From c48a9f33adc950aa0af7d2c71ce83aa31f0eaf37 Mon Sep 17 00:00:00 2001
From: Ziming Huang <hzm414167@alibaba-inc.com>
Date: Thu, 29 Jan 2026 21:10:09 +0800
Subject: [PATCH] [CI] Change Qwen3 Omni stage placement strategy  (#1072)

Signed-off-by: ZeldaHuang <hzm414167@alibaba-inc.com>
---
 .../offline_inference/stage_configs/qwen3_omni_ci.yaml | 10 +++++-----
 .../stage_configs/rocm/qwen3_omni_ci.yaml              | 10 +++++-----
 .../online_serving/stage_configs/qwen3_omni_ci.yaml    | 10 +++++-----
 .../stage_configs/rocm/qwen3_omni_ci.yaml              | 10 +++++-----
 tests/e2e/stage_configs/qwen3_omni_ci.yaml             | 10 +++++-----
 .../model_executor/stage_configs/qwen3_omni_moe.yaml   | 10 +++++-----
 .../stage_configs/qwen3_omni_moe_async_chunk.yaml      | 10 +++++-----
 .../stage_configs/qwen3_omni_moe_multiconnector.yaml   | 10 +++++-----
 .../platforms/rocm/stage_configs/qwen3_omni_moe.yaml   | 10 +++++-----
 9 files changed, 45 insertions(+), 45 deletions(-)

diff --git a/tests/e2e/offline_inference/stage_configs/qwen3_omni_ci.yaml b/tests/e2e/offline_inference/stage_configs/qwen3_omni_ci.yaml
index 7ac435ea5..477e6e59f 100644
--- a/tests/e2e/offline_inference/stage_configs/qwen3_omni_ci.yaml
+++ b/tests/e2e/offline_inference/stage_configs/qwen3_omni_ci.yaml
@@ -7,21 +7,21 @@
 stage_args:
   - stage_id: 0
     runtime:
-      devices: "0,1"
+      devices: "0"
       max_batch_size: 1
     engine_args:
       model_stage: thinker
       model_arch: Qwen3OmniMoeForConditionalGeneration
       worker_type: ar
       scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-      gpu_memory_utilization: 0.6
+      gpu_memory_utilization: 0.9
       enforce_eager: false
       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
+      tensor_parallel_size: 1
       load_format: dummy
     final_output: true
     final_output_type: text
@@ -44,7 +44,7 @@ stage_args:
        model_arch: Qwen3OmniMoeForConditionalGeneration
        worker_type: ar
        scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-       gpu_memory_utilization: 0.3
+       gpu_memory_utilization: 0.6
        enforce_eager: true
        trust_remote_code: true
        engine_output_type: latent  # Output codec codes for code2wav
@@ -68,7 +68,7 @@ stage_args:
 
   - stage_id: 2
     runtime:
-      devices: "0"
+      devices: "1"
       max_batch_size: 1
     engine_args:
       model_stage: code2wav
diff --git a/tests/e2e/offline_inference/stage_configs/rocm/qwen3_omni_ci.yaml b/tests/e2e/offline_inference/stage_configs/rocm/qwen3_omni_ci.yaml
index 7ac435ea5..477e6e59f 100644
--- a/tests/e2e/offline_inference/stage_configs/rocm/qwen3_omni_ci.yaml
+++ b/tests/e2e/offline_inference/stage_configs/rocm/qwen3_omni_ci.yaml
@@ -7,21 +7,21 @@
 stage_args:
   - stage_id: 0
     runtime:
-      devices: "0,1"
+      devices: "0"
       max_batch_size: 1
     engine_args:
       model_stage: thinker
       model_arch: Qwen3OmniMoeForConditionalGeneration
       worker_type: ar
       scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-      gpu_memory_utilization: 0.6
+      gpu_memory_utilization: 0.9
       enforce_eager: false
       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
+      tensor_parallel_size: 1
       load_format: dummy
     final_output: true
     final_output_type: text
@@ -44,7 +44,7 @@ stage_args:
        model_arch: Qwen3OmniMoeForConditionalGeneration
        worker_type: ar
        scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-       gpu_memory_utilization: 0.3
+       gpu_memory_utilization: 0.6
        enforce_eager: true
        trust_remote_code: true
        engine_output_type: latent  # Output codec codes for code2wav
@@ -68,7 +68,7 @@ stage_args:
 
