[Feature]Use DispatchGmmCombineDecode operator to replace MC2(Optional) (#5040)

### What this PR does / why we need it?

This PR adds model-side integration for the previously introduced
experimental AscendC fused operator DispatchGmmCombineDecode, used in
MoE decoding.

The operator implementation itself was added in a prior PR[#4139
](https://github.com/vllm-project/vllm-ascend/pull/4139).
This change only adapts the model execution path to optionally use the
fused operator.

When the environment variable VLLM_ASCEND_ENABLE_FUSED_MC2=2 is set, the
original MC2 path composed of multiple operators (A8W8 dispatch → GMM →
SwiGLU → GMM → combine) might be replaced by the single fused operator
DispatchGmmCombineDecode.

By default, the existing multi-operator MC2 implementation is preserved.

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

Signed-off-by: wangqiankun <wangqiankun13@huawei.com>

vllm_ascend/ascend_forward_context.py

@@ -253,12 +253,24 @@ def select_moe_comm_method(num_tokens: int,
         ascend_config = get_ascend_config()
         dynamic_eplb = ascend_config.dynamic_eplb or ascend_config.expert_map_record_path
         # TODO: drop the EP-size guard when dispatch_ffn_combine supports larger EP sizes
-        fused_mc2_enable = envs_ascend.VLLM_ASCEND_ENABLE_FUSED_MC2 and quant_type == "w8a8_dynamic" and get_ep_group(
+        # TODO: drop dynamic_eplb guard when dispatch_gmm_combine_decode supports tensor list inputs
-        ).world_size <= 16 and (not dynamic_eplb) and (not is_mtp_model)
+        # TODO: add guard for dispatch_gmm_combine_decode when mtp uses float while moe uses w8a8
+        fused_mc2_enable = envs_ascend.VLLM_ASCEND_ENABLE_FUSED_MC2 and quant_type == "w8a8_dynamic" and (
+            not dynamic_eplb)
         if num_tokens <= mc2_tokens_capacity:
-            moe_comm_type = MoECommType.FUSED_MC2 if fused_mc2_enable else MoECommType.MC2
+            fused_decode_enable = fused_mc2_enable
+            if envs_ascend.VLLM_ASCEND_ENABLE_FUSED_MC2 == 1:
+                fused_decode_enable = fused_mc2_enable and get_ep_group(
+                ).world_size <= 16 and (not is_mtp_model)
+            moe_comm_type = MoECommType.FUSED_MC2 if fused_decode_enable else MoECommType.MC2
         else:
-            moe_comm_type = MoECommType.FUSED_MC2 if fused_mc2_enable else MoECommType.ALLTOALL
+            fused_prefill_enable = fused_mc2_enable
+            if envs_ascend.VLLM_ASCEND_ENABLE_FUSED_MC2 == 1:
+                fused_prefill_enable = fused_mc2_enable and get_ep_group(
+                ).world_size <= 16 and (not is_mtp_model)
+            elif envs_ascend.VLLM_ASCEND_ENABLE_FUSED_MC2 == 2:
+                fused_prefill_enable = False
+            moe_comm_type = MoECommType.FUSED_MC2 if fused_prefill_enable else MoECommType.ALLTOALL
 
     else:
         raise ValueError(f"Unsupported soc_version: {soc_version}")

vllm_ascend/envs.py

@@ -135,7 +135,13 @@ env_variables: Dict[str, Callable[[], Any]] = {
     # Whether to anbale dynamic EPLB
     "DYNAMIC_EPLB":
     lambda: os.getenv("DYNAMIC_EPLB", "false").lower(),
-    # Whether to anbale fused mc2(dispatch_gmm_combine_decode/dispatch_ffn_combine operator)
+    # Whether to enable fused mc2(`dispatch_gmm_combine_decode`/`dispatch_ffn_combine` operator)
+    # 0, or not set: default ALLTOALL and MC2 will be used.
+    # 1: ALLTOALL and MC2 might be replaced by `dispatch_ffn_combine` operator.
+    # `dispatch_ffn_combine` can be used only for moe layer with W8A8, EP<=16, non-mtp, non-dynamic-eplb.
+    # 2: MC2 might be replaced by `dispatch_gmm_combine_decode` operator.
+    # `dispatch_gmm_combine_decode` can be used only for **decode node** moe layer
+    # with W8A8, non-dynamic-eplb. And MTP layer must be W8A8.
     "VLLM_ASCEND_ENABLE_FUSED_MC2":
     lambda: int(os.getenv("VLLM_ASCEND_ENABLE_FUSED_MC2", '0')),
 }

