[XPU] WA of topk_softplus_sqrt arg mismatch on XPU (#49408)
Signed-off-by: xiaolong <xiaolong.guo@intel.com>
Liangqiusong committed
Jul 22, 2026 at 16:16 UTC
d6dbdb9b0d6e77b9ac4ef9b298d6dfd8f308b583
1 file changed
+15
vllm/_custom_ops.py
+15
@@ -2447,6 +2447,21 @@ def topk_hash_softplus_sqrt(
2447
hash_indices_table: torch.Tensor | None = None,
2448
is_padding: torch.Tensor | None = None,
2449
) -> None:
2450
+ if current_platform.is_xpu():
2451
+ # TODO: Remove after vllm-xpu-kernels supports is_padding.
2452
+ torch.ops._moe_C.topk_softplus_sqrt(
2453
+ topk_weights,
2454
+ topk_indices,
2455
+ token_expert_indices,
2456
+ gating_output,
2457
+ renormalize,
2458
+ routed_scaling_factor,
2459
+ e_score_correction_bias,
2460
+ input_tokens,
2461
+ hash_indices_table,
2462
+ )
2463
+
2464
+ return
2465
torch.ops._moe_C.topk_softplus_sqrt(
2466
topk_weights,
2467
topk_indices,