[V0.9.1] Add support for flashcomm_v1 in Qwen2.5

Signed-off-by: rjg-lyh <1318825571@qq.com>

Author: rjg-lyh

vllm_ascend/envs.py

@@ -106,6 +106,8 @@
     "VLLM_ASCEND_MODEL_EXECUTE_TIME_OBSERVE":
     lambda: bool(int(os.getenv("VLLM_ASCEND_MODEL_EXECUTE_TIME_OBSERVE", '0'))
                  ),
+    "VLLM_ENABLE_FlashComm":
+    lambda: int(os.getenv("VLLM_ENABLE_FlashComm", '0')),
     # MOE_ALL2ALL_BUFFER:
     #   0: default, normal init.
     #   1: enable moe_all2all_buffer.

vllm_ascend/models/__init__.py

@@ -11,12 +11,16 @@ def register_model():
     from .qwen2_5_vl import \
         AscendQwen2_5_VLForConditionalGeneration  # noqa: F401
     from .qwen2_vl import AscendQwen2VLForConditionalGeneration  # noqa: F401
+    from .qwen2 import CustomQwen2ForCausalLM  # noqa: F401
     from .qwen3 import CustomQwen3ForCausalLM  # noqa: F401
 
     ModelRegistry.register_model(
         "DeepSeekMTPModel",
         "vllm_ascend.models.deepseek_mtp:CustomDeepSeekMTP")
 
+    ModelRegistry.register_model(
+        "Qwen2ForCausalLM", "vllm_ascend.models.qwen2:CustomQwen2ForCausalLM")
+
     ModelRegistry.register_model(
         "Qwen2VLForConditionalGeneration",
         "vllm_ascend.models.qwen2_vl:AscendQwen2VLForConditionalGeneration")

