[MoE Refactor] Introduce MoERunner abstraction and move execution logic from FusedMoE to DefaultMoERunner (#32344)

Signed-off-by: Bill Nell <bnell@redhat.com>
This commit is contained in:
bnellnm
2026-02-10 19:51:07 -05:00
committed by GitHub
parent dc6de33c3d
commit d1481ba783
25 changed files with 913 additions and 753 deletions
@@ -764,6 +764,7 @@ class AWQMarlinMoEMethod(FusedMoEMethodBase):
x: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
return fused_marlin_moe(
x,
@@ -501,6 +501,7 @@ class BitsAndBytesMoEMethod(FusedMoEMethodBase):
x: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
from vllm.model_executor.layers.fused_moe import fused_experts
@@ -349,6 +349,7 @@ class CompressedTensorsW4A4Mxfp4MoEMethod(CompressedTensorsMoEMethod):
x: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
assert self.moe_mk is not None
return self.moe_mk(
@@ -361,7 +362,7 @@ class CompressedTensorsW4A4Mxfp4MoEMethod(CompressedTensorsMoEMethod):
global_num_experts=layer.global_num_experts,
expert_map=layer.expert_map,
apply_router_weight_on_input=layer.apply_router_weight_on_input,
shared_experts_input=layer._get_shared_experts_input(x),
shared_experts_input=shared_experts_input,
)
@@ -645,6 +646,7 @@ class CompressedTensorsW4A4Nvfp4MoEMethod(CompressedTensorsMoEMethod):
x: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
assert not self.is_monolithic
assert layer.activation == "silu", "Only SiLU activation is supported."
@@ -673,7 +675,7 @@ class CompressedTensorsW4A4Nvfp4MoEMethod(CompressedTensorsMoEMethod):
global_num_experts=layer.global_num_experts,
expert_map=layer.expert_map,
apply_router_weight_on_input=layer.apply_router_weight_on_input,
shared_experts_input=layer._get_shared_experts_input(x),
shared_experts_input=shared_experts_input,
)
@@ -1064,6 +1066,7 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
x: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
assert not self.is_monolithic
assert self.moe_mk is not None
@@ -1079,7 +1082,7 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
# https://github.com/vllm-project/vllm/commit/84166fee9770e6fba71a96978b3e7d149392fb28 # noqa: E501
expert_map=layer.expert_map,
apply_router_weight_on_input=layer.apply_router_weight_on_input,
shared_experts_input=layer._get_shared_experts_input(x),
shared_experts_input=shared_experts_input,
)
@property
@@ -1203,6 +1206,7 @@ class CompressedTensorsW8A8Int8MoEMethod(CompressedTensorsMoEMethod):
x: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
from vllm.model_executor.layers.fused_moe import fused_experts
@@ -1713,6 +1717,7 @@ class CompressedTensorsWNA16MarlinMoEMethod(CompressedTensorsMoEMethod):
x: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
assert self.kernel_backend == "Marlin"
return fused_marlin_moe(
@@ -1961,6 +1966,7 @@ class CompressedTensorsWNA16MoEMethod(CompressedTensorsMoEMethod):
x: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
from vllm.model_executor.layers.fused_moe import fused_experts
@@ -2575,6 +2581,7 @@ class CompressedTensorsW4A8Fp8MoEMethod(CompressedTensorsMoEMethod):
x: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
if layer.enable_eplb:
raise NotImplementedError(
@@ -140,6 +140,7 @@ class ExpertsInt8MoEMethod(FusedMoEMethodBase):
x: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
from vllm.model_executor.layers.fused_moe import fused_experts
@@ -1010,6 +1010,7 @@ class Fp8MoEMethod(FusedMoEMethodBase):
x: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
assert self.moe_mk is not None
assert not self.is_monolithic
@@ -1023,7 +1024,7 @@ class Fp8MoEMethod(FusedMoEMethodBase):
global_num_experts=layer.global_num_experts,
expert_map=layer.expert_map,
apply_router_weight_on_input=layer.apply_router_weight_on_input,
shared_experts_input=layer._get_shared_experts_input(x),
shared_experts_input=shared_experts_input,
)
@@ -635,6 +635,7 @@ class GGUFMoEMethod(FusedMoEMethodBase):
x: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
assert layer.activation == "silu", "Only SiLU activation is supported."
if layer.apply_router_weight_on_input:
@@ -900,6 +900,7 @@ class GPTQMarlinMoEMethod(FusedMoEMethodBase):
x: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
return fused_marlin_moe(
x,
@@ -958,6 +958,7 @@ class ModelOptFp8MoEMethod(FusedMoEMethodBase):
x: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
assert not self.is_monolithic
@@ -980,7 +981,7 @@ class ModelOptFp8MoEMethod(FusedMoEMethodBase):
global_num_experts=layer.global_num_experts,
expert_map=layer.expert_map,
apply_router_weight_on_input=layer.apply_router_weight_on_input,
shared_experts_input=layer._get_shared_experts_input(x),
shared_experts_input=shared_experts_input,
)
@@ -1524,6 +1525,7 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
x: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
assert not self.is_monolithic
@@ -1551,7 +1553,7 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
global_num_experts=layer.global_num_experts,
expert_map=layer.expert_map,
apply_router_weight_on_input=layer.apply_router_weight_on_input,
shared_experts_input=layer._get_shared_experts_input(x),
shared_experts_input=shared_experts_input,
)
@@ -367,6 +367,7 @@ class MoeWNA16Method(FusedMoEMethodBase):
x: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
from vllm.model_executor.layers.fused_moe import fused_experts
@@ -900,6 +900,7 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
x: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
assert not self.is_monolithic
if layer.enable_eplb:
@@ -419,6 +419,7 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
x: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
if self.rocm_aiter_moe_enabled:
from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (
@@ -607,6 +608,7 @@ class QuarkW4A8Fp8MoEMethod(QuarkMoEMethod):
x: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (
rocm_aiter_fused_experts,
@@ -977,6 +979,7 @@ class QuarkOCP_MX_MoEMethod(QuarkMoEMethod):
x: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
if not self.emulate:
if (