Metadata-Version: 2.4
Name: tokenspeed-cutedsl-kda
Version: 0.1.0.post20260919
Summary: AOT CuteDSL KDA prefill kernels for TokenSpeed
License-Expression: LicenseRef-Proprietary
Project-URL: Homepage, https://github.com/lightseekorg/tokenspeed
Classifier: Development Status :: 2 - Pre-Alpha
Classifier: Operating System :: POSIX :: Linux
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3 :: Only
Requires-Python: >=3.10
Description-Content-Type: text/markdown
License-File: LICENSE
License-File: THIRD_PARTY_NOTICES.md
Requires-Dist: apache-tvm-ffi==0.1.13.post3
Requires-Dist: cuda-python<14,>=13
Requires-Dist: nvidia-cutlass-dsl[cu13]==4.7.1
Requires-Dist: torch==2.14.0
Dynamic: license-file

# TokenSpeed CuteDSL KDA

The current package requires PyTorch 2.14.0 and CUDA 13. Its AOT payloads
retain their original hashes and TVM-FFI ABI; the runtime dependency update
does not rebuild or change the kernel math.

The release workflow checks both host architectures, then compares GPU
outputs and final states with the previous Torch 2.13.0 package and verifies
CUDA Graph replay. Publication requires these checks to pass on `main`.
Historical payload validation in `PROVENANCE.json` does not establish
validation of the new runtime.

Dependency validation reports one narrowly checked upstream packaging exception:
NVIDIA's ARM64 `nvidia-cusparselt-cu13==0.8.1` wheel declares an internal `sbsa`
tag. The workflow accepts only that sole `pip check` diagnostic after verifying
Linux AArch64, the exact version/tag, the official library payload hash, its ELF
architecture and successful loading. Other dependency failures block release;
this exception does not establish GPU runtime compatibility.
