[Misc] Update default stage config for qwen3-omni (#124)

Signed-off-by: Roger Wang <hey@rogerw.io>

docs/configuration/stage_configs.md

@@ -2,6 +2,9 @@
 
 In vLLM-Omni, the target model is separated into multiple stages, which are processed by different LLMEngines, DiffusionEngines or other types of engines. Depending on different types of stages, such as Autoregressive (AR) stage or Diffusion transformer (DiT) stage, each can choose corresponding schedulers, model workers to load with the Engines in a plug-in fashion.
 
+!!! note
+    Default stage config YAMLs (for example, `vllm_omni/model_executor/stage_configs/qwen2_5_omni.yaml` and `vllm_omni/model_executor/stage_configs/qwen3_omni_moe.yaml`) are bundled and loaded automatically when `stage_configs_path` is not provided. They have been verified to work on 1xH100 for Qwen2.5-Omni and 2xH100 for Qwen3-Omni.
+
 Therefore, as a core part of vLLM-Omni, the stage configs for a model have several main functions:
 
 - Claim partition of stages and their corresponding class implementation in `model_executor/models`.

docs/user_guide/examples/offline_inference/qwen3_omni.md

@@ -46,10 +46,6 @@ sudo apt install ffmpeg
     ``````py
     --8<-- "/mnt/vllm_open_release/vllm-omni-cursor/vllm-omni/examples/offline_inference/qwen3_omni/end2end.py"
     ``````
-??? abstract "qwen3_omni_moe_tp.yaml"
-    ``````yaml
-    --8<-- "/mnt/vllm_open_release/vllm-omni-cursor/vllm-omni/examples/offline_inference/qwen3_omni/qwen3_omni_moe_tp.yaml"
-    ``````
 ??? abstract "run_multiple_prompts.sh"
     ``````sh
     --8<-- "/mnt/vllm_open_release/vllm-omni-cursor/vllm-omni/examples/offline_inference/qwen3_omni/run_multiple_prompts.sh"

examples/offline_inference/qwen3_omni/qwen3_omni_moe_tp.yaml

@@ -1,94 +0,0 @@
-# Stage config for running Qwen3-Omni-MoE with 3-stage architecture
-# Stage 0: Thinker (multimodal understanding + text generation)
-# Stage 1: Talker (text embeddings → 8-layer RVQ codec codes)
-# Stage 2: Code2Wav (8-layer RVQ codes → audio waveform)
-
-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: 0.8
-      enforce_eager: true
-      trust_remote_code: true
-      engine_output_type: latent  # Output hidden states for talker
-      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
-
-  - stage_id: 1
-    runtime:
-       devices: "1"
-       max_batch_size: 1
-    engine_args:
-       model_stage: talker
-       model_arch: Qwen3OmniMoeForConditionalGeneration
-       worker_cls: vllm_omni.worker.gpu_ar_worker.GPUARWorker
-       scheduler_cls: vllm_omni.core.sched.scheduler.OmniScheduler
-       gpu_memory_utilization: 0.5
-       enforce_eager: true
-       trust_remote_code: true
-       engine_output_type: latent  # Output codec codes for code2wav
-      #  tensor_parallel_size: 2
-       enable_prefix_caching: false
-       distributed_executor_backend: "mp"
-       hf_config_name: talker_config
-    engine_input_source: [0]
-    custom_process_input_func: vllm_omni.model_executor.stage_input_processors.qwen3_omni.thinker2talker
-    # final_output: true
-    # final_output_type: text
-    default_sampling_params:
-      temperature: 0.9
-      top_k: 50
-      max_tokens: 4096
-      seed: 42
-      detokenize: False
-      repetition_penalty: 1.05
-      stop_token_ids: [2150]
-
-  - stage_id: 2
-    runtime:
-      devices: "1"
-      max_batch_size: 1
-    engine_args:
-      model_stage: code2wav
-      model_arch: Qwen3OmniMoeForConditionalGeneration
-      worker_cls: vllm_omni.worker.gpu_generation_worker.GPUGenerationWorker
-      scheduler_cls: vllm_omni.core.sched.generation_scheduler.GenerationScheduler
-      enforce_eager: true
-      trust_remote_code: true
-      enable_prefix_caching: false
-      engine_output_type: audio  # Final output: audio waveform
-      gpu_memory_utilization: 0.1
-      distributed_executor_backend: "mp"
-      max_num_batched_tokens: 1000000
-      hf_config_name: thinker_config
-    engine_input_source: [1]
-    custom_process_input_func: vllm_omni.model_executor.stage_input_processors.qwen3_omni.talker2code2wav
-    final_output: true
-    final_output_type: audio
-    default_sampling_params:
-      temperature: 0.0
-      top_p: 1.0
-      top_k: -1
-      max_tokens: 65536
-      seed: 42
-      detokenize: True
-      repetition_penalty: 1.1

examples/offline_inference/qwen3_omni/run_single_prompt_tp.sh

@@ -1,6 +1,5 @@
 python end2end.py --output-wav output_audio \
                   --query-type use_audio \
-                  --init-sleep-seconds 90 \
-                  --stage-configs-path qwen3_omni_moe_tp.yaml
+                  --init-sleep-seconds 90
 
 # init-sleep-seconds works to avoid two vLLM stages initialized at the same time within a card.

vllm_omni/entrypoints/async_omni.py

@@ -90,7 +90,7 @@ class AsyncOmni(EngineClient):
         stage_configs_path: Optional[str] = None,
         log_stats: bool = False,
         log_file: Optional[str] = None,
-        init_sleep_seconds: int = 20,
+        init_sleep_seconds: int = 30,
         shm_threshold_bytes: int = 65536,
         batch_timeout: int = 10,
         init_timeout: int = 60000,

vllm_omni/model_executor/stage_configs/qwen2_5_omni.yaml

@@ -1,4 +1,6 @@
 # stage config for running qwen2.5-omni with architecture of OmniLLM.
+
+# The following config has been verified on 1x H100-80G GPU.
 stage_args:
   - stage_id: 0
     runtime:

vllm_omni/model_executor/stage_configs/qwen3_omni_moe.yaml

@@ -3,24 +3,25 @@
 # Stage 1: Talker (text embeddings → 8-layer RVQ codec codes)
 # Stage 2: Code2Wav (8-layer RVQ codes → audio waveform)
 
+# The following config has been verified on 2x H100-80G GPUs.
 stage_args:
   - stage_id: 0
     runtime:
-      devices: "0"
+      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: 0.8
+      gpu_memory_utilization: 0.6
       enforce_eager: true
       trust_remote_code: true
       engine_output_type: latent  # Output hidden states for talker
       distributed_executor_backend: "mp"
       enable_prefix_caching: false
       hf_config_name: thinker_config
-      # tensor_parallel_size: 2  # Commented out - using single GPU for thinker
+      tensor_parallel_size: 2
     final_output: true
     final_output_type: text
     is_comprehension: true
@@ -42,7 +43,7 @@ stage_args:
        model_arch: Qwen3OmniMoeForConditionalGeneration
        worker_cls: vllm_omni.worker.gpu_ar_worker.GPUARWorker
        scheduler_cls: vllm_omni.core.sched.scheduler.OmniScheduler
-       gpu_memory_utilization: 0.5
+       gpu_memory_utilization: 0.3
        enforce_eager: true
        trust_remote_code: true
        engine_output_type: latent  # Output codec codes for code2wav
@@ -65,7 +66,7 @@ stage_args:
 
   - stage_id: 2
     runtime:
-      devices: "1"
+      devices: "0"
       max_batch_size: 1
     engine_args:
       model_stage: code2wav