Integrate CuTeDSL MoE for ReLU2 NVFP4 (#49580)

Co-authored-by: OpenAI Codex <codex@openai.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>
This commit is contained in:
danielafrimi
2026-07-28 19:46:38 -07:00
committed by GitHub
co-authored by OpenAI Codex Michael Goin
parent 30c2718eaa
commit 32a423ac0a
3 changed files with 245 additions and 9 deletions
@@ -13,6 +13,9 @@ from vllm.model_executor.layers.fused_moe.config import (
from vllm.model_executor.layers.fused_moe.topk_weight_and_reduce import (
TopKWeightAndReduceNoOP,
)
from vllm.model_executor.layers.quantization.utils.flashinfer_utils import (
activation_to_flashinfer_int,
)
from vllm.model_executor.layers.quantization.utils.quant_utils import (
QuantKey,
kNvfp4Dynamic,
@@ -76,7 +79,7 @@ class FlashInferCuteDSLExperts(mk.FusedMoEExpertsModular):
@staticmethod
def _supports_no_act_and_mul() -> bool:
return False
return True
@staticmethod
def _supports_quant_scheme(
@@ -90,7 +93,7 @@ class FlashInferCuteDSLExperts(mk.FusedMoEExpertsModular):
@staticmethod
def _supports_activation(activation: MoEActivation) -> bool:
return activation == MoEActivation.SILU
return activation in (MoEActivation.SILU, MoEActivation.RELU2_NO_MUL)
@staticmethod
def _supports_parallel_config(
@@ -163,4 +166,5 @@ class FlashInferCuteDSLExperts(mk.FusedMoEExpertsModular):
num_local_experts=self.local_num_experts,
local_expert_offset=self.local_expert_offset,
moe_output=output,
activation_type=activation_to_flashinfer_int(activation),
)