flashcomm2_qwen3_mode_code

Signed-off-by: iKunHvv <puppybala@outlook.com>

Author: iKunHvv

tests/ut/ops/test_sequence_parallel.py

@@ -0,0 +1,49 @@
+import unittest
+import torch
+from unittest.mock import patch
+
+from vllm_ascend.ops.sequence_parallel import init_metadata_for_flashcomm2
+
+
+class TestInitMetadataForFlashcomm2(unittest.TestCase):
+
+    def setUp(self):
+        patcher = patch("vllm_ascend.ops.sequence_parallel.MetadataForPadding")
+        self.MockMetadata = patcher.start()
+        self.addCleanup(patcher.stop)
+
+    def _run_case(self, tp_size, input_len, expected_padding_flag, expected_pad_size):
+        with patch(
+            "vllm_ascend.ops.sequence_parallel.get_tensor_model_parallel_world_size",
+            return_value=tp_size,
+        ):
+            input_ids = torch.arange(input_len)
+
+            result = init_metadata_for_flashcomm2(input_ids)
+
+            # 验证 MetadataForPadding 调用参数
+            self.MockMetadata.assert_called_once_with(
+                lengths_sum_unpadding=input_len,
+                lengths_sum_padding=((input_len + tp_size - 1) // tp_size) * tp_size,
+                padding_flag=expected_padding_flag,
+                pad_size=expected_pad_size,
+                not_dummy_and_is_prefill=False,
+            )
+
+            # 验证返回值
+            self.assertEqual(result, self.MockMetadata.return_value)
+
+    def test_no_padding(self):
+        self._run_case(tp_size=4, input_len=8, expected_padding_flag=False, expected_pad_size=0)
+
+    def test_with_padding(self):
+        self._run_case(tp_size=4, input_len=10, expected_padding_flag=True, expected_pad_size=2)
+
+    def test_with_padding_non_multiple(self):
+        self._run_case(tp_size=3, input_len=7, expected_padding_flag=True, expected_pad_size=2)
+
+    def test_exact_multiple(self):
+        self._run_case(tp_size=5, input_len=5, expected_padding_flag=False, expected_pad_size=0)
+
+    def test_empty_input(self):
+        self._run_case(tp_size=4, input_len=0, expected_padding_flag=False, expected_pad_size=0)

vllm_ascend/envs.py

@@ -133,8 +133,9 @@
     lambda: bool(int(os.getenv("VLLM_ASCEND_ENABLE_MATMUL_ALLREDUCE", '0'))),
     # Whether to enable FlashComm optimization when tensor parallel is enabled.
     # This feature will get better performance when concurrency is large.
+    # FlashComm optimization: Enable v1 and v2 by setting this flag to 1 or 2 respectively
     "VLLM_ASCEND_ENABLE_FLASHCOMM":
-    lambda: bool(int(os.getenv("VLLM_ASCEND_ENABLE_FLASHCOMM", '0'))),
+    lambda: int(os.getenv("VLLM_ASCEND_ENABLE_FLASHCOMM", '0')),
     # Whether to enable dense model and general optimizations for better performance.
     # Since we modified the base parent class `linear`, this optimization is also applicable to other model types.
     # However, there might be hidden issues, and it is currently recommended to prioritize its use with dense models.

vllm_ascend/models/qwen3_moe.py

@@ -17,11 +17,14 @@
 # Adapted from vllm/model_executor/models/qwen3_moe.py
 # This file is a part of the vllm-ascend project.
 
-from typing import Optional, Union
+from typing import Any, Optional, Union
 
 import torch
+import torch_npu
 from torch import nn
 from transformers import PretrainedConfig
+import torch.distributed as dist
+import vllm_ascend.envs as envs_ascend
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, CompilationLevel, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
@@ -30,7 +33,8 @@
 from vllm.forward_context import get_forward_context
 from vllm.model_executor.layers.fused_moe.layer import FusedMoE
 from vllm.model_executor.layers.layernorm import RMSNorm
-from vllm.model_executor.layers.linear import ReplicatedLinear
+from vllm.model_executor.layers.linear import (ReplicatedLinear,
+                                               RowParallelLinear)
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.model_executor.layers.vocab_parallel_embedding import (
@@ -47,9 +51,11 @@
     make_empty_intermediate_tensors_factory, make_layers, maybe_prefix)
 from vllm.sequence import IntermediateTensors
 
