torch.empty_like(x)
-            rms_norm(
-                out,
-                x,
-                self.weight.data,
-                self.variance_epsilon,
-            )
-
-            return out
-
-        return _vllm_rmsnorm_forward
-    else:
-        return None

colossalai/inference/hybridengine/polices/llama.py

@@ -17,7 +17,7 @@ from colossalai.shardformer.policies.base_policy import ModulePolicyDescription,
 from colossalai.shardformer.policies.llama import LlamaForCausalLMPolicy
 
 from ..modeling._utils import init_to_get_rotary
-from ..modeling.llama import LlamaInferenceForwards, get_llama_vllm_rmsnorm_forward
+from ..modeling.llama import LlamaInferenceForwards
 
 try:
     from colossalai.kernel.triton import rmsnorm_forward
@@ -120,9 +120,6 @@ class LlamaModelInferPolicy(LlamaForCausalLMPolicy):
         infer_forward = None
         if HAS_TRITON_RMSNORM:
             infer_forward = get_triton_rmsnorm_forward()
-        else:
-            # NOTE: adding rms_norm from cuda kernels caused precision issue, fix @tiandiao123
-            infer_forward = get_llama_vllm_rmsnorm_forward()
 
         if infer_forward is not None:
             method_replacement = {"forward": partial(infer_forward)}

colossalai/inference/pipeline/__init__.py

@@ -1,3 +1,3 @@
-from .engine import PPInferEngine
+from .microbatch_manager import MicroBatchManager
 
-__all__ = ["PPInferEngine"]
+__all__ = ["MicroBatchManager"]

colossalai/inference/pipeline/microbatch_manager.py

@@ -33,10 +33,9 @@ class MicroBatchDescription:
         max_input_len: int,
         max_output_len: int,
         cache_manager: MemoryManager,
-        new_length: int,
     ) -> None:
         self.mb_length = inputs_dict["input_ids"].shape[-1]
-        self.target_length = self.mb_length + new_length
+        self.target_length = self.mb_length + max_output_len
         self.infer_state = BatchInferState.init_from_batch(
             batch=inputs_dict, max_input_len=max_input_len, max_output_len=max_output_len, cache_manager=cache_manager
         )
@@ -77,7 +76,6 @@ class HeadMicroBatchDescription(MicroBatchDescription):
     Args:
         inputs_dict (Dict[str, torch.Tensor]): the inputs of current stage. The key should have `input_ids` and `attention_mask`.
         output_dict (Dict[str, torch.Tensor]): the outputs of previous stage. The key should have `hidden_states` and `past_key_values`.
-        new_length (int): the new length of the input sequence.
 
     """
 
@@ -87,9 +85,8 @@ class HeadMicroBatchDescription(MicroBatchDescription):
         max_input_len: int,
         max_output_len: int,
         cache_manager: MemoryManager,
-        new_length: int,
     ) -> None:
-        super().__init__(inputs_dict, max_input_len, max_output_len, cache_manager, new_length)
+        super().__init__(inputs_dict, max_input_len, max_output_len, cache_manager)
         assert inputs_dict is not None
         assert inputs_dict.get("input_ids") is not None and inputs_dict.get("attention_mask") is not None
         self.input_ids = inputs_dict["input_ids"]
@@ -139,9 +136,8 @@ class BodyMicroBatchDescription(MicroBatchDescription):
         max_input_len: int,
         max_output_len: int,
         cache_manager: MemoryManager,
-        new_length: int,
     ) -> None:
-        super().__init__(inputs_dict, max_input_len, max_output_len, cache_manager, new_length)
+        super().__init__(inputs_dict, max_input_len, max_output_len, cache_manager)
 
     @property
     def cur_length(self):
@@ -158,7 +154,6 @@ class MicroBatchManager:
 
     Args:
         stage (int): stage id of current stage.
-        new_length (int): the new length of the input sequence.
         micro_batch_size (int): the micro batch size.
         micro_batch_buffer_size (int): the buffer size for micro batch. Normally, it should be the same as the number of pipeline stages.
 
@@ -167,7 +162,6 @@ class MicroBatchManager:
     def __init__(
         self,
         stage: int,
-        new_length: int,
         micro_batch_size: int,
         micro_batch_buffer_size: int,
         max_input_len: int,
@@ -175,7 +169,6 @@ class MicroBatchManager:
         cache_manager_list: MemoryManager,
     ):
         self.stage = stage
-        self.new_length = new_length
         self.micro_batch_size = micro_batch_size
         self.buffer_size = micro_batch_buffer_size
         self.max_input_len = max_input_len
@@ -188,11 +181,11 @@ class MicroBatchManager:
     def add_descrption(self, inputs_dict: Dict[str, torch.Tensor]):
         if self.stage == 0:
             self.mb_descrption_buffer[self.idx] = HeadMicroBatchDescription(
-                inputs_dict, self.max_input_len, self.max_output_len, self.cache_manager_list[self.idx], self.new_length
+                inputs_dict, self.max_input_len, self.max_output_len, self.cache_manager_list[self.idx]
             )
         else:
             self.mb_descrption_buffer[self.idx] = BodyMicroBatchDescription(
-                inputs_dict, self.max_input_len, self.max_output_len, self.cache_manager_list[self.idx], self.new_length
+                inputs_dict, self.max_input_len, self.max_output_len, self.cache_manager_list[self.idx]
             )
 
