[MoE Refactoring][Bugfix]Wrap WNA16 Triton kernel into mk and change compressed tensor kernel selection (#31752)

Signed-off-by: Robert Shaw <robshaw@redhat.com>
Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
Co-authored-by: Robert Shaw <robshaw@redhat.com>
This commit is contained in:
Yongye Zhu
2026-01-08 19:01:30 -05:00
committed by GitHub
co-authored by Robert Shaw
parent 6cdf015c3c
commit d62cfe546d
3 changed files with 148 additions and 4 deletions
@@ -1693,11 +1693,11 @@ class CompressedTensorsWNA16MoEMethod(CompressedTensorsMoEMethod):
from vllm.triton_utils import HAS_TRITON
if HAS_TRITON:
from vllm.model_executor.layers.fused_moe import TritonExperts
from vllm.model_executor.layers.fused_moe import TritonWNA16Experts
layer.w13_weight = layer.w13_weight_packed
layer.w2_weight = layer.w2_weight_packed
return TritonExperts(quant_config=self.moe_quant_config)
return TritonWNA16Experts(quant_config=self.moe_quant_config)
else:
raise NotImplementedError(
"TritonExperts requires Triton. "