[Core]Add GPU Diffusion Runner

[princepride]

Purpose

Related: #800

This PR refactors the GPU diffusion worker architecture to improve code organization and maintainability:

• Separated model runner logic: Extracted GPUDiffusionModelRunner from GPUDiffusionWorker to follow the separation of concerns principle
• Improved naming consistency: Renamed gpu_worker.py → gpu_diffusion_worker.py and test_gpu_worker.py → test_gpu_diffusion_worker.py for better clarity
• Adjust NPU worker: Updated npu_worker.py to align with the new architecture and add missing functionality
• Added comprehensive unit tests: Implemented detailed tests for load_weights, sleep, and wake_up methods with proper mocking

Test Plan

Unit Test

pytest tests/diffusion/test_gpu_diffusion_worker.py -v

Result:

============================================ test session starts ============================================
platform linux -- Python 3.13.11, pytest-9.0.2, pluggy-1.6.0 -- /proj-tango-pvc/users/zhipeng.wang/workspace/vllm-omni/.venv/bin/python3
cachedir: .pytest_cache
rootdir: /proj-tango-pvc/users/zhipeng.wang/workspace/vllm-omni
configfile: pyproject.toml
plugins: cov-7.0.0, anyio-4.12.1
collected 8 items                                                                                           

tests/diffusion/test_gpu_diffusion_worker.py::TestGPUDiffusionWorkerLoadWeights::test_load_weights_calls_pipeline PASSED [ 12%]
tests/diffusion/test_gpu_diffusion_worker.py::TestGPUDiffusionWorkerLoadWeights::test_load_weights_empty_iterable PASSED [ 25%]
tests/diffusion/test_gpu_diffusion_worker.py::TestGPUDiffusionWorkerSleep::test_sleep_level_1 PASSED  [ 37%]
tests/diffusion/test_gpu_diffusion_worker.py::TestGPUDiffusionWorkerSleep::test_sleep_level_2 PASSED  [ 50%]
tests/diffusion/test_gpu_diffusion_worker.py::TestGPUDiffusionWorkerSleep::test_sleep_memory_freed_validation PASSED [ 62%]
tests/diffusion/test_gpu_diffusion_worker.py::TestGPUDiffusionWorkerWakeUp::test_wake_up_without_buffers PASSED [ 75%]
tests/diffusion/test_gpu_diffusion_worker.py::TestGPUDiffusionWorkerWakeUp::test_wake_up_with_buffers PASSED [ 87%]
tests/diffusion/test_gpu_diffusion_worker.py::TestGPUDiffusionWorkerWakeUp::test_wake_up_partial_buffer_restore PASSED [100%]

Test Run Diffusion Model

python examples/offline_inference/text_to_image/text_to_image.py

Result:

[princepride]

@ZJY0516 @hsliuustc0106

[gcanlin]

Hi! Could we please wait for #774, which micro-refactors diffusion_worker to be hardware-agnostic? Then we don't need to modify the platform_utils and npu_worker. And the gpu word will be removed. I'd like to make #774 merged first, but it's also okay to merge this one first. I have some concerns about that #774 is becoming larger and larger.

[hsliuustc0106]

Hi! Could we please wait for #774, which micro-refactors diffusion_worker to be hardware-agnostic? Then we don't need to modify the platform_utils and npu_worker. And the gpu word will be removed. I'd like to make #774 merged first, but it's also okay to merge this one first. I have some concerns about that #774 is becoming larger and larger.

I think #774 may need more discussions

[ZJY0516]

LGTM

[Gaohan123]

Is its function similar to executor? Not now, but do we have plan to refractor it as executor in the future?

[princepride]

@ZJY0516 What do you think?

[hsliuustc0106]

any speed difference before and after this PR?

[princepride]

any speed difference before and after this PR?

I am testing it.

[princepride]

I use this script python examples/offline_inference/text_to_image/text_to_image.py compare the speed on H200, the average e2e time of original version is 15236ms, and the current version is 15233ms.

[Gaohan123]

LGTM. Thanks