     def step(self, new_token: torch.Tensor = None):

colossalai/inference/tensor_parallel/modeling/llama.py

@@ -1,10 +1,9 @@
-from typing import List, Optional, Tuple
 import math
-import copy
+from typing import List, Optional, Tuple
 
 import torch
 from transformers.modeling_outputs import BaseModelOutputWithPast
-from transformers.models.llama.modeling_llama import LlamaAttention, LlamaDecoderLayer, LlamaModel, LlamaRMSNorm
+from transformers.models.llama.modeling_llama import LlamaAttention, LlamaDecoderLayer, LlamaModel
 
 from colossalai.inference.tensor_parallel.batch_infer_state import BatchInferState
 from colossalai.kernel.triton import llama_context_attn_fwd, token_attention_fwd
@@ -16,7 +15,9 @@ try:
     from lightllm.models.llama2.triton_kernel.context_flashattention_nopad import (
         context_attention_fwd as lightllm_llama2_context_attention_fwd,
     )
-    from lightllm.models.llama.triton_kernel.context_flashattention_nopad import context_attention_fwd as lightllm_context_attention_fwd
+    from lightllm.models.llama.triton_kernel.context_flashattention_nopad import (
+        context_attention_fwd as lightllm_context_attention_fwd,
+    )
     from lightllm.models.llama.triton_kernel.rotary_emb import rotary_emb_fwd as llama_rotary_embedding_fwd
 
     HAS_LIGHTLLM_KERNEL = True
@@ -26,6 +27,7 @@ except:
 
 try:
     from flash_attn import flash_attn_with_kvcache
+
     HAS_FLASH_KERNEL = True
 except:
     HAS_FLASH_KERNEL = False
@@ -50,7 +52,10 @@ def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
     k_embed = (k * cos) + (rotate_half(k) * sin)
     return q_embed, k_embed
 
-def llama_triton_context_attention(query_states, key_states, value_states, attn_output, infer_state, num_key_value_groups=1):
+
+def llama_triton_context_attention(
+    query_states, key_states, value_states, attn_output, infer_state, num_key_value_groups=1
+):
     if num_key_value_groups == 1:
         if HAS_LIGHTLLM_KERNEL is False:
             llama_context_attn_fwd(
@@ -87,6 +92,7 @@ def llama_triton_context_attention(query_states, key_states, value_states, attn_
             infer_state.max_len_in_batch,
         )
 
+
 def llama_triton_token_attention(query_states, attn_output, infer_state, num_key_value_groups=1):
     assert HAS_LIGHTLLM_KERNEL is True, "You have to install lightllm kernel to run token attention for llama models"
     if num_key_value_groups == 1:
@@ -265,8 +271,7 @@ class LlamaInferenceForwards:
             hidden_states=all_hidden_states,
             attentions=all_self_attns,
         )
-        
+
-    
     @staticmethod
     def llama_decoder_layer_forward(
         self: LlamaDecoderLayer,
@@ -309,7 +314,6 @@ class LlamaInferenceForwards:
             outputs += (present_key_value,)
 
         return outputs
-    
 
     @staticmethod
     def llama_flash_attn_kvcache_forward(
@@ -358,8 +362,15 @@ class LlamaInferenceForwards:
                 infer_state.cache_manager,
             )
             attn_output = torch.empty_like(query_states)
-            
+
-            llama_triton_context_attention(query_states, key_states, value_states, attn_output, infer_state, num_key_value_groups=self.num_key_value_groups)
+            llama_triton_context_attention(
+                query_states,
+                key_states,
+                value_states,
+                attn_output,
+                infer_state,
+                num_key_value_groups=self.num_key_value_groups,
+            )
         else:
             if infer_state.decode_is_contiguous:
                 # if decode is contiguous, then we copy to key cache and value cache in cache manager directly
@@ -381,26 +392,28 @@ class LlamaInferenceForwards:
                     infer_state.decode_mem_index,
                     infer_state.cache_manager,
                 )
-            
+
-            HAS_LIGHTLLM_KERNEL = False
             if HAS_LIGHTLLM_KERNEL:
                 attn_output = torch.empty_like(query_states)
-                llama_triton_token_attention(query_states, attn_output, infer_state, num_key_value_groups=self.num_key_value_groups)
+                llama_triton_token_attention(
+                    query_states, attn_output, infer_state, num_key_value_groups=self.num_key_value_groups
+                )
             else:
-                heads_per_group = self.num_heads // self.num_key_value_heads
+                self.num_heads // self.num_key_value_heads
                 cache_k = infer_state.cache_manager.key_buffer[infer_state.decode_layer_id]
                 cache_v = infer_state.cache_manager.value_buffer[infer_state.decode_layer_id]
-                
+
                 query_states = query_states.view(bsz, -1, self.num_heads, self.head_dim)
                 copy_cache_k = cache_k.view(bsz, -1, self.num_key_value_heads, self.head_dim)
                 copy_cache_v = cache_v.view(bsz, -1, self.num_key_value_heads, self.head_dim)
-                
-                attn_output = flash_attn_with_kvcache(q = query_states, 
-                                                      k_cache = copy_cache_k, 
-                                                      v_cache = copy_cache_v, 
-                                                      softmax_scale = 1/ math.sqrt(self.head_dim), 
-                                                      causal = True)       
 
