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

python/sglang/srt/model_executor/model_runner.py

@@ -2147,6 +2147,16 @@ class ModelRunner:
 
     def _get_attention_backend(self, init_new_workspace: bool = False):
         """Init attention kernel backend."""
+        draft_attn_backend = self.server_args.speculative_draft_attention_backend
+        if self.is_draft_worker and draft_attn_backend:
+            logger.warning(
+                f"Overriding draft attention backend to {draft_attn_backend}."
+            )
+            return self._get_attention_backend_from_str(
+                draft_attn_backend,
+                init_new_workspace=init_new_workspace,
+            )
+
         self.prefill_attention_backend_str, self.decode_attention_backend_str = (
             self.server_args.get_attention_backends()
         )