[Hardware] Support platforms and plugin system (#774)
Signed-off-by: gcanlin <canlinguosdu@gmail.com>
This commit is contained in:
1 parent
b11d436e19
commit
c9392dff01
103 files changed
+1770
-1218
No files matched your search
@@ -6,13 +6,14 @@
|
||||
# The following config has been verified on 2x H100-80G GPUs.
|
||||
stage_args:
|
||||
- stage_id: 0
|
||||
stage_type: llm # Use llm stage type to launch OmniLLM
|
||||
runtime:
|
||||
devices: "0,1"
|
||||
max_batch_size: 5
|
||||
engine_args:
|
||||
model_stage: thinker
|
||||
model_arch: Qwen3OmniMoeForConditionalGeneration
|
||||
worker_cls: vllm_omni.worker.gpu_ar_worker.GPUARWorker
|
||||
worker_type: ar
|
||||
scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
|
||||
gpu_memory_utilization: 0.6
|
||||
enforce_eager: false
|
||||
@@ -20,6 +21,7 @@ stage_args:
|
||||
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
|
||||
load_format: dummy
|
||||
@@ -36,13 +38,14 @@ stage_args:
|
||||
repetition_penalty: 1.05
|
||||
|
||||
- stage_id: 1
|
||||
stage_type: llm # Use llm stage type to launch OmniLLM
|
||||
runtime:
|
||||
devices: "1"
|
||||
max_batch_size: 5
|
||||
engine_args:
|
||||
model_stage: talker
|
||||
model_arch: Qwen3OmniMoeForConditionalGeneration
|
||||
worker_cls: vllm_omni.worker.gpu_ar_worker.GPUARWorker
|
||||
worker_type: ar
|
||||
scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
|
||||
gpu_memory_utilization: 0.3
|
||||
enforce_eager: false
|
||||
@@ -67,13 +70,14 @@ stage_args:
|
||||
stop_token_ids: [2150]
|
||||
|
||||
- stage_id: 2
|
||||
stage_type: llm # Use llm stage type to launch OmniLLM
|
||||
runtime:
|
||||
devices: "0"
|
||||
max_batch_size: 1
|
||||
engine_args:
|
||||
model_stage: code2wav
|
||||
model_arch: Qwen3OmniMoeForConditionalGeneration
|
||||
worker_cls: vllm_omni.worker.gpu_generation_worker.GPUGenerationWorker
|
||||
worker_type: generation
|
||||
scheduler_cls: vllm_omni.core.sched.omni_generation_scheduler.OmniGenerationScheduler
|
||||
enforce_eager: true
|
||||
trust_remote_code: true
|
||||
|
||||
@@ -7,7 +7,7 @@ stage_args:
|
||||
engine_args:
|
||||
model_stage: thinker
|
||||
model_arch: Qwen3OmniMoeForConditionalGeneration
|
||||
worker_cls: vllm_omni.worker.gpu_ar_worker.GPUARWorker
|
||||
worker_type: ar
|
||||
scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
|
||||
gpu_memory_utilization: 0.6
|
||||
enforce_eager: true
|
||||
|
||||
@@ -12,7 +12,7 @@ stage_args:
|
||||
engine_args:
|
||||
model_stage: thinker
|
||||
model_arch: Qwen3OmniMoeForConditionalGeneration
|
||||
worker_cls: vllm_omni.worker.gpu_ar_worker.GPUARWorker
|
||||
worker_type: ar
|
||||
scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
|
||||
gpu_memory_utilization: 0.6
|
||||
enforce_eager: false
|
||||
@@ -42,7 +42,7 @@ stage_args:
|
||||
engine_args:
|
||||
model_stage: talker
|
||||
model_arch: Qwen3OmniMoeForConditionalGeneration
|
||||
worker_cls: vllm_omni.worker.gpu_ar_worker.GPUARWorker
|
||||
worker_type: ar
|
||||
scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