+from vllm.distributed.communication_op import tensor_model_parallel_all_gather
 from vllm_ascend.ops.fused_moe import AscendFusedMoE
 from vllm_ascend.ops.sequence_parallel import (MetadataForPadding,
                                                init_metadata_for_sp)
+                                               init_metadata_for_sp, init_metadata_for_flashcomm2)
 
 
 class CustomSparseMoeBlock(Qwen3MoeSparseMoeBlock):
@@ -125,6 +131,153 @@ def forward(
         return hidden_states
 
 
+class CustomQwen3MoeMLP(Qwen3MoeMLP):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
+        quant_config: Optional[QuantizationConfig] = None,
+        reduce_results: bool = True,
+        prefix: str = "",
+    ) -> None:
+        super().__init__(hidden_size=hidden_size,
+                    intermediate_size=intermediate_size,
+                    hidden_act=hidden_act,
+                    quant_config=quant_config,
+                    reduce_results=reduce_results,
+                    prefix=prefix)
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.enable_flashcomm2 = envs_ascend.VLLM_ASCEND_ENABLE_FLASHCOMM == 2
+        if self.enable_flashcomm2:
+            # if flashcomm2 enabled, replace Linear+AllReduce with All2All+Linear
+            self.down_proj = ReplicatedLinear(
+                intermediate_size,
+                hidden_size,
+                bias=False,
+                quant_config=quant_config,
+                prefix=f"{prefix}.down_proj",
+            )
+        else:
+            self.down_proj = RowParallelLinear(
+                intermediate_size,
+                hidden_size,
+                bias=False,
+                quant_config=quant_config,
+                prefix=f"{prefix}.down_proj",
+            )
+
+    def forward(self, x, _metadata_for_padding=None):
+        #if flashcomm2 enabled, the input of MLP is DP
+        #so we need allgather hidden_states and then use TP in gate_up and use DP(by all2all) in down_proj
+        if self.enable_flashcomm2:
+            x = tensor_model_parallel_all_gather(x, 0)
+        gate_up, _ = self.gate_up_proj(x)
+        x = self.act_fn(gate_up)
+        if self.enable_flashcomm2:
+            #Do not need pad input, because the input of mlp is the output of the attn, which is padded
+            output = torch.empty(x.shape, dtype=x.dtype, device=x.device)
+            dist.all_to_all_single(output,
+                                   x,
+                                   group=get_tp_group().device_group)
+            x = output.reshape(self.tp_size, -1, output.size(-1)) \
+                        .transpose(0, 1) \
+                        .reshape(-1, output.size(-1)*self.tp_size)
+        x, _ = self.down_proj(x)
+        return x
+
+
+class CustomQwen3MoeAttention(Qwen3MoeAttention):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        num_kv_heads: int,
+        rope_theta: float = 10000,
+        rope_scaling: Optional[dict[str, Any]] = None,
+        max_position_embeddings: int = 8192,
+        head_dim: Optional[int] = None,
+        rms_norm_eps: float = 1e-06,
+        qkv_bias: bool = False,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+        dual_chunk_attention_config: Optional[dict[str, Any]] = None,
+    ) -> None:
+        super().__init__(hidden_size=hidden_size,
+                         num_heads=num_heads,
+                         num_kv_heads=num_kv_heads,
+                         rope_theta=rope_theta,
+                         rope_scaling=rope_scaling,
+                         max_position_embeddings=max_position_embeddings,
+                         head_dim=head_dim,
+                         rms_norm_eps=rms_norm_eps,
+                         qkv_bias=qkv_bias,
+                         cache_config=cache_config,
+                         quant_config=quant_config,
+                         prefix=prefix,
+                         dual_chunk_attention_config=dual_chunk_attention_config)
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.enable_flashcomm2 = envs_ascend.VLLM_ASCEND_ENABLE_FLASHCOMM == 2
+        if self.enable_flashcomm2:
+            self.o_proj = ReplicatedLinear(
+                self.total_num_heads * self.head_dim,
+                hidden_size,
+                bias=False,
+                quant_config=quant_config,
+                prefix=f"{prefix}.o_proj",
+            )
+        else:
+            self.o_proj = RowParallelLinear(self.total_num_heads * self.head_dim,
+                                            hidden_size,
+                                            bias=False,
+                                            quant_config=quant_config,
+                                            prefix=f"{prefix}.o_proj")
+
+    def attn_output_all_to_all(self,
+                               attn_output: torch.Tensor,
+                               _metadata_for_padding: Optional[MetadataForPadding] = None) -> torch.Tensor:
+        assert _metadata_for_padding is not None, "Metadata for padding is required for FlashComm2."
+        # pad input because AllGather requires token_num to be divisible by tp_size
+        attn_output = _metadata_for_padding.padding_full(attn_output)
+        output = torch.empty(attn_output.shape,
+                            dtype=attn_output.dtype,
+                            device=attn_output.device)
+        dist.all_to_all_single(output,
+                            attn_output,
+                            group=get_tp_group().device_group)
+        attn_output = output.reshape(self.tp_size, -1, output.size(-1)) \
+                            .transpose(0, 1) \
+                            .reshape(-1, output.size(-1)*self.tp_size)
+        return attn_output
+
+    def forward(
+            self,
+            positions: torch.Tensor,
+            hidden_states: torch.Tensor,
+            _metadata_for_padding: Optional[MetadataForPadding] = None) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+        # Add qk-norm
+        q_by_head = q.view(*q.shape[:-1], q.shape[-1] // self.head_dim,
+                           self.head_dim)
+        q_by_head = self.q_norm(q_by_head)
+        q = q_by_head.view(q.shape)
+
+        k_by_head = k.view(*k.shape[:-1], k.shape[-1] // self.head_dim,
+                           self.head_dim)
+        k_by_head = self.k_norm(k_by_head)
+        k = k_by_head.view(k.shape)
+        q, k = self.rotary_emb(positions, q, k)
+        attn_output = self.attn(q, k, v)
+        if self.enable_flashcomm2:
+            attn_output = self.attn_output_all_to_all(attn_output, _metadata_for_padding)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
 class CustomQwen3MoeDecoderLayer(Qwen3MoeDecoderLayer):
 
