From 1406c6ec7194defb5abdc268176a18e65279cb94 Mon Sep 17 00:00:00 2001 From: Yueqian Lin <70319226+linyueqian@users.noreply.github.com> Date: Thu, 4 Dec 2025 11:41:58 -0500 Subject: [PATCH] [Misc] Add stage config for Qwen3-Omni-30B-A3B-Thinking (#172) Signed-off-by: linyueqian --- .../qwen3_omni/qwen3_omni_moe_thinking.yaml | 36 +++++++++++++++++++ 1 file changed, 36 insertions(+) create mode 100644 examples/online_serving/qwen3_omni/qwen3_omni_moe_thinking.yaml diff --git a/examples/online_serving/qwen3_omni/qwen3_omni_moe_thinking.yaml b/examples/online_serving/qwen3_omni/qwen3_omni_moe_thinking.yaml new file mode 100644 index 000000000..9205e5dd9 --- /dev/null +++ b/examples/online_serving/qwen3_omni/qwen3_omni_moe_thinking.yaml @@ -0,0 +1,36 @@ +# Stage config for running Qwen3-Omni-MoE-Thinking (text-only output) +# This config is for models like Qwen3-Omni-30B-A3B-Thinking that only have the +# thinker component and do not support audio output. +# +# Single stage: Thinker (multimodal understanding + text generation) + +# The following config has been verified on 2x H100-80G GPUs. +stage_args: + - stage_id: 0 + runtime: + devices: "0,1" + max_batch_size: 1 + engine_args: + model_stage: thinker + model_arch: Qwen3OmniMoeForConditionalGeneration + worker_cls: vllm_omni.worker.gpu_ar_worker.GPUARWorker + scheduler_cls: vllm_omni.core.sched.scheduler.OmniScheduler + gpu_memory_utilization: 0.9 + enforce_eager: true + trust_remote_code: true + engine_output_type: text + distributed_executor_backend: "mp" + enable_prefix_caching: false + hf_config_name: thinker_config + tensor_parallel_size: 2 + final_output: true + final_output_type: text + is_comprehension: true + default_sampling_params: + temperature: 0.4 + top_p: 0.9 + top_k: 1 + max_tokens: 2048 + seed: 42 + detokenize: True + repetition_penalty: 1.05