[Kernel] Applies routed_scaling_factor internally (#47408)

Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>
This commit is contained in:
Jee Jee Li
2026-07-07 02:00:54 -07:00
committed by GitHub
parent e55cc59e52
commit 5d23ca47ab
8 changed files with 56 additions and 44 deletions
@@ -43,6 +43,7 @@ def vllm_topk_sigmoid(
gating_output: torch.Tensor,
renormalize: bool = False,
e_score_correction_bias: torch.Tensor | None = None,
routed_scaling_factor: float = 1.0,
) -> tuple[torch.Tensor, ...]:
ops.topk_sigmoid(
topk_weights,
@@ -51,6 +52,7 @@ def vllm_topk_sigmoid(
gating_output,
renormalize,
e_score_correction_bias,
routed_scaling_factor,
)
return topk_weights, topk_indices
@@ -216,9 +218,8 @@ def fused_topk_bias(
gating_output,
renormalize,
e_score_correction_bias,
routed_scaling_factor,
)
if routed_scaling_factor != 1.0:
topk_weights *= routed_scaling_factor
return topk_weights, topk_ids
elif scoring_func == "sqrtsoftplus":
return vllm_topk_softplus_sqrt(