[Quantization] Prefer Humming before Marlin backends on SM90 (#56997)

Signed-off-by: jinzhen.ljz <jinzhen.ljz@antgroup.com>
Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Jinzhen Lin
2026-10-06 12:31:35 -07:00
committed by GitHub
co-authored by Codex mgoin Claude Opus 5.5
parent bb87d227d4
commit d1f3d8b870
13 changed files with 159 additions and 15 deletions
+2 -1
View File
@@ -34,6 +34,7 @@ from vllm.forward_context import set_forward_context
from vllm.model_executor.kernels.linear.mxfp8.emulation import (
EmulationMxfp8LinearKernel,
)
from vllm.model_executor.kernels.linear.mxfp8.humming import HummingMxfp8LinearKernel
from vllm.model_executor.kernels.linear.mxfp8.marlin import (
MarlinMxfp8LinearKernel,
)
@@ -1060,7 +1061,7 @@ def test_online_quantization(
assert isinstance(moe._quant_method, expected_moe_cls)
if model_name == PARTIALLY_PREQUANTIZED_MODEL_NAME and isinstance(
o_proj.quant_method.kernel, MarlinMxfp8LinearKernel
o_proj.quant_method.kernel, (MarlinMxfp8LinearKernel, HummingMxfp8LinearKernel)
):
assert o_proj.weight.dtype == torch.int32
elif model_name == PARTIALLY_PREQUANTIZED_MODEL_NAME and isinstance(