|
||||
gpu_memory_utilization: 0.3
|
||||
enforce_eager: true
|
||||
@@ -73,7 +73,7 @@ stage_args:
|
||||
engine_args:
|
||||
model_stage: code2wav
|
||||
model_arch: Qwen3OmniMoeForConditionalGeneration
|
||||
worker_cls: vllm_omni.worker.gpu_generation_worker.GPUGenerationWorker
|
||||
worker_type: generation
|
||||
scheduler_cls: vllm_omni.core.sched.omni_generation_scheduler.OmniGenerationScheduler
|
||||
enforce_eager: true
|
||||
trust_remote_code: true
|
||||
|
||||
@@ -27,8 +27,6 @@ from PIL import Image
|
||||
from safetensors.torch import save_file
|
||||
from vllm.utils.network_utils import get_open_port
|
||||
|
||||
from vllm_omni.utils.platform_utils import is_npu
|
||||
|
||||
os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
|
||||
|
||||
MODEL = "Tongyi-MAI/Z-Image-Turbo"
|
||||
@@ -122,8 +120,6 @@ class OmniServer:
|
||||
|
||||
@pytest.fixture(scope="module")
|
||||
def omni_server():
|
||||
if is_npu():
|
||||
pytest.skip("Tongyi-MAI/Z-Image-Turbo is not supported on NPU yet.")
|
||||
with OmniServer(MODEL, ["--num-gpus", "1"]) as server:
|
||||
yield server
|
||||
|
||||
|
||||
@@ -27,7 +27,7 @@ from tests.conftest import (
|
||||
generate_synthetic_video,
|
||||
modify_stage_config,
|
||||
)
|
||||
from vllm_omni.utils import is_rocm
|
||||
from vllm_omni.platforms import current_omni_platform
|
||||
|
||||
models = ["Qwen/Qwen3-Omni-30B-A3B-Instruct"]
|
||||
|
||||
@@ -57,7 +57,7 @@ def get_chunk_config():
|
||||
|
||||
CHUNK_CONFIG_PATH = get_chunk_config()
|
||||
# CI stage config for 2xH100-80G GPUs or AMD GPU MI325
|
||||
if is_rocm():
|
||||
if current_omni_platform.is_rocm():
|
||||
# ROCm stage config optimized for MI325 GPU
|
||||
stage_configs = [str(Path(__file__).parent / "stage_configs" / "rocm" / "qwen3_omni_ci.yaml")]
|
||||
else:
|
||||
@@ -144,7 +144,9 @@ def get_max_batch_size(size_type="few"):
|
||||
return batch_sizes.get(size_type, 5)
|
||||
|
||||
|
||||
@pytest.mark.skipif(is_rocm(), reason="Test skipped on AMD environment due to known output issues")
|
||||
@pytest.mark.skipif(
|
||||
current_omni_platform.is_rocm(), reason="Test skipped on AMD environment due to known output issues"
|
||||
)
|
||||
@pytest.mark.parametrize("omni_server", test_params, indirect=True)
|
||||
def test_mix_to_text_audio_001(client: openai.OpenAI, omni_server, request) -> None:
|
||||
"""
|
||||
@@ -224,7 +226,9 @@ def test_mix_to_text_audio_001(client: openai.OpenAI, omni_server, request) -> N
|
||||
assert similarity > 0.9, "The audio content is not same as the text"
|
||||
|
||||
|
||||
@pytest.mark.skipif(is_rocm(), reason="Test skipped on AMD environment due to known output issues")
|
||||
@pytest.mark.skipif(
|
||||
current_omni_platform.is_rocm(), reason="Test skipped on AMD environment due to known output issues"
|
||||
)
|
||||
@pytest.mark.parametrize("omni_server", test_params, indirect=True)
|
||||
def test_text_to_text_audio_001(client: openai.OpenAI, omni_server) -> None:
|
||||
"""
|
||||
|
||||
Reference in new issue
Block a user