TransformerEngine 2.11.0

TransformerEngine 2.11.0+c188b533

Built with:

• Python 3.12
• PyTorch 2.10.0
• CUDA 13.0
• CUDA architectures: sm_75, sm_80, sm_89, sm_90, sm_100, sm_120

TransformerEngine 2.10.0

TransformerEngine 2.10.0+769ed778

Built with:

• Python 3.12
• PyTorch 2.10.0
• CUDA 12.9
• CUDA architectures: sm_70, sm_80, sm_89, sm_90, sm_100, sm_120

TileLang 0.1.10 + #2303 (CUDA 13.1, PyTorch 2.10)

TileLang built from stock upstream tile-ai/tilelang at commit a8d93798 (post-v0.1.10), which includes PR #2303 ([CUDA] Support preferred copy instruction lowering, d0937562) — i.e. the T.copy(..., prefer_instruction="tma") API. The released PyPI tilelang==0.1.10 predates #2303 and lacks it.

• Source: tile-ai/tilelang@a8d93798 (unmodified upstream)
• Build: CUDA 13.1, for PyTorch 2.10 (cu129 runtime); cp38-abi3 → installs on CPython ≥3.10
• Why hosted: consumed by xorl's FlashQLA GDN backend (XORL_GDN_BACKEND=flashqla). FlashQLA needs both the tl_gemm builtin (in stock ≥0.1.10) and prefer_instruction="tma" (#2303). The fast gemm_v1 path is re-added in-repo by xorl's tilelang_gemm_v1 shim (no tilelang source fork); this wheel is plain upstream.

Replace with PyPI tilelang>=0.1.11 once a release carrying #2303 ships.

Mamba SSM 2.3.1 + Causal Conv1d 1.6.1 (CUDA 12.9, PyTorch 2.10)

Pre-compiled wheels for mamba-ssm 2.3.1 and causal-conv1d 1.6.1.

• Python 3.12, Linux x86_64
• Built against PyTorch 2.10.0+cu129, CUDA 12.9
• Source: https://github.com/state-spaces/mamba

FlashAttention 3.0.0b1

FlashAttention-3 v3.0.0b1

• Requires CUDA 12.3+
• Python 3.9+ (stable ABI)
• Hopper (SM90) optimized

DeepGEMM 2.3.0

DeepGEMM wheel for CUDA 13, PyTorch 2.10, Python 3.12

DeepEP 1.2.1 (CUDA 13.0)

DeepEP 1.2.1+567632d

Built with:

• Python 3.12
• PyTorch 2.11.0
• CUDA 13.0

DeepEP 1.2.1

Pre-built DeepEP wheel (commit 567632d, torch 2.9, CUDA 12, Python 3.12)