Add speculator attention backend switch (#9981)

python/sglang/srt/server_args.py

@@ -262,6 +262,7 @@ class ServerArgs:
     speculative_accept_threshold_single: float = 1.0
     speculative_accept_threshold_acc: float = 1.0
     speculative_token_map: Optional[str] = None
+    speculative_attention_backend: str = "prefill"
 
     # Expert parallelism
     ep_size: int = 1
@@ -1561,6 +1562,13 @@ class ServerArgs:
             help="The path of the draft model's small vocab table.",
             default=ServerArgs.speculative_token_map,
         )
+        parser.add_argument(
+            "--speculative-attention-backend",
+            type=str,
+            choices=["prefill", "decode"],
+            help="Attention backend to use for speculative decoding operations (both target verify and draft extend). 'prefill' (default) or 'decode'.",
+            default=ServerArgs.speculative_attention_backend,
+        )
 
         # Expert parallelism
         parser.add_argument(