add fp8 quantization kernels

[baodii]

add a) static_scaled_fp8_quant; b) dynamic_scaled_fp8_quant; c) dynamic_per_token_scaled_fp8_quant kernels.

[Copilot]

Pull Request Overview

This PR adds FP8 quantization kernels to the vLLM XPU implementation, providing three different quantization methods for FP8 (float8) data types. The implementation includes comprehensive test coverage and follows the existing kernel architecture pattern.

• Adds static, dynamic per-tensor, and dynamic per-token FP8 quantization kernels
• Provides comprehensive test suite with reference implementations for validation
• Extends the dispatch utilities to support FP8_e5m2 data type in addition to existing FP8_e4m3fn

Reviewed Changes

Copilot reviewed 10 out of 10 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
tests/test_fp8_quant.py	Main test file with reference implementations and parametrized test cases
tests/register_ops.py	Python bindings for the three new FP8 quantization operations
tests/ops/fp8_quant_op.py	High-level Python wrapper function for FP8 quantization operations
csrc/xpu/torch_bindings.cpp	PyTorch operator registration for the three FP8 quantization kernels
csrc/xpu/quantization/fp8/utils.h	Common utilities and type definitions for FP8 quantization
csrc/xpu/quantization/fp8/fp8_quant.h	Core FP8 quantization kernel implementations and helper functions
csrc/xpu/quantization/fp8/fp8_quant.cpp	C++ implementation of the three FP8 quantization functions
csrc/xpu/ops.h	Function declarations for the new FP8 quantization operations
csrc/xpu/dispatch_utils.h	Extended dispatch macros to include FP8_e5m2 data type
CMakeLists.txt	Build configuration update to include the new FP8 quantization source file

Comments suppressed due to low confidence (1)

csrc/xpu/quantization/fp8/fp8_quant.h:73

• Commented-out debug printf statements should be removed to improve code cleanliness and maintainability.

        static_cast<float>(input[i]), scale);

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[Copilot]

The comment mentions 'cuda kernel output' but this is for XPU implementation. It should say 'xpu kernel output' to be consistent with the context.

Suggested change

	# For fp8, in order to match the cuda kernel output, we have to do exactly
	# For fp8, in order to match the xpu kernel output, we have to do exactly

[Copilot]

Commented-out debug printf statements should be removed to improve code cleanliness and maintainability.

Suggested change

// sycl::ext::oneapi::experimental::printf("scaled_fp8_conversion: %f\n", r);

[Copilot]

Commented-out sys.path manipulation code should be removed if it's not needed, as it can be confusing and reduces code maintainability.

Suggested change

[zufangzhu]

Hi @baodii, can we merge this file with quant_utils.h into one?