   - stage_id: 2
     runtime:
-      devices: "0"
+      devices: "1"
       max_batch_size: 1
     engine_args:
       model_stage: code2wav
diff --git a/tests/e2e/online_serving/stage_configs/qwen3_omni_ci.yaml b/tests/e2e/online_serving/stage_configs/qwen3_omni_ci.yaml
index 4db11e11f..8f0161edd 100644
--- a/tests/e2e/online_serving/stage_configs/qwen3_omni_ci.yaml
+++ b/tests/e2e/online_serving/stage_configs/qwen3_omni_ci.yaml
@@ -8,14 +8,14 @@ stage_args:
   - stage_id: 0
     stage_type: llm  # Use llm stage type to launch OmniLLM
     runtime:
-      devices: "0,1"
+      devices: "0"
       max_batch_size: 5
     engine_args:
       model_stage: thinker
       model_arch: Qwen3OmniMoeForConditionalGeneration
       worker_type: ar
       scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-      gpu_memory_utilization: 0.6
+      gpu_memory_utilization: 0.9
       enforce_eager: false
       trust_remote_code: true
       engine_output_type: latent  # Output hidden states for talker
@@ -23,7 +23,7 @@ stage_args:
       enable_prefix_caching: false
       max_num_batched_tokens: 32768
       hf_config_name: thinker_config
-      tensor_parallel_size: 2
+      tensor_parallel_size: 1
       load_format: dummy
     final_output: true
     final_output_type: text
@@ -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: 0.3
+       gpu_memory_utilization: 0.6
        enforce_eager: false
        trust_remote_code: true
        engine_output_type: latent  # Output codec codes for code2wav
@@ -72,7 +72,7 @@ stage_args:
   - stage_id: 2
     stage_type: llm  # Use llm stage type to launch OmniLLM
     runtime:
-      devices: "0"
+      devices: "1"
       max_batch_size: 1
     engine_args:
       model_stage: code2wav
diff --git a/tests/e2e/online_serving/stage_configs/rocm/qwen3_omni_ci.yaml b/tests/e2e/online_serving/stage_configs/rocm/qwen3_omni_ci.yaml
index 3297347c4..59642a77b 100644
--- a/tests/e2e/online_serving/stage_configs/rocm/qwen3_omni_ci.yaml
+++ b/tests/e2e/online_serving/stage_configs/rocm/qwen3_omni_ci.yaml
@@ -6,21 +6,21 @@
 stage_args:
   - stage_id: 0
     runtime:
-      devices: "0,1"
+      devices: "0"
       max_batch_size: 5
     engine_args:
       model_stage: thinker
       model_arch: Qwen3OmniMoeForConditionalGeneration
       worker_type: ar
       scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-      gpu_memory_utilization: 0.6
+      gpu_memory_utilization: 0.9
       enforce_eager: false
       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
+      tensor_parallel_size: 1
     final_output: true
     final_output_type: text
     is_comprehension: true
@@ -42,7 +42,7 @@ stage_args:
        model_arch: Qwen3OmniMoeForConditionalGeneration
        worker_type: ar
        scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-       gpu_memory_utilization: 0.3
+       gpu_memory_utilization: 0.6
        enforce_eager: true
        trust_remote_code: true
        engine_output_type: latent  # Output codec codes for code2wav
@@ -65,7 +65,7 @@ stage_args:
 
