szl-maskmod

Original SZL score_mod + block-mask attention (FlexAttention silhouette). Inspired by Dong et al. https://arxiv.org/abs/2412.05496. Not PyTorch flex_attention.py. No CuTeDSL. No speedup claim. Doctrine v11. Λ = Conjecture 1 OPEN. Apache-2.0. Copyright 2026 SZL Holdings.

Pure-Python (noarch) kernel published from GitHub szl-holdings/szl-maskmod by its hf-kernel-publish workflow. GitHub is the source of truth; this kernel repository is written only by that workflow, and this card is published from CARD.md there.

Load it

# pip install "kernels>=0.17.2" torch
from kernels import get_kernel

km = get_kernel("SZLHOLDINGS/szl-maskmod", version=1, trust_remote_code=True)

• version=1 resolves the v1 branch, the stable API channel. It is republished on every release. The kernels library refuses a call that names neither a version nor a revision.
• trust_remote_code is required: SZLHOLDINGS is not a trusted kernel publisher on the Hub. The narrowest form is trust_remote_code=["SZLHOLDINGS/szl-maskmod"].
• CPU hosts load build/torch-cpu; CUDA and other hosts load build/torch-universal. Both variants carry identical code, built from the same GitHub commit.

Pin an exact build (production)

from huggingface_hub import HfApi
from kernels import get_kernel

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

Record pin (a commit of this kernel repository) in your lockfile. SZL_KERNEL_SOURCE.json in this repository names the GitHub commit each build came from. Commits of the model repository huggingface.co/SZLHOLDINGS/szl-maskmod are not kernel revisions and do not resolve here.

Use

import torch

q = k = v = torch.randn(1, 2, 8, 16)
chain = km.ReceiptChain()
keep = km.create_dense_mask(8, 8, causal=True, sliding_window=4)
y = km.maskmod_attn(q, k, v, block_mask=keep, chain=chain)
print(km.selfcheck())

Each receipt binds the block-mask topology (shape and sha256 of the keep-mask), the score_mod name and the causal flag into a SHA3-256 hash chain. Correctness is checked against SDPA on CPU; no CUDA benchmark, tokens/s or energy figure is claimed.