From 2babf88f245ddd93bc8d6ad14670a82c8ab5177b Mon Sep 17 00:00:00 2001 From: ybyang <10629930+whybeyoung@users.noreply.github.com> Date: Fri, 9 Jan 2026 19:20:17 +0800 Subject: [PATCH] fix spec qwen3 pd error (#16708) --- python/sglang/srt/mem_cache/memory_pool.py | 9 +++++++++ 1 file changed, 9 insertions(+) diff --git a/python/sglang/srt/mem_cache/memory_pool.py b/python/sglang/srt/mem_cache/memory_pool.py index 65d562a27..cb7061440 100644 --- a/python/sglang/srt/mem_cache/memory_pool.py +++ b/python/sglang/srt/mem_cache/memory_pool.py @@ -304,8 +304,17 @@ class MambaPool: return dst_index def get_contiguous_buf_infos(self): + """ + Get buffer info for RDMA registration. + Only returns conv and temporal state buffers, excluding intermediate buffers + used for speculative decoding (intermediate_ssm, intermediate_conv_window). + """ state_tensors = [] for field in vars(self.mamba_cache): + # Skip intermediate buffers used only for speculative decoding + # These buffers have different size (spec_state_size + 1) and should not be transferred + if field in ("intermediate_ssm", "intermediate_conv_window"): + continue value = getattr(self.mamba_cache, field) if isinstance(value, list): state_tensors.extend(value)