SGlang run error: ValueError: Speculative decoding for Qwen3_5MoeForConditionalGeneration is not compatible with radix cache when using --mamba-scheduler-strategy no_buffer. #108

Description

/sgl-workspace/sglang/python/sglang/launch_server.py:51: UserWarning: 'python -m sglang.launch_server' is still supported, but 'sglang serve' is the recommended entrypoint.
Example: sglang serve --model-path [options]
warnings.warn(
[2026-03-23 14:01:38] WARNING model_config.py:1098: Transformers version 5.3.0 is used for model type qwen3_5_moe. If you experience issues related to RoPE parameters, they may be due to incompatibilities between Transformers >=5.0.0 and some models. You can try downgrading to transf
ormers==4.57.1 as a workaround.
Traceback (most recent call last):
File "", line 198, in _run_module_as_main
File "", line 88, in _run_code
File "/sgl-workspace/sglang/python/sglang/launch_server.py", line 59, in
server_args = prepare_server_args(sys.argv[1:]) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/sgl-workspace/sglang/python/sglang/srt/server_args.py", line 6384, in prepare_server_args
return ServerArgs.from_cli_args(raw_args)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/sgl-workspace/sglang/python/sglang/srt/server_args.py", line 5858, in from_cli_args return cls(**{attr: getattr(args, attr) for attr in attrs})
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "", line 350, in init File "/sgl-workspace/sglang/python/sglang/srt/server_args.py", line 779, in post_init self._handle_model_specific_adjustments() File "/sgl-workspace/sglang/python/sglang/srt/server_args.py", line 1965, in _handle_model_specific_adjustments self._handle_mamba_radix_cache( File "/sgl-workspace/sglang/python/sglang/srt/server_args.py", line 2159, in _handle_mamba_radix_cache raise ValueError( ValueError: Speculative decoding for Qwen3_5MoeForConditionalGeneration is not compatible with radix cache when using --mamba-scheduler-strategy no_buffer.To use radix cache with speculative decoding, please use --mamba-scheduler-strategy extra_buffer and set SGLANG_ENABLE_SPEC_V2=1.

Reproduction

python -m sglang.launch_server --model-path $MODEL_PATH --served-model-name qwen3p5 --host 0.0.0.0 --port 7672 --tp-size 8 --mem-fraction-static 0.8 --context-length 262144 --reasoning-parser qwen3 --speculative-algo NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4

Logs

Environment Information

nightly-dev-20260323-999bad5a

Actually， every sglang docker failed

Known Issue

• The issue hasn't been already addressed in Documentation, Issues, and Discussions.