Support offline generation scenario for prefill delayer (#16363)

python/sglang/srt/managers/schedule_policy.py

@@ -1,5 +1,7 @@
 from __future__ import annotations
 
+from sglang.srt.managers.prefill_delayer import PrefillDelayerSinglePassExecutor
+
 # Copyright 2023-2024 SGLang Team
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -327,6 +329,7 @@ class PrefillAdder:
         mixed_with_decode_tokens: int = 0,
         priority_scheduling_preemption_threshold: int = 0,
         prefill_max_requests: Optional[int] = None,
+        prefill_delayer_single_pass: Optional[PrefillDelayerSinglePassExecutor] = None,
     ):
         self.page_size = page_size
         self.tree_cache = tree_cache
@@ -366,6 +369,7 @@ class PrefillAdder:
         )
         self.nsa_prefill_cp_in_seq_split = is_nsa_prefill_cp_in_seq_split()
         self.prefill_max_requests = prefill_max_requests
+        self.prefill_delayer_single_pass = prefill_delayer_single_pass
 
     def _get_running_request_total_token_offset(self, req: Req) -> int:
         return (
@@ -614,6 +618,13 @@ class PrefillAdder:
             if input_tokens >= self.rem_input_tokens and len(self.can_run_list) != 0:
                 return AddReqResult.OTHER
 
+            if (self.prefill_delayer_single_pass is not None) and (
+                not self.prefill_delayer_single_pass.negotiate_should_allow_prefill(
+                    local_prefillable=True
+                )
+            ):
+                return AddReqResult.OTHER
+
             if self.rem_chunk_tokens is None or input_tokens <= self.rem_chunk_tokens:
                 # Non-chunked prefill
                 self.can_run_list.append(req)