     def __init__(
@@ -142,7 +295,7 @@ def __init__(
         rope_scaling = getattr(config, "rope_scaling", None)
         max_position_embeddings = getattr(config, "max_position_embeddings",
                                           8192)
-        self.self_attn = Qwen3MoeAttention(
+        self.self_attn = CustomQwen3MoeAttention(
             hidden_size=self.hidden_size,
             num_heads=config.num_attention_heads,
             num_kv_heads=config.num_key_value_heads,
@@ -178,7 +331,7 @@ def __init__(
                                                   quant_config=quant_config,
                                                   prefix=f"{prefix}.mlp")
         else:
-            self.mlp = Qwen3MoeMLP(hidden_size=config.hidden_size,
+            self.mlp = CustomQwen3MoeMLP(hidden_size=config.hidden_size,
                                    intermediate_size=config.intermediate_size,
                                    hidden_act=config.hidden_act,
                                    quant_config=quant_config,
@@ -191,6 +344,7 @@ def __init__(
         self.enable_sequence_parallelism = (
             vllm_config.compilation_config.pass_config.
             enable_sequence_parallelism if vllm_config is not None else False)
+        self.enable_flashcomm2 = envs_ascend.VLLM_ASCEND_ENABLE_FLASHCOMM == 2
 
     def forward(
         self,
@@ -201,34 +355,37 @@ def forward(
     ) -> torch.Tensor:
 
         # To prevent precision issues during the decoder phase when only prefilling enables SP
-        if not self.enable_sequence_parallelism:
-            self.self_attn.o_proj.reduce_results = True
-        else:
-            self.self_attn.o_proj.reduce_results = not _metadata_for_padding.not_dummy_and_is_prefill if _metadata_for_padding is not None else True
+        if not self.enable_flashcomm2:
+            if not self.enable_sequence_parallelism:
+                self.self_attn.o_proj.reduce_results = True
+            else:
+                self.self_attn.o_proj.reduce_results = not _metadata_for_padding.not_dummy_and_is_prefill if _metadata_for_padding is not None else True
 
         # Self Attention
         if residual is None:
             residual = hidden_states
-            if _metadata_for_padding and _metadata_for_padding.not_dummy_and_is_prefill:
+            if _metadata_for_padding and (_metadata_for_padding.not_dummy_and_is_prefill or self.enable_flashcomm2):
                 residual = _metadata_for_padding.padding_slice(residual)
 
             hidden_states = self.input_layernorm(hidden_states)
         else:
             hidden_states, residual = self.input_layernorm(
                 hidden_states, residual)
 
