From d5fbbfd988bee594846a553e58fbcdb675fa2b43 Mon Sep 17 00:00:00 2001 From: Liangsheng Yin Date: Wed, 24 Dec 2025 21:13:26 +0800 Subject: [PATCH] Retry removing wrong logic about max total token in spec decoding (#15748) --- python/sglang/srt/model_executor/model_runner.py | 16 ---------------- 1 file changed, 16 deletions(-) diff --git a/python/sglang/srt/model_executor/model_runner.py b/python/sglang/srt/model_executor/model_runner.py index 0660c9687..f788dee80 100644 --- a/python/sglang/srt/model_executor/model_runner.py +++ b/python/sglang/srt/model_executor/model_runner.py @@ -1876,23 +1876,7 @@ class ModelRunner: self.max_total_num_tokens = self.server_args.draft_runner_cache_size max_num_reqs = self.server_args.max_num_reqs else: - # We are sharing the `token_to_kv_pool`, and both verify and draft tokens - # can be concurrently allocated, so we should give a headroom for it. - extra_tokens = ( - # draft - max_num_reqs - * self.server_args.speculative_num_steps - * self.server_args.speculative_eagle_topk - # verify - + max_num_reqs * self.server_args.speculative_num_draft_tokens - # buffer - + 100 - ) - # Target worker and draft worker shares the same indices for the - # token_to_kv_pool, so we should make sure to match max_total_num_tokens. - self.max_total_num_tokens += extra_tokens self.server_args.draft_runner_cache_size = self.max_total_num_tokens - self.server_args.max_num_reqs = max_num_reqs if max_total_tokens is not None: