[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
+2
View File
@@ -27,6 +27,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
kMxfp4Dynamic,
kMxfp4Static,
kMxfp8Dynamic,
kNvfp4DynamicToken,
kNvfp4Static,
)
@@ -40,6 +41,7 @@ QUANT_KEY_NAMES: dict[str, QuantKey] = {
"fp8_per_block_dynamic": kFp8Dynamic128Sym,
"mxfp8": kMxfp8Dynamic,
"mxfp4": kMxfp4Dynamic,
"nvfp4_per_token": kNvfp4DynamicToken,
"int8_per_channel_static": kInt8StaticChannelSym,
# Load-time MXFP4-to-int4 for Kimi-K3 on gfx942.
"int4_per_group_32": kInt4Static32,