[RFC]: Make device agnostic for diverse hardware support

[wangshuai09]

Motivation.

vLLM has already been adapted to many hardware devices, such as GPU, TPU, and XPU. However, adapting these backends requires implementing separate Worker/Executor/Model Runner frameworks for each, which leads to code redundancy and maintenance difficulties.
In fact, these hardware framework codes can be abstracted at the device layer, forming a unified framework. This way, only one set of code would need to be maintained, and different backends would only need to implement the device layer interfaces and any device-specific logic if necessary.
I also found some new features are only updated on GPU-related codes. In fact, these codes are also applicable to other hardware, but it is difficult for other hardware to perceive and follow these updates.

Proposed Change.

This RFC is intended to establish a unified framework.
Maybe there will be diffuculty for intergrating hardware framework to common framework, It makes sense to work towards this direction, the diagram below represents a proposed solution:

Taking Executor as example, for third-party hardware devices based on the pytorch ecosystem, the basic interfaces of torch have been well adapted, so after abstracting the device-related hard coding, such as torch.cuda, torch.xpu, GPU Executor could be used as the Common Executor of all third-party devices.

Following #6080, different hardware backends can put their own device-specific code in NewBackendPlatform, so that the framework can be device-agnostic through current_platform. For example, torch.cuda.synchronize could use current_platform.synchronize.

Feedback Period.

To realize this idea will involve more files, so the following steps are currently sorted out to finally achieve the above purpose：

• BackendPlatform
  • Neuron
  • Openvino
• Backend Type Check
  • is_cpu -> current_platform.is_cpu
  • is_xpu -> current_platform.is_xpu
  • is_openvino -> current_platform.is_openvino
  • is_neuron -> current_platform.is_neuron
• Backend Releated Func
  • seed_everything -> current_platform.seed_everything
  • is_pin_memory_available -> current_platform.is_pin_memory_available
  • DeviceMemoryProfiler -> current_platform.memory_profiler
  • wrap_device -> current_platform.wrap_device
• Backend Releated Hard Coding
  • torch.xxx.get_device_name -> current_platform.get_device
  • torch.xxx.Event -> current_platform.Event
  • torch.xxx.synchronize -> current_platform.synchronize
  • torch.xxx.Stream -> current_platform.Stream
  • torch.xxx.stream -> current_platform.stream
  • torch.xxx.empty_cache -> current_platform.empty_cache
  • torch.xxx.device_count -> current_platform.device_count
  • torch.xxx.memory_allocated -> current_platform.memroy_allocated
  • torch.xxx.set_device -> current_paltform.set_device
  • torch.xxx.current_device -> current_platform.current_device
  • torch.xxx.get_device_capability -> current_platform.get_device_capability
• Try to unify hardware framework, cpu releated framework may have problem to intergrate.
  • gpu(neuron,openvino,tpu,xpu,..)_executor -> common_backend_executor
  • gpu(neuron,openvino,tpu,xpu,..)_worker -> common_backend_worker
  • gpu(neuron,openvino,tpu,xpu,..)_model_runner -> common_backend_model_runner

There must be omissions or difficulties in actual implementation here, keep updating.

CC List.

@youkaichao @WoosukKwon

Any Other Things.

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[youkaichao]

we can do it step by step.

is_cpu -> current_platform.is_cpu
is_xpu -> current_platform.is_xpu
is_openvino -> current_platform.is_openvino
is_neuron -> current_platform.is_neuron

this can be the first step, and should be easy to do.

the rest might need some case-by-case discussion.