minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Disable swa memory for gpt-oss with spec (#17517)

Browse Source
This commit is contained in:
Ke Bao
2026-01-22 01:04:19 +08:00
committed by GitHub
parent a95c9f5b81
commit d725487dc8
1 changed files with 7 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/server_args.py
+7
View File
@@ -1285,6 +1285,13 @@ class ServerArgs:
"Disable hybrid SWA memory for GPT-OSS model with trtllm_mha attention backend." "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) quant_method = get_quantization_config(hf_config)
is_mxfp4_quant_format = quant_method == "mxfp4" is_mxfp4_quant_format = quant_method == "mxfp4"
if is_mxfp4_quant_format: if is_mxfp4_quant_format: