[Feat] Refactor for parallel_config in FusedMoEModularKernel (vllm-project#30282)

Signed-off-by: yewentao256 <zhyanwentao@126.com>
Signed-off-by: Robert Shaw <robshaw@redhat.com>
Co-authored-by: Robert Shaw <robshaw@redhat.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>

Author: 3 people

tests/kernels/moe/modular_kernel_tools/common.py

@@ -594,7 +594,8 @@ def next_power_of_2(x):
     )
 
     modular_kernel = mk.FusedMoEModularKernel(
-        prepare_finalize=prepare_finalize, fused_experts=fused_experts
+        prepare_finalize=prepare_finalize,
+        fused_experts=fused_experts,
     )
 
     return modular_kernel

tests/kernels/moe/test_flashinfer.py

@@ -5,6 +5,7 @@
 import pytest
 import torch
 
+import vllm.model_executor.layers.fused_moe.modular_kernel as mk
 from vllm.config import ParallelConfig, VllmConfig, set_current_vllm_config
 from vllm.model_executor.layers.fused_moe.config import (
     FusedMoEQuantConfig,
@@ -107,6 +108,19 @@ def make_moe_tensors_8bit(
         layer.w2_input_scale = a2_scale
         layer.w13_weight_scale = w13_weight_scale
         layer.w2_weight_scale = w2_weight_scale
+        # Setup dummy config.
+        layer.moe_parallel_config = mk.FusedMoEParallelConfig(
+            tp_size=1,
+            pcp_size=1,
+            dp_size=1,
+            ep_size=1,
+            tp_rank=1,
+            pcp_rank=1,
+            dp_rank=1,
+            ep_rank=1,
+            use_ep=False,
+            all2all_backend="naive",
+        )
 
         register_moe_scaling_factors(layer)
 

vllm/model_executor/layers/fused_moe/cutlass_moe.py

@@ -460,7 +460,6 @@ def cutlass_moe_fp8(
     expert_map: torch.Tensor | None = None,
     apply_router_weight_on_input: bool = False,
     global_num_experts: int = -1,
-    parallel_config=None,
 ) -> torch.Tensor:
     """
     This function computes a a8w8-quantized Mixture of Experts (MoE) layer
@@ -538,7 +537,6 @@ def cutlass_moe_fp8(
             c_strides2=c_strides2,
             quant_config=quant_config,
         ),
-        parallel_config=parallel_config,
     )
 
     return fn(

vllm/model_executor/layers/fused_moe/deep_gemm_moe.py

@@ -293,7 +293,7 @@ def deep_gemm_moe_fp8(
     expert_map: torch.Tensor | None = None,
     a1_scale: torch.Tensor | None = None,
     a2_scale: torch.Tensor | None = None,
-    apply_router_weight_on_input=False,
+    apply_router_weight_on_input: bool = False,
 ) -> torch.Tensor:
     """
     This function computes a a8w8-quantized Mixture of Experts (MoE) layer

vllm/model_executor/layers/fused_moe/fused_moe_modular_method.py

@@ -43,19 +43,14 @@ def make(
         prepare_finalize: FusedMoEPrepareAndFinalize,
         shared_experts: torch.nn.Module | None,
     ) -> "FusedMoEModularMethod":
-        parallel_config = getattr(
-            getattr(moe_layer, "vllm_config", None),
-            "parallel_config",
-            None,
-        )
         return FusedMoEModularMethod(
             old_quant_method,
             FusedMoEModularKernel(
                 prepare_finalize,
                 old_quant_method.select_gemm_impl(prepare_finalize, moe_layer),
                 shared_experts,
                 getattr(moe_layer, "shared_experts_stream", None),
-                parallel_config=parallel_config,
+                moe_parallel_config=moe_layer.moe_parallel_config,
             ),
         )
 

vllm/model_executor/layers/fused_moe/modular_kernel.py

@@ -10,10 +10,12 @@
 import torch
 
 import vllm.envs as envs
-from vllm.config import ParallelConfig, get_current_vllm_config
 from vllm.forward_context import get_forward_context, is_forward_context_available
 from vllm.logger import init_logger
-from vllm.model_executor.layers.fused_moe.config import FusedMoEQuantConfig
+from vllm.model_executor.layers.fused_moe.config import (
+    FusedMoEParallelConfig,
+    FusedMoEQuantConfig,
+)
 from vllm.model_executor.layers.fused_moe.utils import (
     _resize_cache,
     count_expert_num_tokens,
@@ -681,20 +683,23 @@ def __init__(
         fused_experts: FusedMoEPermuteExpertsUnpermute,
         shared_experts: torch.nn.Module | None = None,
         shared_experts_stream: torch.cuda.Stream | None = None,
-        parallel_config: ParallelConfig | None = None,
+        moe_parallel_config: FusedMoEParallelConfig | None = None,
     ):
         super().__init__()
         self.prepare_finalize = prepare_finalize
         self.fused_experts = fused_experts
         self.shared_experts = shared_experts
         self.shared_experts_stream = shared_experts_stream
 
-        # cache whether this worker is using DP+EP
-        if parallel_config is None:
-            parallel_config = get_current_vllm_config().parallel_config
+        # prefer an explicit FusedMoEParallelConfig when available (from
+        # FusedMoE layers / tests).
+        # if not provided, assume this kernel is
+        # running in a non-DP+EP context
+        self.moe_parallel_config: FusedMoEParallelConfig | None = moe_parallel_config
         self.is_dp_ep = (
-            parallel_config.data_parallel_size > 1
-            and parallel_config.enable_expert_parallel
+            moe_parallel_config is not None
+            and moe_parallel_config.dp_size > 1
+            and moe_parallel_config.use_ep
         )
 
         self._post_init_setup()

vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py

@@ -1266,9 +1266,6 @@ def apply(
                     ab_strides2=self.ab_strides2,
                     c_strides1=self.c_strides1,
                     c_strides2=self.ab_strides1_c_strides2,
-                    parallel_config=getattr(
-                        getattr(layer, "vllm_config", None), "parallel_config", None
-                    ),
                 )
 
         else:

vllm/model_executor/layers/quantization/utils/flashinfer_utils.py

@@ -247,11 +247,6 @@ def flashinfer_cutlass_moe_fp8(
     assert quant_config is not None
 
     # Construct modular kernel with block-scale support when requested.
-    parallel_config = getattr(
-        getattr(layer, "vllm_config", None),
-        "parallel_config",
-        None,
-    )
     fused_experts = mk.FusedMoEModularKernel(
         build_flashinfer_fp8_cutlass_moe_prepare_finalize(
             moe=moe, use_deepseek_fp8_block_scale=use_deepseek_fp8_block_scale
@@ -262,7 +257,7 @@ def flashinfer_cutlass_moe_fp8(
             out_dtype=hidden_states.dtype,
             use_deepseek_fp8_block_scale=use_deepseek_fp8_block_scale,
         ),
-        parallel_config=parallel_config,
+        moe_parallel_config=layer.moe_parallel_config,
     )
 
     return fused_experts(