[Auto Sync] Update schedule_batch.py, common.py, eagle_info... (20260105) (#16519)

Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Co-authored-by: cctry <shiyang@x.ai>

python/sglang/srt/managers/schedule_batch.py

@@ -791,20 +791,11 @@ class Req:
 
     def pop_committed_kv_cache(self) -> int:
         """Return the length of committed KV cache and mark them as freed."""
-
+        assert (
-        # NOTE: This function is called exactly once after the request is finished.
+            not self.kv_committed_freed
-        global_server_args = get_global_server_args()
+        ), f"Committed KV cache already freed ({self.kv_committed_len=})"
-        topk = global_server_args.speculative_eagle_topk
+        self.kv_committed_freed = True
-
+        return self.kv_committed_len
-        enable_kv_committed_len = topk is None or topk == 1
-        if enable_kv_committed_len:
-            assert (
-                not self.kv_committed_freed
-            ), f"Committed KV cache already freed ({self.kv_committed_len=})"
-            self.kv_committed_freed = True
-            return self.kv_committed_len
-        else:
-            return len(self.origin_input_ids) + max(len(self.output_ids) - 1, 0)
 
     def pop_overallocated_kv_cache(self) -> Tuple[int, int]:
         """Return the range of over-allocated KV cache and mark them as freed."""

python/sglang/srt/mem_cache/common.py

@@ -484,6 +484,14 @@ def alloc_for_decode(batch: ScheduleBatch, token_per_req: int) -> torch.Tensor:
 
 def release_kv_cache(req: Req, tree_cache: BasePrefixCache, is_insert: bool = True):
     tree_cache.cache_finished_req(req, is_insert=is_insert)
+
+    # MambaRadixCache may alloc mamba state before alloc KV cache
+    if req.req_pool_idx is None:
+        assert isinstance(
+            tree_cache, MambaRadixCache
+        ), "Only MambaRadixCache can handle abort with prefix cache hit before alloc"
+        return
+
     start_p, end_p = req.pop_overallocated_kv_cache()
 
     global_server_args = get_global_server_args()

python/sglang/srt/speculative/eagle_info.py

@@ -394,9 +394,11 @@ class EagleVerifyInput(SpecInput, EagleVerifyInputV2Mixin):
         # Iterate every accepted token and check if req has finished after append the token
         # should be checked BEFORE free kv cache slots
         for i, (req, accept_index_row) in enumerate(zip(batch.reqs, accept_index_cpu)):
+            num_accepted = 0
             for j, idx in enumerate(accept_index_row):
                 if idx == -1:
                     break
+                num_accepted += 1
                 id = predict_cpu[idx]
                 req.output_ids.append(id)
                 req.check_finished()
@@ -414,6 +416,9 @@ class EagleVerifyInput(SpecInput, EagleVerifyInputV2Mixin):
                                 f"{i=}, {req=}\n" f"{accept_index=}\n" f"{predict=}\n"
                             )
                             raise e
+            # Update KV cache tracking for the accepted tokens
+            req.kv_committed_len += num_accepted
+            req.kv_allocated_len = req.kv_committed_len
             if not req.finished():
                 unfinished_index.append(i)
                 if idx == -1:
@@ -442,9 +447,6 @@ class EagleVerifyInput(SpecInput, EagleVerifyInputV2Mixin):
         if page_size == 1:
             # TODO: boolean array index leads to a device sync. Remove it.
             token_to_kv_pool_allocator.free(batch.out_cache_loc[evict_mask])
-            for i, req in enumerate(batch.reqs):
-                req.kv_committed_len += accept_length_list[i] + 1
-                req.kv_allocated_len = req.kv_committed_len
         else:
             if self.topk == 1:
                 # Only evict full empty page. Do not evict partial empty page
@@ -456,9 +458,6 @@ class EagleVerifyInput(SpecInput, EagleVerifyInputV2Mixin):
                     next_power_of_2(self.draft_token_num),
                 )
                 token_to_kv_pool_allocator.free(batch.out_cache_loc[evict_mask])
-                for i, req in enumerate(batch.reqs):
-                    req.kv_committed_len += accept_length_list[i] + 1
-                    req.kv_allocated_len = req.kv_committed_len
             else:
                 # Shift the accepted tokens to the beginning.
                 # Only evict the last part