[Spec] Add explicit prefill shared-read capability for plugins (#39502)

python/sglang/srt/model_executor/runner_utils/shared_read_event.py

@@ -35,10 +35,7 @@ def maybe_publish_prefill_shared_read_done(
     declared = model_runner.attn_backend.shared_read_ends(forward_batch.forward_mode)
     if declared is not SharedReadEnds.PRE_REPLAY:
         return
-    if (
-        not model_runner.spec_algorithm.is_none()
-        and not model_runner.spec_algorithm.is_dflash_family()
-    ):
+    if not model_runner.spec_algorithm.supports_prefill_shared_read_done():
         # Stage the draft's shared reads before publishing the read-done event.
         stage = getattr(model_runner, "prefill_shared_read_stager", None)
         if stage is None or not stage(forward_batch):

python/sglang/srt/speculative/spec_info.py

@@ -136,6 +136,14 @@ class SpeculativeAlgorithm(Enum):
     def supports_target_verify_for_draft(self) -> bool:
         return self.is_dflash_family()
 
+    def supports_prefill_shared_read_done(self) -> bool:
+        """Whether target EXTEND has no later speculative shared-buffer reader.
+
+        The backend must still declare a pre-replay read end. Other algorithms
+        must stage their draft's shared reads before publishing the target event.
+        """
+        return self.is_none() or self.is_dflash_family()
+
     def supports_mixed_chunk(self) -> bool:
         """Whether mixed chunk prefill may stay enabled with this algorithm.
 

python/sglang/srt/speculative/spec_registry.py

@@ -100,6 +100,10 @@ class CustomSpecAlgo:
     def supports_target_verify_for_draft(self) -> bool:
         return False
 
+    def supports_prefill_shared_read_done(self) -> bool:
+        # Whether target EXTEND has no later speculative shared-buffer reader.
+        return False
+
     def supports_ragged_verify(self) -> bool:
         return False