   - stage_id: 2
     runtime:
-      devices: "0"
+      devices: "1"
       max_batch_size: 1
     engine_args:
       model_stage: code2wav
diff --git a/tests/e2e/stage_configs/qwen3_omni_ci.yaml b/tests/e2e/stage_configs/qwen3_omni_ci.yaml
index 148a949d8..479e4d6e9 100644
--- a/tests/e2e/stage_configs/qwen3_omni_ci.yaml
+++ b/tests/e2e/stage_configs/qwen3_omni_ci.yaml
@@ -7,14 +7,14 @@
 stage_args:
 - stage_id: 0
   runtime:
-    devices: "0,1"
+    devices: "0"
     max_batch_size: 5
   engine_args:
     model_stage: thinker
     model_arch: Qwen3OmniMoeForConditionalGeneration
     worker_type: ar
     scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-    gpu_memory_utilization: 0.6
+    gpu_memory_utilization: 0.9
     enforce_eager: false
     trust_remote_code: true
     engine_output_type: latent  # Output hidden states for talker
@@ -23,7 +23,7 @@ stage_args:
     max_model_len: 32768
     enable_prefix_caching: false
     hf_config_name: thinker_config
-    tensor_parallel_size: 2
+    tensor_parallel_size: 1
   final_output: true
   final_output_type: text
   is_comprehension: true
@@ -46,7 +46,7 @@ stage_args:
     model_arch: Qwen3OmniMoeForConditionalGeneration
     worker_type: ar
     scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-    gpu_memory_utilization: 0.3
+    gpu_memory_utilization: 0.6
     enforce_eager: false
     trust_remote_code: true
     engine_output_type: latent  # Output codec codes for code2wav
@@ -68,7 +68,7 @@ stage_args:
 
