[Feature] Spec-Overlap supporting DP-ATTN; PD-Disaggregation; npugraph mode (#12443)

python/sglang/srt/managers/scheduler.py

@@ -846,6 +846,15 @@ class Scheduler(
             self.server_args.disaggregation_transfer_backend
         )
 
+        if self.draft_worker is None or self.spec_algorithm.is_ngram():
+            draft_token_to_kv_pool = None
+        elif self.spec_algorithm.is_eagle() and self.enable_overlap:
+            draft_token_to_kv_pool = (
+                self.draft_worker.draft_worker.draft_runner.token_to_kv_pool
+            )
+        else:
+            draft_token_to_kv_pool = self.draft_worker.model_runner.token_to_kv_pool
+
         if (
             self.disaggregation_mode == DisaggregationMode.DECODE
         ):  # *2 for the headroom.
@@ -874,11 +883,7 @@ class Scheduler(
             self.disagg_decode_prealloc_queue = DecodePreallocQueue(
                 req_to_token_pool=self.req_to_token_pool,
                 token_to_kv_pool_allocator=self.token_to_kv_pool_allocator,
-                draft_token_to_kv_pool=(
-                    None
-                    if self.draft_worker is None or self.spec_algorithm.is_ngram()
-                    else self.draft_worker.model_runner.token_to_kv_pool
-                ),
+                draft_token_to_kv_pool=draft_token_to_kv_pool,
                 req_to_metadata_buffer_idx_allocator=self.req_to_metadata_buffer_idx_allocator,
                 metadata_buffers=self.disagg_metadata_buffers,
                 scheduler=self,
@@ -911,11 +916,7 @@ class Scheduler(
 
             self.disagg_prefill_bootstrap_queue = PrefillBootstrapQueue(
                 token_to_kv_pool=self.token_to_kv_pool_allocator.get_kvcache(),
-                draft_token_to_kv_pool=(
-                    None
-                    if self.draft_worker is None or self.spec_algorithm.is_ngram()
-                    else self.draft_worker.model_runner.token_to_kv_pool
-                ),
+                draft_token_to_kv_pool=draft_token_to_kv_pool,
                 req_to_metadata_buffer_idx_allocator=self.req_to_metadata_buffer_idx_allocator,
                 metadata_buffers=self.disagg_metadata_buffers,
                 tp_rank=self.tp_rank,
@@ -935,6 +936,8 @@ class Scheduler(
             self.disagg_prefill_inflight_queue: List[Req] = []
 
     def init_overlap(self):
+        self.future_map = None
+
         if not self.enable_overlap:
             return