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

2026-01-22 01:04:19 +08:00
parent a95c9f5b81
commit d725487dc8
1 changed files with 7 additions and 0 deletions
--- 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: