From 2b6c4257a01ccb5ffba87b4a1ee7d2b88301c8a3 Mon Sep 17 00:00:00 2001 From: fzyzcjy <5236035+fzyzcjy@users.noreply.github.com> Date: Sun, 9 Nov 2025 00:18:53 +0800 Subject: [PATCH] Fix sending all requests to the first rank in DP attention (#12832) --- python/sglang/srt/managers/tokenizer_manager.py | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/python/sglang/srt/managers/tokenizer_manager.py b/python/sglang/srt/managers/tokenizer_manager.py index 41f92e299..c5a2e35ed 100644 --- a/python/sglang/srt/managers/tokenizer_manager.py +++ b/python/sglang/srt/managers/tokenizer_manager.py @@ -902,10 +902,14 @@ class TokenizerManager(TokenizerCommunicatorMixin): Current policy: - Respect explicit server flag `enable_tokenizer_batch_encode`. - Or, if no request has text or multimodal input (all use pre-tokenized input_ids or input_embeds), batch the requests without tokenization. + - Batch tokenization does not support DP attention yet, and it will make everything goes to the first rank currently """ return batch_size > 0 and ( self.server_args.enable_tokenizer_batch_encode - or not self._batch_has_text(batch_size, requests) + or ( + (not self.server_args.enable_dp_attention) + and (not self._batch_has_text(batch_size, requests)) + ) ) def _send_one_request(