+                attn_output = flash_attn_with_kvcache(
+                    q=query_states,
+                    k_cache=copy_cache_k,
+                    v_cache=copy_cache_v,
+                    softmax_scale=1 / math.sqrt(self.head_dim),
+                    causal=True,
+                )
 
         attn_output = attn_output.view(bsz, q_len, self.hidden_size)
 
@@ -408,4 +421,3 @@ class LlamaInferenceForwards:
 
         # return past_key_value as None
         return attn_output, None, None
-

tests/test_infer/test_pipeline_infer.py

@@ -5,8 +5,7 @@ import transformers
 from packaging import version
 
 import colossalai
-from colossalai.inference.pipeline import PPInferEngine
+from colossalai.inference import CaiInferEngine, LlamaModelInferPolicy
-from colossalai.inference.pipeline.policies import LlamaModelInferPolicy
 from colossalai.testing import clear_cache_before_run, parameterize, rerun_if_address_is_in_use, spawn
 
 CUDA_SUPPORT = version.parse(torch.version.cuda) > version.parse("11.5")
@@ -26,27 +25,43 @@ for k, v in inputs.items():
         inputs[k] = v.to("cuda").repeat(*new_shape)
 
 
-def pipeline_inference_test(pp_size, new_length, micro_batch_size):
+def pipeline_inference_test(tp_size, pp_size, max_output_len, micro_batch_size):
-    model = transformers.LlamaForCausalLM(transformers.LlamaConfig(num_hidden_layers=4))
+    model = transformers.LlamaForCausalLM(
+        transformers.LlamaConfig(
+            vocab_size=20000, hidden_size=512, intermediate_size=1536, num_attention_heads=4, num_hidden_layers=4
+        )
+    )
 
-    engine = PPInferEngine(
+    engine = CaiInferEngine(
+        tp_size=tp_size,
         pp_size=pp_size,
         model=model,
         model_policy=LlamaModelInferPolicy(),
-        new_length=new_length,
+        max_output_len=max_output_len,
         micro_batch_size=micro_batch_size,
     )
     output = engine.inference(inputs)
     if dist.get_rank() == 0:
-        assert len(output[0]) == new_length, f"{len(output)}, {new_length}"
+        assert len(output[0]) == max_output_len, f"{len(output)}, {max_output_len}"
 
 
+@parameterize("tp_size", [1])
 @parameterize("pp_size", [2])
-@parameterize("new_length", [4, 8, 16])
+@parameterize("max_output_len", [4])
-@parameterize("micro_batch_size", [1, 4])
+@parameterize("micro_batch_size", [1])
 @clear_cache_before_run()
-def run_pipeline_inference_test(pp_size, new_length, micro_batch_size):
+def run_pipeline_inference_test(tp_size, pp_size, max_output_len, micro_batch_size):
-    pipeline_inference_test(pp_size, new_length, micro_batch_size)
+    pipeline_inference_test(tp_size, pp_size, max_output_len, micro_batch_size)
+    torch.cuda.empty_cache()
+
+
+@parameterize("tp_size", [2])
+@parameterize("pp_size", [2])
+@parameterize("max_output_len", [4])
+@parameterize("micro_batch_size", [1])
+@clear_cache_before_run()
+def run_tp_pipeline_inference_test(tp_size, pp_size, max_output_len, micro_batch_size):
+    pipeline_inference_test(tp_size, pp_size, max_output_len, micro_batch_size)
     torch.cuda.empty_cache()
 
 
@@ -55,12 +70,18 @@ def check_pipeline_inference(rank, world_size, port):
     run_pipeline_inference_test()
 
 
+def check_tp_pipeline_inference(rank, world_size, port):
+    colossalai.launch(config={}, rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl")
+    run_tp_pipeline_inference_test()
+
+
 @pytest.mark.skipif(not CUDA_SUPPORT, reason="kv-cache manager engine requires cuda version to be higher than 11.5")
 @pytest.mark.dist
 @rerun_if_address_is_in_use()
 @clear_cache_before_run()
 def test_pipeline_inference():
     spawn(check_pipeline_inference, nprocs=2)
+    spawn(check_tp_pipeline_inference, nprocs=4)
 
 
 if __name__ == "__main__":