 - stage_id: 2
   runtime:
-    devices: "0"
+    devices: "1"
     max_batch_size: 1
   engine_args:
     model_stage: code2wav
diff --git a/vllm_omni/model_executor/stage_configs/qwen3_omni_moe.yaml b/vllm_omni/model_executor/stage_configs/qwen3_omni_moe.yaml
index d0380a39d..e3dcf940f 100644
--- a/vllm_omni/model_executor/stage_configs/qwen3_omni_moe.yaml
+++ b/vllm_omni/model_executor/stage_configs/qwen3_omni_moe.yaml
@@ -9,14 +9,14 @@ stage_args:
   - stage_id: 0
     stage_type: llm  # Use llm stage type to launch OmniLLM
     runtime:
-      devices: "0,1"
+      devices: "0"
       max_batch_size: 64
     engine_args:
       model_stage: thinker
       model_arch: Qwen3OmniMoeForConditionalGeneration
       worker_type: ar
       scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-      gpu_memory_utilization: 0.6
+      gpu_memory_utilization: 0.9
       enforce_eager: true
       trust_remote_code: true
       engine_output_type: latent  # Output hidden states for talker
@@ -24,7 +24,7 @@ stage_args:
       enable_prefix_caching: false
       max_num_batched_tokens: 32768
       hf_config_name: thinker_config
-      tensor_parallel_size: 2
+      tensor_parallel_size: 1
     final_output: true
     final_output_type: text
     is_comprehension: true
@@ -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: 0.3
+      gpu_memory_utilization: 0.6
       enforce_eager: true
       trust_remote_code: true
       engine_output_type: latent  # Output codec codes for code2wav
@@ -71,7 +71,7 @@ stage_args:
   - stage_id: 2
     stage_type: llm  # Use llm stage type to launch OmniLLM
     runtime:
-      devices: "0"
+      devices: "1"
       max_batch_size: 1
     engine_args:
       model_stage: code2wav
diff --git a/vllm_omni/model_executor/stage_configs/qwen3_omni_moe_async_chunk.yaml b/vllm_omni/model_executor/stage_configs/qwen3_omni_moe_async_chunk.yaml
index 42748960e..1a8b60aa1 100644
--- a/vllm_omni/model_executor/stage_configs/qwen3_omni_moe_async_chunk.yaml
+++ b/vllm_omni/model_executor/stage_configs/qwen3_omni_moe_async_chunk.yaml
@@ -9,14 +9,14 @@ stage_args:
   - stage_id: 0
     stage_type: llm  # Use llm stage type to launch OmniLLM
     runtime:
-      devices: "0,1"
+      devices: "0"
       max_batch_size: 64
     engine_args:
       model_stage: thinker
       model_arch: Qwen3OmniMoeForConditionalGeneration
       worker_type: ar
       scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-      gpu_memory_utilization: 0.6
+      gpu_memory_utilization: 0.9
       enforce_eager: false
       trust_remote_code: true
       engine_output_type: latent  # Output hidden states for talker
@@ -24,7 +24,7 @@ stage_args:
       enable_prefix_caching: false
       max_num_batched_tokens: 32768
       hf_config_name: thinker_config
-      tensor_parallel_size: 2
+      tensor_parallel_size: 1
       custom_process_next_stage_input_func: vllm_omni.model_executor.stage_input_processors.qwen3_omni.thinker2talker_async_chunk
     final_output: true
     final_output_type: text
@@ -41,14 +41,14 @@ stage_args:
   - stage_id: 1
     stage_type: llm  # Use llm stage type to launch OmniLLM
     runtime:
-      devices: "0"
+      devices: "1"
       max_batch_size: 64
     engine_args:
       model_stage: talker
       model_arch: Qwen3OmniMoeForConditionalGeneration
       worker_type: ar
       scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-      gpu_memory_utilization: 0.3
+      gpu_memory_utilization: 0.6
       enforce_eager: false
       trust_remote_code: true
       engine_output_type: latent  # Output codec codes for code2wav
diff --git a/vllm_omni/model_executor/stage_configs/qwen3_omni_moe_multiconnector.yaml b/vllm_omni/model_executor/stage_configs/qwen3_omni_moe_multiconnector.yaml
index 69102df6f..0bc6e48e5 100644
--- a/vllm_omni/model_executor/stage_configs/qwen3_omni_moe_multiconnector.yaml
+++ b/vllm_omni/model_executor/stage_configs/qwen3_omni_moe_multiconnector.yaml
@@ -8,21 +8,21 @@ stage_args:
   - stage_id: 0
     stage_type: llm  # Use llm stage type to launch OmniLLM
     runtime:
-      devices: "0,1"
+      devices: "0"
       max_batch_size: 1
     engine_args:
       model_stage: thinker
       model_arch: Qwen3OmniMoeForConditionalGeneration
       worker_type: ar
       scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-      gpu_memory_utilization: 0.6
+      gpu_memory_utilization: 0.9
       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
+      tensor_parallel_size: 1
     final_output: true
     final_output_type: text
     is_comprehension: true
@@ -48,7 +48,7 @@ stage_args:
        model_arch: Qwen3OmniMoeForConditionalGeneration
        worker_type: ar
        scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-       gpu_memory_utilization: 0.3
+       gpu_memory_utilization: 0.6
        enforce_eager: true
        trust_remote_code: true
        engine_output_type: latent  # Output codec codes for code2wav
@@ -77,7 +77,7 @@ stage_args:
   - stage_id: 2
     stage_type: llm  # Use llm stage type to launch OmniLLM
     runtime:
-      devices: "0"
+      devices: "1"
       max_batch_size: 1
     engine_args:
       model_stage: code2wav
diff --git a/vllm_omni/platforms/rocm/stage_configs/qwen3_omni_moe.yaml b/vllm_omni/platforms/rocm/stage_configs/qwen3_omni_moe.yaml
index 93e05ea9c..31312673a 100644
--- a/vllm_omni/platforms/rocm/stage_configs/qwen3_omni_moe.yaml
+++ b/vllm_omni/platforms/rocm/stage_configs/qwen3_omni_moe.yaml
@@ -7,14 +7,14 @@
 stage_args:
   - stage_id: 0
     runtime:
-      devices: "0,1"
+      devices: "0"
       max_batch_size: 1
     engine_args:
       model_stage: thinker
       model_arch: Qwen3OmniMoeForConditionalGeneration
       worker_type: ar
       scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-      gpu_memory_utilization: 0.6
+      gpu_memory_utilization: 0.9
       enforce_eager: true
       trust_remote_code: true
       engine_output_type: latent  # Output hidden states for talker
@@ -22,7 +22,7 @@ stage_args:
       enable_prefix_caching: false
       max_num_batched_tokens: 32768
       hf_config_name: thinker_config
-      tensor_parallel_size: 2
+      tensor_parallel_size: 1
     final_output: true
     final_output_type: text
     is_comprehension: true
@@ -44,7 +44,7 @@ stage_args:
       model_arch: Qwen3OmniMoeForConditionalGeneration
       worker_type: ar
       scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-      gpu_memory_utilization: 0.3
+      gpu_memory_utilization: 0.6
       enforce_eager: true
       trust_remote_code: true
       engine_output_type: latent  # Output codec codes for code2wav
@@ -68,7 +68,7 @@ stage_args:
 
   - stage_id: 2
     runtime:
-      devices: "0"
+      devices: "1"
       max_batch_size: 1
     engine_args:
       model_stage: code2wav