From 5273f11fd8c65709e3d48bf9fba44c4caf658858 Mon Sep 17 00:00:00 2001
From: wxzhoucs <66296518+Misaka9468@users.noreply.github.com>
Date: Wed, 22 Apr 2026 04:47:01 +0800
Subject: [PATCH] [PD] Resolve missing bootstrap_room problem about fake-decode
 in load-balance method  (#18399)

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: hnyls2002 <lsyincs@gmail.com>
---
 .../sglang/srt/managers/data_parallel_controller.py  | 10 ----------
 python/sglang/srt/managers/tokenizer_manager.py      | 12 +++++++++++-
 2 files changed, 11 insertions(+), 11 deletions(-)

diff --git a/python/sglang/srt/managers/data_parallel_controller.py b/python/sglang/srt/managers/data_parallel_controller.py
index 73db1f498..1b71b1853 100644
--- a/python/sglang/srt/managers/data_parallel_controller.py
+++ b/python/sglang/srt/managers/data_parallel_controller.py
@@ -559,16 +559,6 @@ class DataParallelController:
         if self.maybe_external_dp_rank_routing(req):
             return
 
-        # Set default bootstrap_room if in FAKE auto mode and room is None
-        if (
-            req.bootstrap_room is None
-            and self.server_args.disaggregation_transfer_backend == "fake"
-        ):
-            req.bootstrap_room = self.round_robin_counter
-            self.round_robin_counter = (self.round_robin_counter + 1) % len(
-                self.workers
-            )
-
         assert req.bootstrap_room is not None, (
             "req.bootstrap_room should not be None. Do not send requests directly to "
             "prefill or decode instances; send to the router instead."
diff --git a/python/sglang/srt/managers/tokenizer_manager.py b/python/sglang/srt/managers/tokenizer_manager.py
index aa9567d07..729ed3cd7 100644
--- a/python/sglang/srt/managers/tokenizer_manager.py
+++ b/python/sglang/srt/managers/tokenizer_manager.py
@@ -440,6 +440,8 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
             self.server_args.disaggregation_mode
         )
         self.bootstrap_server = start_disagg_service(self.server_args)
+        # Single-source counter for auto-assigning fake bootstrap_room.
+        self.fake_bootstrap_room_counter = 0
 
         # Encoder Disaggregation
         if self.server_args.language_only:
@@ -974,6 +976,14 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
                 SessionParams(**obj.session_params) if obj.session_params else None
             )
 
+            bootstrap_room = obj.bootstrap_room
+            if (
+                bootstrap_room is None
+                and self.server_args.disaggregation_transfer_backend == "fake"
+            ):
+                bootstrap_room = self.fake_bootstrap_room_counter
+                self.fake_bootstrap_room_counter += 1
+
             tokenized_obj = TokenizedGenerateReqInput(
                 input_text,
                 input_ids,
@@ -988,7 +998,7 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
                 http_worker_ipc=obj.http_worker_ipc,
                 bootstrap_host=obj.bootstrap_host,
                 bootstrap_port=obj.bootstrap_port,
-                bootstrap_room=obj.bootstrap_room,
+                bootstrap_room=bootstrap_room,
                 lora_id=obj.lora_id,
                 input_embeds=input_embeds,
                 positional_embed_overrides=obj.positional_embed_overrides,