szl-block-kv

KANCHAY · Doctrine v11 · Lean 749/14/163 · Λ = Conjecture 1 (advisory) · a-11-oy.com

Kernel sources are on this repo. CPU get_kernel import-LIVE is MEASURED. GPU UNAVAILABLE (no cubin + timed run). Not a model. Not listed next to Chaski or Qantu.

Canonical GitHub source for SZLHOLDINGS/szl-block-kv.

The cut

Blocked KV was invented for speed. We use the blocking to forget on purpose. Memory is a privilege.

A cache that cannot be subpoenaed for what it was never allowed to hold.

Silhouette → leave → SZL

Leader	Take, then tweak
Anthropic	No persistent unauthorized memory.
NVIDIA	PagedAttention / blocked KV — then we spend it on governance.
Unsloth	No.

Nobody else ships this combination. That is the point of a one-of-one.

Intended use

Governed decode memory.

Limitations

• Kernel. Not a drop-in vLLM replacement.

Canonical GitHub: szl-holdings/szl-khipu

Original SZL construction in the paged-KV category. Inspired by Kwon et al. PagedAttention SOSP 2023 https://arxiv.org/abs/2309.06180. NOT a rehost of vLLM or kernels-community/paged-attention. Distinct from a11oy MODELED H2O eviction.

Doctrine v11. Λ = Conjecture 1 OPEN (advisory; uniqueness unproven).

Status

STATUS: import-LIVE on CPU Kernel Hub get_kernel (kernels 0.16.1). Triton page kernel is UNAVAILABLE.

Thing	Label	Method / N / date / what-NOT
Kernel Hub get_kernel	import-LIVE	MEASURED 2026-08-28 2:29pm ET on kernels 0.16.1. HEAD d3ede3e (d3ede3e471b51080492b1c69306283507dcf507e). Legal name szl-block-kv (Python module szl_block_kv). Variants: build/torch-universal (default get_kernel) and build/torch-cpu (backend="cpu"). Working calls: get_kernel("SZLHOLDINGS/szl-block-kv", revision="main", trust_remote_code=True) and the same with backend="cpu". selfcheck ok. max_abs_vs_contiguous=2.38e-07 (full 2.384185791015625e-07), path=torch_gather, chain_ok=true. What-NOT: no tokens/s; no joules.
Triton page kernel	UNAVAILABLE	MEASURED 2026-08-28 7:01pm ET this session. Host cursor (Linux 6.12.94+ x86_64, Intel Xeon 8-core). torch 2.13.0+cu130 compiled CUDA 13.0. torch.cuda.is_available()=false. nvidia-smi UNAVAILABLE. device_count=0. Triton 3.7.1 present with no CUDA device. No cubin. No timed GPU run. No tokens/s. No joules.

v0 is a labeled torch gather over a block table. GPU paged Triton is UNAVAILABLE. No speedup claim.

Load

from kernels import get_kernel

kv = get_kernel("SZLHOLDINGS/szl-block-kv", revision="main", trust_remote_code=True)
print(kv.selfcheck())

Local checkout:

from szl_block_kv import PagedCache, paged_attn, reshape_and_cache, selfcheck
print(selfcheck())

Correctness (documented): paged gather matches contiguous KV SDPA within atol/rtol 1e-5 on float32. Skip CUDA Triton — that kernel is not in v0.

Apache-2.0. Copyright 2026 SZL Holdings.