support fp8 grouped gemm

[mayuyuace]

Support wfp8a16.
And only support per tensor scale.

Now cutlass-sycl uses inefficient data type convert function from fp8 to other dtype:
https://github.com/intel/sycl-tla/blob/5ac97000432951b2812ade99235fbf539958c5d4/include/cutlass/float8.h#L336

As the picture shows, cuda will uses hardware instruction.
So the fp8 grouped gemm kernel has low performance.
Cutlass needs provide more efficient convert function.

Update:
Optimize dequantize fp8 by replacing native dtype convert function:
For example when m=1, n 5120, k=8192, e=16, topk=1 on PVC:
fp16 @ e5m2: 6.321ms --> 1.358ms
bf16 @ e5m2: 6.319ms --> 3.931ms
fp16 @ e4m3: 7.007ms --> 2.648ms
bf16 @ e4m3: 7.109ms --> 5.072ms

[Copilot]

Pull Request Overview

This PR adds support for FP8 (float8) data types in grouped GEMM operations, specifically enabling weight quantization with FP8 formats (e4m3 and e5m2) while keeping activations in FP16/BF16. The implementation introduces a scaling parameter for dequantization of FP8 weights during computation.

Key Changes:

• Added scale_B parameter throughout the grouped GEMM call chain for FP8 weight dequantization
• Implemented four new FP8 policy classes supporting e4m3/e5m2 formats with fp16/bf16 activations
• Added comprehensive test coverage for FP8 grouped GEMM operations

Reviewed Changes

Copilot reviewed 9 out of 9 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
vllm_xpu_kernels/fused_moe_interface.py	Added scale_B parameter to cutlass_grouped_gemm and xpu_fused_moe function signatures
tests/fused_moe/test_fused_moe.py	Added new test_grouped_gemm_fp8 test case and updated existing tests to pass None for scale_B
csrc/xpu/torch_bindings.cpp	Updated PyTorch binding signature to include optional ptr_B_scale parameter
csrc/xpu/ops.h	Added ptr_B_scale parameter to cutlass_grouped_gemm function declaration
csrc/xpu/cutlass_kernels/grouped_gemm_kernel.cpp	Integrated scale_B parameter into kernel execution and added FP8 policy template instantiations
csrc/xpu/cutlass_kernels/grouped_gemm.hpp	Added FP8 data type detection and routing logic with new policy-based kernel dispatching
csrc/xpu/cutlass_kernels/collective/gemm/moe_gemm_array_cooperative.hpp	Updated to handle scale tensor in mainloop tensors tuple
csrc/xpu/cutlass_kernels/collective/gemm/moe_dtype_policy.hpp	Added four new FP8 policy classes and GmemTiledCopy definitions for all policies
csrc/xpu/cutlass_kernels/collective/gemm/moe_array_mma.hpp	Implemented FP8 conversion logic and scale application in the MMA collective

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[mayuyuace]

@Liangliang-Ma Please help review this.

[mayuyuace]

@jikunshang @baodii

[jikunshang]

I think we can on hold this now that we will upgrade sycl-tla v0.6 which may break current fp16/bf16 moe gemm..

[mayuyuace]

I think we can on hold this now that we will upgrade sycl-tla v0.6 which may break current fp16/bf16 moe gemm..

Sure that this PR can be merged after sycl-tla is upgraded.