[ROCm][Quantization][2/N] Refactor quark_moe w4a8 w/ oracle (#39136)

Signed-off-by: Bowen Bao <bowenbao@amd.com>
This commit is contained in:
Bowen Bao
2026-05-04 19:50:38 -07:00
committed by GitHub
parent 416f9cdede
commit 1e9500410a
6 changed files with 875 additions and 347 deletions
@@ -24,7 +24,7 @@ from vllm.model_executor.layers.fused_moe.oracle.mxfp4 import (
make_mxfp4_moe_kernel,
make_mxfp4_moe_quant_config,
mxfp4_round_up_hidden_size_and_intermediate_size,
select_gpt_oss_mxfp4_moe_backend,
select_deepseek_v4_mxfp4_moe_backend,
select_mxfp4_moe_backend,
)
from vllm.model_executor.layers.linear import LinearBase, UnquantizedLinearMethod
@@ -140,7 +140,7 @@ class GptOssMxfp4MoEMethod(FusedMoEMethodBase):
def __init__(self, moe: FusedMoEConfig):
super().__init__(moe)
self.weight_dtype = "gpt_oss_mxfp4"
self.mxfp4_backend, self.experts_cls = select_gpt_oss_mxfp4_moe_backend(moe)
self.mxfp4_backend, self.experts_cls = select_mxfp4_moe_backend(moe)
self.max_capture_size = (
get_current_vllm_config().compilation_config.max_cudagraph_capture_size
@@ -468,7 +468,7 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
def __init__(self, moe: FusedMoEConfig):
super().__init__(moe)
self.weight_dtype = "mxfp4"
self.mxfp4_backend, self.experts_cls = select_mxfp4_moe_backend(moe)
self.mxfp4_backend, self.experts_cls = select_deepseek_v4_mxfp4_moe_backend(moe)
self.max_capture_size = (
get_current_vllm_config().compilation_config.max_cudagraph_capture_size