vllm_ascend/ops/fused_moe/fused_moe.py

@@ -345,7 +345,7 @@ class AscendFusedMoE(FusedMoE):
                 shared_out = fc3_context.shared_experts(hidden_states)
                 # NOTE: This is exactly the opposite of `maybe_all_reduce_tensor_model_parallel`
                 moe_comm_type = forward_context.moe_comm_type
-                if moe_comm_type in {MoECommType.ALLTOALL, MoECommType.MC2} \
+                if moe_comm_type in {MoECommType.ALLTOALL, MoECommType.MC2, MoECommType.FUSED_MC2} \
                         and not shared_expert_dp_enabled():
                     shared_out = tensor_model_parallel_all_reduce(shared_out)
                 set_flash_common3_context(shared_out=shared_out)

vllm_ascend/ops/fused_moe/moe_comm_method.py

@@ -291,9 +291,9 @@ class FusedMC2CommImpl(MoECommMethod):
         assert not (
             w1_scale is None or w2_scale is None
         ), "w1_scale and w2_scale cannot be None for FusedMC2CommImpl."
-        out = torch.empty_like(hidden_states)
 
         if envs_ascend.VLLM_ASCEND_ENABLE_FUSED_MC2 == 1:
+            out = torch.empty_like(hidden_states)
             torch.ops._C_ascend.dispatch_ffn_combine(
                 x=hidden_states,
                 weight1=w1[0],
@@ -307,7 +307,21 @@ class FusedMC2CommImpl(MoECommMethod):
                 out=out,
             )
         elif envs_ascend.VLLM_ASCEND_ENABLE_FUSED_MC2 == 2:
-            raise NotImplementedError()
+            assert expert_map is not None, "expert_map cannot be None."
+            out, _ = torch.ops._C_ascend.dispatch_gmm_combine_decode(
+                x=hidden_states,
+                expert_ids=topk_ids,
+                gmm1_permuted_weight=w1[0],
+                gmm1_permuted_weight_scale=w1_scale[0],
+                gmm2_weight=w2[0],
+                gmm2_weight_scale=w2_scale[0],
+                expert_smooth_scales=None,
+                expert_scales=topk_weights.to(torch.float32),
+                group_ep=self.token_dispatcher.moe_all_to_all_group_name,
+                ep_rank_size=self.token_dispatcher.ep_world_size,
+                ep_rank_id=self.token_dispatcher.ep_rank_id,
+                moe_expert_num=len(expert_map),
+                global_bs=self.token_dispatcher.fused_global_bs)
         else:
             raise ValueError(
                 f"Wrong value of {envs_ascend.VLLM_ASCEND_ENABLE_FUSED_MC2=}")

vllm_ascend/ops/fused_moe/token_dispatcher.py

@@ -125,6 +125,7 @@ class TokenDispatcherWithMC2(MoETokenDispatcher):
             max_num_tokens = min(max_num_reqs * uniform_decode_query_len, 512)
         num_tokens_per_tp_rank = (max_num_tokens + tp_size - 1) // tp_size
         self.global_bs = num_tokens_per_tp_rank * self.ep_world_size
+        self.fused_global_bs = max_num_tokens * self.ep_world_size
 
     def get_dispatch_mc2_kwargs(
         self,

vllm_ascend/quantization/w8a8_dynamic.py

@@ -231,6 +231,10 @@ class AscendW8A8DynamicFusedMoEMethod:
         topk_weights = topk_weights.to(self.in_dtype)
 
         moe_comm_method = get_forward_context().moe_comm_method
+        # When VLLM_ASCEND_ENABLE_FUSED_MC2 == 2, use dispatch_gmm_combine_decode, need fp32 scale
+        w2_weight_scale_fp32_flag = (
+            get_forward_context().moe_comm_type == MoECommType.FUSED_MC2
+            and envs_ascend.VLLM_ASCEND_ENABLE_FUSED_MC2 == 2)
         if self.dynamic_eplb:
             w1 = layer.w13_weight_list
             w1_scale = layer.w13_weight_scale_fp32_list
@@ -240,7 +244,10 @@ class AscendW8A8DynamicFusedMoEMethod:
             w1 = [layer.w13_weight]
             w1_scale = [layer.w13_weight_scale_fp32]
             w2 = [layer.w2_weight]
-            w2_scale = [layer.w2_weight_scale]
+            w2_scale = [
+                layer.w2_weight_scale_fp32
+                if w2_weight_scale_fp32_flag else layer.w2_weight_scale
+            ]
 
         fused_scale_flag = (get_forward_context().moe_comm_type
                             == MoECommType.FUSED_MC2
@@ -279,6 +286,8 @@ class AscendW8A8DynamicFusedMoEMethod:
             layer.w13_weight_offset.data.shape[0], -1)
         layer.w2_weight_scale.data = layer.w2_weight_scale.data.view(
             layer.w2_weight_scale.data.shape[0], -1)
+        layer.w2_weight_scale_fp32 = layer.w2_weight_scale.data.to(
+            torch.float32)
         layer.w2_weight_offset.data = layer.w2_weight_offset.data.view(
             layer.w2_weight_offset.data.shape[0], -1)