[server_args] compute mem_fraction_static after dp chunked-prefill division (#28884)

2026-06-22 17:59:05 -07:00
parent ba9d5aed98
commit de3ec2c437
1 changed files with 4 additions and 1 deletions
@@ -3266,9 +3266,12 @@ class ServerArgs:
        if self.mem_fraction_static is None:
            # Constant meta data (e.g., from attention backend)
            reserved_mem = 512
+            effective_chunked_prefill_size = self.chunked_prefill_size // (
+                self.dp_size if self.enable_dp_attention and self.dp_size > 1 else 1
+            )
            # For activation during large prefill
            if self.chunked_prefill_size > 0:
-                reserved_mem += max(self.chunked_prefill_size, 2048) * 1.5
+                reserved_mem += max(effective_chunked_prefill_size, 2048) * 1.5
            else:
                reserved_mem += max(self.max_prefill_tokens, 2048) * 1.5
            # For cuda graphs