From d7b942552927e2df4689577ca434d20f497e349b Mon Sep 17 00:00:00 2001 From: chx96642264 Date: Sat, 18 Jul 2026 19:32:03 +0800 Subject: [PATCH] [NPU] fix: skip Triton embedding kernel on NPU to avoid kernel launch failure (#31636) --- python/sglang/srt/layers/vocab_parallel_embedding.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/python/sglang/srt/layers/vocab_parallel_embedding.py b/python/sglang/srt/layers/vocab_parallel_embedding.py index adc7a3d53..42fdf6a21 100644 --- a/python/sglang/srt/layers/vocab_parallel_embedding.py +++ b/python/sglang/srt/layers/vocab_parallel_embedding.py @@ -503,6 +503,8 @@ class VocabParallelEmbedding(torch.nn.Module): def _use_triton_embedding(self, input_: torch.Tensor) -> bool: """Whether the fused Triton kernel can replace the mask+gather+fill unit.""" + if _is_npu: + return False if self.tp_size == 1: return False if not isinstance(self.quant_method, UnquantizedEmbeddingMethod):