Kernel

# pip install "kernels>=0.17.2" torch
# This loads and executes remote Python code; review the source first.
from kernels import get_kernel

attn = get_kernel("SZLHOLDINGS/YARQA-ATTN", version=1, trust_remote_code=True)

What it is not

Not a Fall 2026 ATELIER weight. No tensors in this kernel. The model twin SZLHOLDINGS/YARQA-ATTN keeps yarqa_attn.npz, a small SYNTHETIC attention-gate silhouette described by its TRAINING_RECEIPT.json; the kernel never loads it, and it carries no benchmark or capability claim. Not an alias of szl-receipt-attn. Not a pointer at the Triton trio (szl-receipt-attn, szl-maskmod, szl-block-kv); those three stay separate, and a11oy-net must not list this as a fourth Flash / Flex / paged stack. Do not list this next to Chaski, Qantu, Waman, Chakana, or Tinku.

We do not copy Dao hopper, Sage csrc, vLLM paged .cu, cuDNN FMHA, TRT cubins, CuTeDSL, or flex_attention.py.

Doctrine

Doctrine v11 LOCKED. Λ = Conjecture 1 (advisory; uniqueness OPEN; never a theorem). GitHub bytes are the artifact; the Hub is the publish mirror.

API

yarqa_attn(q, k, v, n_canals, chain=None): q,k,v are (batch, heads, seq, dim) on CPU. Sequence length is split into n_canals contiguous canals (earlier canals receive the remainder). Each canal is independent attention. Outputs are concatenated along seq.

ReceiptChain: SHA3-256. One receipt for partition boundaries, one for the attention-output digest. verify() returns (ok, depth, first_break).

canal_bounds(seq_len, n_canals): exclusive end-points [0, ..., seq_len].

Source tree (put torch-ext/ on PYTHONPATH):

import torch
from yarqa_attn import yarqa_attn, ReceiptChain, selfcheck, canal_bounds

q = k = v = torch.randn(1, 2, 16, 32)
chain = ReceiptChain()
y = yarqa_attn(q, k, v, 4, chain=chain)
print(canal_bounds(16, 4), chain.verify(), selfcheck())

selfcheck() never fabricates a pass. It runs a small CPU check: slice-and-attend vs a naive block-diagonal reference, receipt tamper detection, and that n_canals > 1 actually splits.

Source and Hub release scope

GitHub szl-holdings/YARQA-ATTN is the only writer of the Kernel Hub repository. Its hf-kernel-publish workflow builds build/torch-cpu and build/torch-universal (identical bytes) from torch-ext/yarqa_attn at one GitHub commit, writes them to the main and v1 branches, and then loads them anonymously with get_kernel("SZLHOLDINGS/YARQA-ATTN", version=1, trust_remote_code=True) in a fresh runner and compares the served files with the GitHub source. SZL_KERNEL_SOURCE.json in the Kernel Hub repository names the GitHub commit each build came from.

version=1 follows the v1 branch, the stable API channel. Loading with trust_remote_code=True executes code from the Kernel Hub repository; the narrowest form is trust_remote_code=["SZLHOLDINGS/YARQA-ATTN"]. To pin an exact build, resolve the v1 branch to its commit and pass that as revision:

from huggingface_hub import HfApi
from kernels import get_kernel

refs = HfApi().list_repo_refs("SZLHOLDINGS/YARQA-ATTN", repo_type="kernel")
pin = next(b.target_commit for b in refs.branches if b.name == "v1")
attn = get_kernel("SZLHOLDINGS/YARQA-ATTN", revision=pin, trust_remote_code=["SZLHOLDINGS/YARQA-ATTN"])

A GitHub source commit or a model-twin revision is not a Kernel Hub revision. The model twin carries only a pointer to the Kernel Hub repository. Publication does not qualify performance, GPU execution, the model-twin NPZ artifact or production use.

Publication context is recorded in hf/README.md.

Evidence record

MEASURED — Compartment correctness. tests/test_yarqa_attn.py asserts that fp32 output matches a naive within-compartment (block-diagonal) reference, and with one canal matches full SDPA, within atol=1e-5, rtol=1e-5; that more than one canal actually splits; that one canal's output does not depend on another canal's values; that earlier canals take the remainder; and that selfcheck() reports ok. CI runs it on every pull request and every push to main (cpu-tests.yml). Receipt.

MEASURED — Receipts. tests/test_receipt.py asserts one receipt for the partition and one for the output, that tampering breaks the chain at the first row, and that the digests change with the values and with the canal count. Receipt.

MEASURED — Immutable CPU get_kernel load from Kernel Hub revision 4e0828e517e84377d0024ab3f9d0cdab7d31520d using kernels==0.16.1 on 2026-10-05 UTC. selfcheck() passed. The three canonical Git object blobs at protected source cc895b47e7a3d1e7d5f246812a573dfd1205b103 matched all six provider copies across the CPU and universal build variants. Neither Hub twin exposes SZL_SOURCE_BINDING.json, and both provider cards remain behind the source-owned card. Source binding is UNAVAILABLE and card projection is SOURCE_AHEAD. No performance, GPU, provider-SLA or production claim is made. Receipt.

UNAVAILABLE — GPU cubins. None shipped; the 2026-08-28 session had no CUDA device. v0 refuses CUDA tensors, and the refusal test is skipped where no GPU exists. (no receipt linked).

NOT_CLAIMED — Throughput, tokens/s, joules, or performance versus FlashAttention. A speed claim needs a timed run on named hardware; none exists. (no receipt linked).

Labels follow the SZL claim language. A claim is only as strong as the receipt it links.

Source and provenance

Rendering this card does not establish Hub publication. Verify any Hub copy against its source revision, publication evidence, and exact Hub revision.