[MoE Refactor] Mxfp4 oracle rebased (#37128)

Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
Yongye Zhu
2026-03-21 03:37:04 +00:00
committed by GitHub
co-authored by Claude Opus 4.6
parent c7f98b4d0a
commit 87bd91892f
18 changed files with 1707 additions and 1381 deletions
@@ -25,9 +25,9 @@ from vllm.model_executor.layers.fused_moe.config import (
ocp_mx_moe_quant_config,
)
from vllm.model_executor.layers.fused_moe.fused_marlin_moe import fused_marlin_moe
from vllm.model_executor.layers.quantization.mxfp4 import (
Mxfp4Backend,
get_mxfp4_backend,
from vllm.model_executor.layers.fused_moe.oracle.mxfp4 import (
Mxfp4MoeBackend,
select_mxfp4_moe_backend,
)
from vllm.model_executor.layers.quantization.utils.marlin_utils_fp8 import (
prepare_fp8_moe_layer_for_marlin,
@@ -699,9 +699,9 @@ class QuarkOCP_MX_MoEMethod(QuarkMoEMethod):
f"Please check that the combination is supported in OCP_MX_Scheme."
)
self.mxfp4_backend: Mxfp4Backend | None = None
self.mxfp4_backend: Mxfp4MoeBackend | None = None
if self.ocp_mx_scheme == "w_mxfp4":
self.mxfp4_backend = get_mxfp4_backend(moe.is_lora_enabled)
self.mxfp4_backend, _ = select_mxfp4_moe_backend(moe)
if self.input_quant is not None:
self.static_input_scales = not self.input_quant.get("is_dynamic")