Use attn tp group in embedding for more models (#17570)

2026-01-24 13:37:44 +08:00
parent 176da1bbdd
commit fb683be6eb
19 changed files with 19 additions and 19 deletions
--- a/python/sglang/srt/models/longcat_flash.py
+++ b/python/sglang/srt/models/longcat_flash.py
@@ -503,7 +503,7 @@ class LongcatFlashModel(nn.Module):
        self.embed_tokens = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
-            enable_tp=not is_dp_attention_enabled(),
+            use_attn_tp_group=is_dp_attention_enabled(),
        )

        self.alt_stream = torch.cuda.Stream()