[Kernel][MoE] Allow FlashInfer MXINT4 MoE for gated SiLU (#46518)
Signed-off-by: Kaihang Jiang <kaihangj@nvidia.com>
This commit is contained in:
@@ -5,9 +5,13 @@
|
||||
import pytest
|
||||
import torch
|
||||
|
||||
from vllm.model_executor.layers.fused_moe.activation import MoEActivation
|
||||
from vllm.model_executor.layers.fused_moe.experts.marlin_moe import (
|
||||
fused_marlin_moe,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.experts.trtllm_mxint4_moe import (
|
||||
TrtLlmMxint4ExpertsMonolithic,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.router.grouped_topk_router import (
|
||||
grouped_topk,
|
||||
)
|
||||
@@ -77,6 +81,14 @@ __all__ = [
|
||||
]
|
||||
|
||||
|
||||
def test_trtllm_mxint4_activation_supports_vllm_gated_silu():
|
||||
assert TrtLlmMxint4ExpertsMonolithic._supports_activation(MoEActivation.SILU)
|
||||
assert TrtLlmMxint4ExpertsMonolithic._supports_activation(MoEActivation.SWIGLUOAI)
|
||||
assert not TrtLlmMxint4ExpertsMonolithic._supports_activation(
|
||||
MoEActivation.RELU2_NO_MUL
|
||||
)
|
||||
|
||||
|
||||
def marlin_quantize_moe_weights(
|
||||
weights_bf16: torch.Tensor, group_size: int = 32
|
||||
) -> tuple[torch.Tensor, torch.Tensor]:
|
||||
|
||||
@@ -68,8 +68,10 @@ class TrtLlmMxint4ExpertsMonolithic(mk.FusedMoEExpertsMonolithic):
|
||||
|
||||
@staticmethod
|
||||
def _supports_activation(activation: MoEActivation) -> bool:
|
||||
# FlashInfer MxInt4 uses a fused SwiGLU activation.
|
||||
return activation == MoEActivation.SWIGLUOAI
|
||||
# FlashInfer MxInt4 names the standard gated SiLU path "SwiGLU".
|
||||
# In vLLM MoE configs that maps to SILU/silu_and_mul; SWIGLUOAI is
|
||||
# kept as an alias for consistency with other FlashInfer backends.
|
||||
return activation in (MoEActivation.SILU, MoEActivation.SWIGLUOAI)
|
||||
|
||||
@staticmethod
|
||||
def _supports_parallel_config(
|
||||
|
||||
Reference in New Issue
Block a user