vllm_ascend/models/qwen2.py

@@ -0,0 +1,261 @@
+from collections.abc import Iterable
+from typing import Any, Optional, Union
+
+import torch
+from torch import nn
+import torch.nn.functional as F
+from transformers import Qwen2Config
+
+from vllm.attention import Attention, AttentionType
+from vllm.compilation.decorators import support_torch_compile
+from vllm.config import CacheConfig, VllmConfig
+from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
+from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
+                                               QKVParallelLinear,
+                                               RowParallelLinear)
+from vllm.model_executor.layers.logits_processor import LogitsProcessor
+from vllm.model_executor.layers.quantization import QuantizationConfig
+from vllm.model_executor.layers.rotary_embedding import get_rope
+from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.sequence import IntermediateTensors
+
+from vllm.model_executor.models.interfaces import SupportsLoRA, SupportsPP
+from vllm.model_executor.models.utils import (AutoWeightsLoader, PPMissingLayer, extract_layer_index, maybe_prefix)
+
+from vllm.model_executor.models.qwen2 import Qwen2Model, Qwen2DecoderLayer
+from vllm.distributed import (
+                            get_pp_group,
+                            get_tensor_model_parallel_world_size,
+                            get_tensor_model_parallel_rank,
+                            tensor_model_parallel_all_gather,
+                            tensor_model_parallel_all_reduce,
+                            tensor_model_parallel_reduce_scatter)
+from vllm.forward_context import get_forward_context
+import vllm_ascend.envs as ascend_envs
+from vllm_ascend.ops.linear import MaybeScatterRowParallelLinear
+
+
+def all_gather_and_maybe_unpad(
+    hidden_states: torch.Tensor,
+    pad_size: int,
+) -> torch.Tensor:
+    hidden_states = tensor_model_parallel_all_gather(hidden_states, 0)
+    if pad_size > 0:
+        return hidden_states[:-pad_size, :]
+    return hidden_states
+
+def maybe_pad_and_reduce_scatter(
+    hidden_states: torch.Tensor,
+    pad_size: int,
+) -> torch.Tensor:
+    if pad_size > 0:
+        hidden_states = F.pad(hidden_states, (0, 0, 0, pad_size))
+    hidden_states = tensor_model_parallel_reduce_scatter(hidden_states, 0)
+    return hidden_states
+
+class CustomQwen2DecoderLayer(Qwen2DecoderLayer):
+
+    def __init__(
+        self,
+        config: Qwen2Config,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__(config=config,
+                         cache_config=cache_config,
+                         quant_config=quant_config,
+                         prefix=prefix)
+        self.tp_rank = get_tensor_model_parallel_rank()
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.fc_level = ascend_envs.VLLM_ENABLE_FlashComm
+        if self.fc_level:
+            self.self_attn.reduce_results=False
+            self.mlp.reduce_results=False
+        
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        residual: Optional[torch.Tensor],
+        with_prefill: bool,
+        pad_size: int,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        # Self Attention
+        if residual is None:
+            residual = hidden_states
+            if self.fc_level == 1 and with_prefill:
+                if pad_size > 0:
+                    residual = F.pad(residual, (0, 0, 0, pad_size))
+                residual = torch.chunk(residual, self.tp_size, dim=0)[self.tp_rank]
+            hidden_states = self.input_layernorm(hidden_states)
+        else:
+            hidden_states, residual = self.input_layernorm(
+                hidden_states, residual)
+            if self.fc_level == 1 and with_prefill:
+                hidden_states = all_gather_and_maybe_unpad(hidden_states, pad_size)
+        hidden_states = self.self_attn(
+            positions=positions,
+            hidden_states=hidden_states,
+            with_prefill=with_prefill,
+            pad_size=pad_size,
+        )
+        if self.fc_level == 1 and with_prefill:
+            if pad_size > 0:
+                hidden_states = F.pad(hidden_states, (0, 0, 0, pad_size))
+            hidden_states = tensor_model_parallel_reduce_scatter(hidden_states, 0)
+        else:
+            hidden_states = tensor_model_parallel_all_reduce(hidden_states)
+        # Fully Connected
+        hidden_states, residual = self.post_attention_layernorm(
+            hidden_states, residual)
+        if self.fc_level == 1 and with_prefill:
+            hidden_states = all_gather_and_maybe_unpad(hidden_states, pad_size)
+        hidden_states = self.mlp(hidden_states, with_prefill, pad_size)
+        if self.fc_level == 1 and with_prefill:
+            maybe_pad_and_reduce_scatter(hidden_states, pad_size)
+        else:
+            hidden_states = tensor_model_parallel_all_reduce(hidden_states)
+        return hidden_states, residual
+
+
+@support_torch_compile(
+    dynamic_arg_dims={
+        "input_ids": 0,
+        # positions is of shape (3, seq_len) if mrope is enabled for qwen2-vl,
+        # otherwise (seq_len, ).
+        "positions": -1,
+        "intermediate_tensors": 0,
+        "inputs_embeds": 0,
+    })
+class CustomQwen2Model(Qwen2Model):
+
+    def __init__(self,
+                 *,
+                 vllm_config: VllmConfig,
+                 prefix: str = "",
+                 decoder_layer_type: type[nn.Module] = CustomQwen2DecoderLayer):
+        super().__init__(vllm_config=vllm_config,
+                         prefix=prefix,
+                         decoder_layer_type=decoder_layer_type)
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.fc_level = ascend_envs.VLLM_ENABLE_FlashComm
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        if get_pp_group().is_first_rank:
+            if inputs_embeds is not None:
+                hidden_states = inputs_embeds
+            else:
+                hidden_states = self.get_input_embeddings(input_ids)
+            residual = None
+        else:
+            assert intermediate_tensors is not None
+            hidden_states = intermediate_tensors["hidden_states"]
+            residual = intermediate_tensors["residual"]
+        pad_size = 0
+        with_prefill = get_forward_context().with_prefill
+        if self.fc_level == 1 and with_prefill:
+            num_tokens = hidden_states.size(0)
+            pad_size = (self.tp_size - (num_tokens % self.tp_size)) % self.tp_size
+        for layer in self.layers[self.start_layer:self.end_layer]:
+            hidden_states, residual = layer(
+                positions,
+                hidden_states,
+                residual,
+                with_prefill,
+                pad_size,
+            )
+        if not get_pp_group().is_last_rank:
+            return IntermediateTensors({
+                "hidden_states": hidden_states,
+                "residual": residual
+            })
+        hidden_states, _ = self.norm(hidden_states, residual)
+        if self.fc_level == 1 and with_prefill:
+            hidden_states = all_gather_and_maybe_unpad(hidden_states, pad_size)
+        return hidden_states
+
+
+class CustomQwen2ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
+    # add `CustomQwen2Model` to init self.model
+    packed_modules_mapping = {
+        "qkv_proj": [
+            "q_proj",
+            "k_proj",
+            "v_proj",
+        ],
+        "gate_up_proj": [
+            "gate_proj",
+            "up_proj",
+        ],
+    }
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+        config = vllm_config.model_config.hf_config
+        quant_config = vllm_config.quant_config
+        lora_config = vllm_config.lora_config
+
+        self.config = config
+        self.lora_config = lora_config
+
+        self.quant_config = quant_config
+        self.model = CustomQwen2Model(vllm_config=vllm_config,
+                                prefix=maybe_prefix(prefix, "model"))
+
+        if get_pp_group().is_last_rank:
+            if config.tie_word_embeddings:
+                self.lm_head = self.model.embed_tokens
+            else:
+                self.lm_head = ParallelLMHead(config.vocab_size,
+                                              config.hidden_size,
+                                              quant_config=quant_config,
+                                              prefix=maybe_prefix(
+                                                  prefix, "lm_head"))
+        else:
+            self.lm_head = PPMissingLayer()
+
+        self.logits_processor = LogitsProcessor(config.vocab_size)
+
+        self.make_empty_intermediate_tensors = (
+            self.model.make_empty_intermediate_tensors)
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.model.get_input_embeddings(input_ids)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        hidden_states = self.model(input_ids, positions, intermediate_tensors,
+                                   inputs_embeds)
+        return hidden_states
+
+    def compute_logits(
+        self,
+        hidden_states: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> Optional[torch.Tensor]:
+        logits = self.logits_processor(self.lm_head, hidden_states,
+                                       sampling_metadata)
+        return logits
+
+    def load_weights(self, weights: Iterable[tuple[str,
+                                                   torch.Tensor]]) -> set[str]:
+        loader = AutoWeightsLoader(
+            self,
+            skip_prefixes=(["lm_head."]
+                           if self.config.tie_word_embeddings else None),
+        )
+        return loader.load_weights(weights)

vllm_ascend/worker/model_runner_v1.py

@@ -2033,8 +2033,8 @@ def capture_model(self) -> None:
                 for num_tokens in reversed(self.aclgraph_batch_sizes):
                     for _ in range(self.vllm_config.compilation_config.
                                    cudagraph_num_of_warmups):
-                        self._dummy_run(num_tokens, skip_attn=skip_attn)
-                    self._dummy_run(num_tokens, skip_attn=skip_attn)
+                        self._dummy_run(num_tokens, skip_attn=skip_attn, with_prefill=False)
+                    self._dummy_run(num_tokens, skip_attn=skip_attn, with_prefill=False)
         else:
             logger.info("Skipping NPU graph capture for eager mode.")
             return