Support using different attention backend for draft decoding. (#14843)

python/sglang/srt/server_args.py

@@ -424,6 +424,7 @@ class ServerArgs:
     speculative_accept_threshold_acc: float = 1.0
     speculative_token_map: Optional[str] = None
     speculative_attention_mode: str = "prefill"
+    speculative_draft_attention_backend: Optional[str] = None
     speculative_moe_runner_backend: Optional[str] = None
     speculative_moe_a2a_backend: Optional[str] = None
     speculative_draft_model_quantization: Optional[str] = None
@@ -3331,6 +3332,12 @@ class ServerArgs:
             help="Attention backend for speculative decoding operations (both target verify and draft extend). Can be one of 'prefill' (default) or 'decode'.",
             default=ServerArgs.speculative_attention_mode,
         )
+        parser.add_argument(
+            "--speculative-draft-attention-backend",
+            type=str,
+            help="Attention backend for speculative decoding drafting.",
+            default=ServerArgs.speculative_draft_attention_backend,
+        )
         parser.add_argument(
             "--speculative-moe-runner-backend",
             type=str,