feat: add Intel GPU (XPU) support via auto-round-lib dispatch

[luoyu-intel]

Summary

Adds XPU support to SageAttention by detecting Intel GPUs and dispatching sageattn() and sageattn_varlen() to the ARK kernel backend (auto-round-lib). No CUDA compilation is required on XPU systems. The dispatch is transparent — users call sageattn() as usual and it automatically routes to the correct backend.

Changes

3 files, 85 insertions, 10 deletions

File	Change
setup.py	Auto-detect XPU → skip CUDA compilation, print install hint for auto-round-lib
sageattention/core.py	sageattn(): XPU dispatch with head_dim auto-padding (<64→64, 64-128→128), bool mask → FP32 conversion. sageattn_varlen(): XPU dispatch with head_dim auto-padding. CUDA quant imports guarded with try/except for XPU-only environments.
README.md	Add Intel GPU (XPU) installation section and API documentation.

Design Decisions

1. No CUDA kernel porting — XPU uses auto-round-lib (ARK) as an external dependency, installed via pip install auto-round-lib. No SYCL code is added to this repository.
2. Zero API breakage — sageattn(q, k, v, ...) works identically on CUDA and XPU. The same parameters are accepted.
3. head_dim auto-padding — Models using head_dim=96 (e.g., some diffusion models) are automatically padded to 128 before dispatching to ARK, matching upstream behavior.
4. bool mask support — torch.bool attention masks are auto-converted to FP32 additive masks for ARK compatibility.

Test Results

Same as ARK PR above (the fork routes to ARK, so performance is identical). Verified on Intel Arc Pro B60:

sageattn:            [1, 16, 4096, 128]
sageattn (hd=96):    [1, 16, 4096, 96]       # auto-padding
sageattn return_lse: O=[1, 16, 4096, 128], LSE=[1, 16, 4096]
sageattn_varlen:     [4096, 16, 128]

Installation (XPU)

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/xpu
pip install auto-round-lib
git clone https://github.com/luoyu-intel/SageAttention.git # for this PR
git clone https://github.com/thu-ml/SageAttention.git # after merge
cd SageAttention
python setup.py install   # CUDA compilation skipped automatically

Performance (TOPS)

FLOPs = 4 × B × Hq × Sq × Sk × D (non-causal attention) · GPU: Intel Arc Pro B60 (BMG-G21) · FP16, non-causal
BMG theoretical peak: ~197 TOPS (INT8) / ~98 TFLOPS (FP16).

Config	torch.sdpa	ARK sdpa	ARK sagev1	sagev1/ARK	sagev1/torch
Sq=4096 Sk=4096 H=16	56.2 TOPS	71.0 TOPS	74.6 TOPS	1.05×	1.33×
Sq=8192 Sk=8192 H=16	57.5 TOPS	81.1 TOPS	91.0 TOPS	1.12×	1.58×
Sq=16384 Sk=16384 H=16	58.4 TOPS	82.0 TOPS	96.0 TOPS	1.17×	1.64×
Sq=4096 Sk=4096 H=96(8) GQA	57.3 TOPS	81.3 TOPS	92.3 TOPS	1.13×	1.61×
Sq=4096 Sk=8192 H=96(8) GQA	57.9 TOPS	82.1 TOPS	96.1 TOPS	1.17×	1.66×
Sq=4096 Sk=16384 H=96(8) GQA	58.0 TOPS	82.0 TOPS	98.1 TOPS	1.20×	1.69×
Sq=4096 Sk=32768 H=96(8) GQA	56.5 TOPS	81.9 TOPS	98.9 TOPS	1.21×	1.75×