Skip to content

[CI] Change Qwen3 Omni stage placement strategy #1072

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
hsliuustc0106 merged 3 commits into from
Jan 29, 2026
Merged

[CI] Change Qwen3 Omni stage placement strategy #1072

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
4e429ad
change qwen3 omni yaml placement strategy
ZeldaHuang Jan 29, 2026
515dfdd
update
ZeldaHuang Jan 29, 2026
01f5e08
update
ZeldaHuang Jan 29, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 5 additions & 5 deletions tests/e2e/offline_inference/stage_configs/qwen3_omni_ci.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -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
Expand All @@ -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
Expand All @@ -68,7 +68,7 @@ stage_args:

- stage_id: 2
runtime:
devices: "0"
devices: "1"
max_batch_size: 1
engine_args:
model_stage: code2wav
Expand Down
10 changes: 5 additions & 5 deletions tests/e2e/offline_inference/stage_configs/rocm/qwen3_omni_ci.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -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
Expand All @@ -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
Expand All @@ -68,7 +68,7 @@ stage_args:

- stage_id: 2
runtime:
devices: "0"
devices: "1"
max_batch_size: 1
engine_args:
model_stage: code2wav
Expand Down
10 changes: 5 additions & 5 deletions tests/e2e/online_serving/stage_configs/qwen3_omni_ci.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,22 +8,22 @@ 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
distributed_executor_backend: "mp"
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
Expand All @@ -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
Expand All @@ -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
Expand Down
10 changes: 5 additions & 5 deletions tests/e2e/online_serving/stage_configs/rocm/qwen3_omni_ci.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -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
Expand All @@ -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
Expand All @@ -65,7 +65,7 @@ stage_args:

- stage_id: 2
runtime:
devices: "0"
devices: "1"
max_batch_size: 1
engine_args:
model_stage: code2wav
Expand Down
10 changes: 5 additions & 5 deletions tests/e2e/stage_configs/qwen3_omni_ci.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -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
Expand All @@ -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
Expand All @@ -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
Expand All @@ -68,7 +68,7 @@ stage_args:

- stage_id: 2
runtime:
devices: "0"
devices: "1"
max_batch_size: 1
engine_args:
model_stage: code2wav
Expand Down
10 changes: 5 additions & 5 deletions vllm_omni/model_executor/stage_configs/qwen3_omni_moe.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,22 +9,22 @@ 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
distributed_executor_backend: "mp"
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
Expand All @@ -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
Expand All @@ -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
Expand Down
10 changes: 5 additions & 5 deletions vllm_omni/model_executor/stage_configs/qwen3_omni_moe_async_chunk.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,22 +9,22 @@ 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
distributed_executor_backend: "mp"
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
Expand All @@ -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
Expand Down
10 changes: 5 additions & 5 deletions vllm_omni/model_executor/stage_configs/qwen3_omni_moe_multiconnector.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -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
Expand All @@ -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
Expand Down Expand Up @@ -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
Expand Down
10 changes: 5 additions & 5 deletions vllm_omni/platforms/rocm/stage_configs/qwen3_omni_moe.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,22 +7,22 @@
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
distributed_executor_backend: "mp"
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
Expand All @@ -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
Expand All @@ -68,7 +68,7 @@ stage_args:

- stage_id: 2
runtime:
devices: "0"
devices: "1"
max_batch_size: 1
engine_args:
model_stage: code2wav
Expand Down