From a9b4367808a14ae03b5514b7f07b45402756e806 Mon Sep 17 00:00:00 2001 From: tzhouam Date: Mon, 9 Mar 2026 08:22:26 +0000 Subject: [PATCH] Update GPU memory utilization and max number of batched tokens in qwen3_omni.yaml to optimize resource allocation during model execution. Signed-off-by: tzhouam --- tests/perf/stage_configs/qwen3_omni.yaml | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/tests/perf/stage_configs/qwen3_omni.yaml b/tests/perf/stage_configs/qwen3_omni.yaml index fd5586b94..01ca7d9fb 100644 --- a/tests/perf/stage_configs/qwen3_omni.yaml +++ b/tests/perf/stage_configs/qwen3_omni.yaml @@ -47,7 +47,7 @@ stage_args: model_arch: Qwen3OmniMoeForConditionalGeneration worker_type: ar scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler - gpu_memory_utilization: 0.45 + gpu_memory_utilization: 0.6 enforce_eager: false trust_remote_code: true engine_output_type: latent # Output codec codes for code2wav @@ -85,7 +85,7 @@ stage_args: engine_output_type: audio # Final output: audio waveform gpu_memory_utilization: 0.1 distributed_executor_backend: "mp" - max_num_batched_tokens: 1000000 + max_num_batched_tokens: 100000 hf_config_name: thinker_config engine_input_source: [1] custom_process_input_func: vllm_omni.model_executor.stage_input_processors.qwen3_omni.talker2code2wav