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
@@ -570,7 +570,7 @@ class Qwen2MoeModel(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(),
                prefix=add_prefix("embed_tokens", prefix),
            )
        else: