Spaces:

---

kernels-community

/

README

like 5

Running

We are reopening this request with the corrected organization-owned publishing flow.

Ownership and provenance

• Target kernel repository: Efficient-Large-Model/Sol-Attn
• Canonical NVIDIA source: NVlabs/Sana@8a26fb0
• Reviewable Kernel Builder source: Efficient-Large-Model/Sol-Attn-Kernel-Source@8e9f35a

The kernel implementation remains maintained under NVIDIA's NVlabs/Sana. The Kernel Builder source is owned by the Efficient-Large-Model Hub organization. No personal GitHub fork is involved, and no files or workflows are added to NVlabs/Sana.

Publishing flow

We use the official HF Jobs/Kernel Builder flow directly:

1. Run the pinned Nix/Cachix image used by huggingface/kernel-builder-job.
2. Clone the exact organization-owned Builder source commit above.
3. Run kernel-builder check-config.
4. Run kernel-builder build-and-upload --repo-id Efficient-Large-Model/Sol-Attn.

The Job namespace is explicitly Efficient-Large-Model, and HF_TOKEN is injected only as a Job secret.

Validation evidence

• The packaged tree is verified against the pinned NVlabs/Sana commit: all 51 implementation files match, with only semantically equivalent package-relative internal imports in 17 Python files for Kernel Hub's version-isolated loader.
• H100 full exact sink vs. PyTorch SDPA: max absolute error 0.001953125, cosine similarity 0.999997854.
• H100 architecture-selected sparse CuTe path vs. the Triton Sol-Attn reference: max absolute error 0.0009765625, cosine similarity 1.0.

Platform access needed

Could you please enable kernel creation for the Efficient-Large-Model organization and help us run the publishing Job? The Hub currently rejects Job creation before scheduling with HTTP 402: Pre-paid credit balance is insufficient. No build Job was created and no compute was consumed.