[Bugfix] Bind routed-experts capture to the MoE layer, not the kernel (#59455)

Signed-off-by: aoshen02 <aoshen@inferact.ai>
Co-authored-by: aoshen02 <aoshen@inferact.ai>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Eric Tang
2026-10-02 15:50:38 +00:00
committed by GitHub
co-authored by aoshen02 Claude Opus 5.5
parent 0979892992
commit 4c4003bb5f
32 changed files with 423 additions and 898 deletions
@@ -680,6 +680,7 @@ class ModelOptFp8MoEMethod(FusedMoEMethodBase):
topk_group=layer.topk_group,
e_score_correction_bias=layer.e_score_correction_bias,
routed_scaling_factor=layer.routed_scaling_factor,
routing_sink=layer.routing_sink,
)
def apply(
@@ -1092,6 +1093,7 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
topk_group=layer.topk_group,
e_score_correction_bias=layer.e_score_correction_bias,
routed_scaling_factor=layer.routed_scaling_factor,
routing_sink=layer.routing_sink,
)
def apply(
@@ -1461,6 +1463,7 @@ class ModelOptMxFp8FusedMoE(FusedMoEMethodBase):
topk_group=layer.topk_group,
e_score_correction_bias=layer.e_score_correction_bias,
routed_scaling_factor=layer.routed_scaling_factor,
routing_sink=layer.routing_sink,
)
def apply(