[Quantization] [Refactor] Create special "GptOssMxfp4MoeMethod" (#39604)

Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
This commit is contained in:
Yongye Zhu
2026-04-13 12:53:58 -04:00
committed by GitHub
parent 4d042ed85f
commit 739e5945dc
18 changed files with 129 additions and 34 deletions
@@ -30,11 +30,11 @@ from vllm.model_executor.layers.fused_moe.fused_marlin_moe import fused_marlin_m
from vllm.model_executor.layers.fused_moe.oracle.mxfp4 import (
TRITON_BACKENDS,
Mxfp4MoeBackend,
convert_to_mxfp4_moe_kernel_format,
convert_gpt_oss_weight_to_mxfp4_moe_kernel_format,
make_mxfp4_moe_kernel,
make_mxfp4_moe_quant_config,
mxfp4_round_up_hidden_size_and_intermediate_size,
select_mxfp4_moe_backend,
select_gpt_oss_mxfp4_moe_backend,
)
from vllm.model_executor.layers.quantization.utils.marlin_utils_fp8 import (
prepare_fp8_moe_layer_for_marlin,
@@ -995,7 +995,7 @@ class QuarkOCP_MX_MoEMethod(QuarkMoEMethod):
self.w2_precision_config = None
if self.ocp_mx_scheme == "w_mxfp4":
self.mxfp4_backend, self.experts_cls = select_mxfp4_moe_backend(moe)
self.mxfp4_backend, self.experts_cls = select_gpt_oss_mxfp4_moe_backend(moe)
elif self.ocp_mx_scheme.startswith("w_mxfp4"):
# TODO(bowenbao): refactor and introduce backends for other OCP MX schemes.
self.mxfp4_backend = Mxfp4MoeBackend.NONE
@@ -1300,7 +1300,7 @@ class QuarkOCP_MX_MoEMethod(QuarkMoEMethod):
# Convert weights to kernel format
w13, w2, w13_scale, w2_scale, w13_bias, w2_bias = (
convert_to_mxfp4_moe_kernel_format(
convert_gpt_oss_weight_to_mxfp4_moe_kernel_format(
mxfp4_backend=self.mxfp4_backend,
layer=layer,
w13_weight=w13,