support using fa4 on deepseek on blackwell (#9928)

2025-09-17 07:16:06 +08:00
parent 311de47bb7
commit a2f7218a2e
7 changed files with 136 additions and 0 deletions
--- a/python/sglang/srt/models/deepseek_v2.py
+++ b/python/sglang/srt/models/deepseek_v2.py
@@ -1124,6 +1124,9 @@ class DeepseekV2AttentionMLA(nn.Module):
                return AttnForwardMethod.MHA_CHUNKED_KV
            else:
                return _dispatch_mla_subtype()
+        elif attention_backend == "fa4":
+            # TODO(cicirori): use FA4 MHA for DeepSeekV3 for now
+            return AttnForwardMethod.MHA_CHUNKED_KV
        elif attention_backend == "trtllm_mla":
            original_mode = getattr(forward_batch, "_original_forward_mode", None)
            if (