[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
+2
View File
@@ -295,6 +295,7 @@ def test_flashinfer_per_tensor_moe_fp8_no_graph(
expert_map=None,
apply_router_weight_on_input=True,
routed_scaling_factor=1.0,
routing_sink=None,
)
check_accuracy(
@@ -968,6 +969,7 @@ def test_trtllm_mxfp8_minimax2_routing_applies_routed_scale(m: int, workspace_in
router_logits=logits,
e_score_correction_bias=bias,
routed_scaling_factor=scale,
routing_sink=None,
**common,
)
topk_weights, topk_ids = fused_topk_bias(
+2
View File
@@ -781,6 +781,7 @@ def test_trtllm_mxfp4_deferred_finalize(
global_num_experts=num_experts,
expert_map=None,
apply_router_weight_on_input=False,
routing_sink=None,
),
)
return
@@ -1780,6 +1781,7 @@ def test_rocm_mxfp4_moe_oracle(
global_num_experts=num_experts,
expert_map=None,
apply_router_weight_on_input=False,
routing_sink=None,
)
else:
# Modular impl uses topk_weights and topk_ids
File diff suppressed because it is too large Load Diff