-            if _metadata_for_padding and _metadata_for_padding.not_dummy_and_is_prefill:
+            if _metadata_for_padding and (_metadata_for_padding.not_dummy_and_is_prefill or self.enable_flashcomm2):
                 hidden_states = _metadata_for_padding.allgather_unpadding_aligned(
                     hidden_states)
 
         hidden_states = self.self_attn(
             positions=positions,
             hidden_states=hidden_states,
+            _metadata_for_padding=_metadata_for_padding
         )
 
-        if _metadata_for_padding and _metadata_for_padding.not_dummy_and_is_prefill:
-            hidden_states = _metadata_for_padding.padding_aligned_reduce_scatter(
-                hidden_states)
+        if not self.enable_flashcomm2:
+            if _metadata_for_padding and _metadata_for_padding.not_dummy_and_is_prefill:
+                hidden_states = _metadata_for_padding.padding_aligned_reduce_scatter(
+                    hidden_states)
 
         # Fully Connected
         hidden_states, residual = self.post_attention_layernorm(
@@ -276,6 +433,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         self.make_empty_intermediate_tensors = (
             make_empty_intermediate_tensors_factory(
                 ["hidden_states", "residual"], config.hidden_size))
+        self.enable_flashcomm2 = envs_ascend.VLLM_ASCEND_ENABLE_FLASHCOMM == 2
 
     def forward(
         self,
@@ -310,7 +468,7 @@ def forward(
 
         hidden_states, _ = self.norm(hidden_states, residual)
 
-        if _metadata_for_padding and _metadata_for_padding.not_dummy_and_is_prefill:
+        if _metadata_for_padding and (_metadata_for_padding.not_dummy_and_is_prefill or self.enable_flashcomm2):
             hidden_states = _metadata_for_padding.allgather_unpadding_aligned(
                 hidden_states)
 
@@ -354,6 +512,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
             self.model.make_empty_intermediate_tensors)
 
         self.enable_sequence_parallelism = vllm_config.compilation_config.pass_config.enable_sequence_parallelism
+        self.enable_flashcomm2 = envs_ascend.VLLM_ASCEND_ENABLE_FLASHCOMM == 2
         # Set MoE hyperparameters
         self.expert_weights: list[torch.Tensor] = []
 
@@ -382,8 +541,13 @@ def forward(
         intermediate_tensors: Optional[IntermediateTensors] = None,
         inputs_embeds: Optional[torch.Tensor] = None,
     ) -> Union[torch.Tensor, IntermediateTensors]:
-        _metadata_for_padding = init_metadata_for_sp(
-            input_ids, self.enable_sequence_parallelism)
+        if self.enable_flashcomm2:
+            if self.enable_sequence_parallelism:
+                raise ValueError(f"Sequence parallelism and FlashComm2 cannot be enabled simultaneously.")
+            _metadata_for_padding = init_metadata_for_flashcomm2(input_ids)
+        else:
+            _metadata_for_padding = init_metadata_for_sp(
+                input_ids, self.enable_sequence_parallelism)
         hidden_states = self.model(input_ids, positions, intermediate_tensors,
                                    inputs_embeds, _metadata_for_padding)
         return hidden_states

vllm_ascend/ops/fused_moe.py

@@ -21,6 +21,7 @@
 import torch
 import torch.distributed as dist
 import torch_npu
+import vllm_ascend.envs as envs_ascend
 from torch import nn
 from vllm.config import get_current_vllm_config
 from vllm.distributed import (get_tensor_model_parallel_rank,
@@ -366,6 +367,7 @@ def __init__(
             num_experts=self.global_num_experts,
             num_global_redundant_experts=self.global_redundant_expert_num,
             num_local_experts=self.local_num_experts)
+        self.enable_flashcomm2 = envs_ascend.VLLM_ASCEND_ENABLE_FLASHCOMM == 2
 
     def naive_multicast(self, x: torch.Tensor,
                         cu_tokens_across_dp_cpu: torch.Tensor):
@@ -415,7 +417,7 @@ def forward(self,
 
         enable_sp = _metadata_for_padding is not None and _metadata_for_padding.not_dummy_and_is_prefill
         tp_size = get_tensor_model_parallel_world_size()
-        if enable_sp:
+        if enable_sp or self.enable_flashcomm2:
             tp_rank = get_tensor_model_parallel_rank()
             mc2_mask_sp = _metadata_for_padding.mc2_mask if _metadata_for_padding is not None else forward_context.mc2_mask
             chunk_mc2_mask = torch.tensor_split(mc2_mask_sp, tp_size, dim=0)