[quantization] Add per-token NVFP4 quantization MoE support for ReLU2 (#56740)

Signed-off-by: Xuanteng Huang <xuantengh@nvidia.com>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Co-authored-by: OpenAI Codex <noreply@openai.com>
This commit is contained in:
Xuanteng Huang
2026-10-07 22:01:09 +00:00
committed by GitHub
co-authored by Claude Fable 5.1 OpenAI Codex
parent b803edbc9f
commit f38c6778ce
18 changed files with 450 additions and 61 deletions
@@ -334,6 +334,7 @@ def convert_to_nvfp4_moe_kernel_format(
a2_scale: torch.Tensor | None,
is_act_and_mul: bool,
use_a16: bool = False,
trtllm_hidden_alignment: int = 256,
) -> tuple[
torch.Tensor,
torch.Tensor,
@@ -419,6 +420,7 @@ def convert_to_nvfp4_moe_kernel_format(
w2_scale_2=w2_scale_2,
a2_scale=a2_scale,
is_act_and_mul=is_act_and_mul,
trtllm_hidden_alignment=trtllm_hidden_alignment,
)
elif nvfp4_backend == NvFp4MoeBackend.HUMMING:
from vllm.model_executor.layers.quantization.utils.humming import (