feat: Add Windows support with triton-windows and PyTorch fallback

[tori29umai0123]

Summary

Add Windows support for AngelSlim with FP8 Triton kernels using triton-windows.

Changes

New Files

• angelslim/compressor/_platform.py - Platform detection and automatic backend selection
• angelslim/compressor/quant/core/quant_func_torch.py - PyTorch fallback for weight quantization
• angelslim/compressor/diffusion/kernels/python/quantizers/fp8_per_block_torch.py - PyTorch fallback for FP8
  per-block quantization
• angelslim/compressor/diffusion/kernels/python/quantizers/fp8_per_token_group_torch.py - PyTorch fallback for FP8
  per-token-group quantization
• angelslim/compressor/diffusion/kernels/python/gemm/fp8_gemm_torch.py - PyTorch fallback for FP8 GEMM

Modified Files

• requirements/requirements.txt - Platform-specific triton dependency (triton for Linux, triton-windows for Windows)
• setup.py - Version string includes CUDA and PyTorch version (e.g., 0.0.0_dev+cu128.torch2.10)
• angelslim/compressor/diffusion/cache/taylorcache_helper.py - Conditional torch.compile (disabled on Windows)
• angelslim/compressor/quant/core/quant_func.py - Lazy Triton import with automatic backend selection
• angelslim/compressor/diffusion/quant/quant_func.py - Remove CUDA requirement for per-token-group quantization
• angelslim/compressor/diffusion/kernels/python/quantizers/__init__.py - Conditional imports based on backend
• angelslim/compressor/diffusion/kernels/python/gemm/__init__.py - Conditional imports based on backend
• README.md - Windows build instructions

Features

• Automatic Backend Selection: Automatically detects if Triton is available and functional; falls back to PyTorch
  if not
• Environment Variables:
  • ANGELSLIM_BACKEND=pytorch|triton - Force backend selection
  • ANGELSLIM_TORCH_COMPILE=0|1 - Enable/disable torch.compile
• Cross-Platform: Works on both Linux (with triton) and Windows (with triton-windows or PyTorch fallback)

Test

python -c "import torch; from angelslim.compressor.diffusion.kernels.python.quantizers import
fp8_per_block_quant_triton; from angelslim.compressor.diffusion.kernels.python.gemm import fp8_gemm_triton_block;
a,b=torch.randn(128,256,device='cuda'),torch.randn(512,256,device='cuda'); aq,a_s=fp8_per_block_quant_triton(a);
bq,b_s=fp8_per_block_quant_triton(b); c=fp8_gemm_triton_block(aq,a_s,bq,b_s); print(f'FP8 GEMM OK: {c.shape},
{c.dtype}')"

Expected output:
FP8 GEMM OK: torch.Size([128, 512]), torch.bfloat16

Tested Environment

- Windows 11 + NVIDIA RTX A6000
- CUDA 12.8 + PyTorch 2.10.0 + triton-windows

[yghstill]

Please mv to docs/source/getting_started/installation.md