[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
@@ -790,6 +790,7 @@ class AutoAWQMoEMethod(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,
)
@@ -779,4 +779,5 @@ class AutoGPTQMoEMethod(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,
)
@@ -293,6 +293,7 @@ class CompressedTensorsW4A4Nvfp4MoEMethod(CompressedTensorsMoEMethod):
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(
@@ -287,6 +287,7 @@ class CompressedTensorsW4A8Int8MoEMethod(CompressedTensorsMoEMethod):
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(
@@ -366,6 +366,7 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
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(
@@ -291,4 +291,5 @@ class CompressedTensorsW8A8Int8MoEMethod(CompressedTensorsMoEMethod):
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,
)
@@ -179,6 +179,7 @@ class CompressedTensorsW8A8Mxfp8MoEMethod(CompressedTensorsMoEMethod):
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(
@@ -528,6 +528,7 @@ class CompressedTensorsWNA16MoEMethod(CompressedTensorsMoEMethod):
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(
@@ -818,6 +818,7 @@ class Fp8MoEMethod(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(
@@ -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(
@@ -531,6 +531,7 @@ class MoeWNA16Method(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,
)
@staticmethod
@@ -472,6 +472,7 @@ class GptOssMxfp4MoEMethod(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,
)
@@ -900,4 +901,5 @@ class Mxfp4MoEMethod(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,
)
@@ -156,6 +156,7 @@ class OnlineMoEMethodBase(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(
@@ -1980,6 +1980,7 @@ class QuarkOCP_MX_MoEMethod(QuarkMoEMethod):
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,
)