From d725487dc8eaa56eaef39d248d46b2cc1c678b27 Mon Sep 17 00:00:00 2001 From: Ke Bao Date: Thu, 22 Jan 2026 01:04:19 +0800 Subject: [PATCH] Disable swa memory for gpt-oss with spec (#17517) --- python/sglang/srt/server_args.py | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/python/sglang/srt/server_args.py b/python/sglang/srt/server_args.py index 2797e0670..678274a4c 100644 --- a/python/sglang/srt/server_args.py +++ b/python/sglang/srt/server_args.py @@ -1285,6 +1285,13 @@ class ServerArgs: "Disable hybrid SWA memory for GPT-OSS model with trtllm_mha attention backend." ) + if self.speculative_algorithm is not None: + # TODO: fix spec with SWA memory cache + self.disable_hybrid_swa_memory = True + logger.warning( + "Disable hybrid SWA memory for GPT-OSS model with speculative decoding." + ) + quant_method = get_quantization_config(hf_config) is_mxfp4_quant_format = quant_method == "mxfp4" if is_mxfp4_quant_format: