SageAttention 2.2 SM75/Turing Backend — Torch 2.12 + CUDA 13.0 — Linux

SageAttention 2.2-derived SM75 / NVIDIA Turing Backend

This release provides SageAttention 2.2-derived attention backends for
NVIDIA Turing / Compute Capability 7.5 (SM75) GPUs.

Linux and Windows wheels are platform-specific. Install only the wheel that
matches your operating system, Python version, PyTorch version, and CUDA stack.

What this build is

This is NOT an unmodified upstream SageAttention 2.2 build.

The SM75/Turing backend combines:

• SageAttention 2.2 package/API and dispatcher
• SageAttention 2.2 native INT8 preprocessing
• SM75 BLOCK_SIZE=32 / BLKK32 preprocessing support
• Proven SageAttention 2.1.2-derived SM75 qattn CUDA kernel
• Public sageattn() SM75 dispatch integration
• FP16 and BF16 support
• Safe PyTorch SDPA fallback for attention-mask cases that are not handled
  by the native SM75 fast path

A separate SageAttention 2.1.2 runtime installation is NOT required.

Linux

Verified environment

• Linux x86-64
• CPython 3.12
• PyTorch 2.12.0+cu130
• PyTorch CUDA: CUDA 13.0 (cu130)
• Validated GPU: NVIDIA Tesla T4
• Compute Capability: 7.5 (SM75)

Verified validation

The Linux SM75 build has undergone extensive runtime and compatibility testing
on an NVIDIA Tesla T4.

Verified:

• Native SM75 runtime: PASS
• Public SageAttention API: PASS
• FP16: PASS
• BF16: PASS
• HND and NHD tensor layouts: PASS
• Causal and non-causal attention: PASS
• Head dimensions 64 and 128: PASS
• Sequence-length boundary testing: PASS
• GQA / MQA: PASS
• Cross-attention: PASS
• return_lse: PASS
• Custom attention scale: PASS
• CUDA graph capture: PASS
• PyTorch autocast FP16/BF16: PASS
• Two CUDA streams: PASS
• Two-GPU alternating-device testing: PASS
• Long sequence N=16384: PASS
• 500-call FP16/BF16 stress testing: PASS
• Boolean attention masks: PASS
• Additive floating-point attention masks: PASS
• Float32 additive masks with FP16/BF16 Q/K/V: PASS
• Normal unmasked native SM75 path regression: PASS

The final attention-mask validation matched the PyTorch SDPA reference exactly
for the tested boolean and additive-mask configurations.

Performance

Earlier SM75 attention microbenchmarks measured approximately:

• 1.36x to 1.95x speedup over PyTorch SDPA in tested FP16 workloads
• ~1.87x geometric-mean speedup in the original representative benchmark set

These are attention-kernel microbenchmarks and should NOT be interpreted as
equivalent end-to-end application speedups.

Numerical accuracy

The native SM75 Sage attention path has shown approximately 1.2% or less
relative mean error versus PyTorch SDPA in the tested workloads.

Attention-mask calls that require the compatibility fallback use PyTorch SDPA
for correctness.

Known SM75 limitations

• qk_quant_gran="per_thread" is not supported by the native SM75 qattn path.
  The verified native Turing path uses per_warp.
• Native FP8 attention is not supported on SM75/Turing.
• Direct Triton attention-mask kernels are not supported on SM75.
  Public masked-attention calls use the correctness-first PyTorch SDPA
  fallback where required.

These limitations do not affect the verified normal FP16/BF16 SM75
per_warp attention path.

Linux compatibility

Treat the Linux wheel as verified only for:

Linux x86-64

• CPython 3.12
• PyTorch 2.12.0+cu130
• CUDA 13
• NVIDIA SM75/Turing

Compatibility with other Python versions, PyTorch versions, CUDA versions,
operating systems, or GPU architectures should NOT be assumed unless
separately tested.

Linux wheel

sageattention-2.2.0+sm75.torch2.12cu130-cp312-cp312-linux_x86_64.whl

Windows

Windows wheels are separate builds and are NOT interchangeable with the Linux
wheel.

Use only a Windows wheel whose filename explicitly ends in:

-cp312-cp312-win_amd64.whl

The Windows build should be treated as verified only for the exact Python,
PyTorch, CUDA, and NVIDIA GPU configuration stated with that wheel.

Do NOT install the Linux x86_64 wheel on Windows, and do NOT assume that Linux
T4 validation automatically validates a Windows build.

Windows compatibility and validation results should be listed separately for
each Windows wheel once that wheel has completed its own runtime testing.

Important

This is a custom SM75/Turing compatibility build derived from SageAttention
2.2 and is not an official unmodified upstream SageAttention 2.2 release.

For the tested Linux configuration above, the SM75 backend has passed the
current correctness, compatibility, attention-mask, stress, and regression
validation suite.

Causal Conv1d & Mamba SSM for Windows using Torch 2.7.1, Cuda12.8, Python 3.10

Causal Conv1d & Mamba SSM for Windows using Torch 2.7.1, Cuda 12.8 & Python 3.10

Thanks to sdbds for the Code to build these .whl files for Windows.