[ROCm][Kimi-K3] Add opt-in gfx942 MXFP4-to-int4 conversion (#51274)

Signed-off-by: Markus Hartikainen <markus.hartikainen@amd.com>
Signed-off-by: Hartikainen <mahartik@amd.com>
This commit is contained in:
Markus Hartikainen
2026-10-03 14:28:02 +08:00
committed by GitHub
parent ca0df1f122
commit b0e21b3083
5 changed files with 346 additions and 2 deletions
@@ -65,6 +65,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
kFp8Static128BlockSym,
kFp8StaticChannelSym,
kFp8StaticTensorSym,
kInt4Static32,
kInt8StaticChannelSym,
kMxfp4Static,
kMxfp8Dynamic,
@@ -231,6 +232,9 @@ class OnlineQuantizationConfig(QuantizationConfig):
"""
if spec is None or spec.weight is None:
return None
# Load-time gfx942 requant, not online conversion. Mxfp4MoEMethod owns it.
if spec.weight == kInt4Static32:
return None
cls = table.get(spec.weight)
if cls is None:
raise ValueError(