szl-khipu-kernels

Kernel-Hub card for the SZL KHIPU NumPy silhouettes.

CPU numpy path LIVE. CUDA UNAVAILABLE this session — that is honesty, not a missing bench.

from kernels import get_kernel

k = get_kernel(
    "SZLHOLDINGS/szl-khipu-kernels",
    revision="main",
    trust_remote_code=True,


<!-- SZL-ATELIER-CUT:v1:START -->
## The cut

A suite, not a zoo. One doctrine, several ops.

The cuDNN of governed inference — smaller, stricter, signed.

### Silhouette → leave → SZL

| Leader | Take, then tweak |
|---|---|
| Anthropic | Principles compiled to ops. |
| NVIDIA | Direct take: kernel packaging. |
| Unsloth | No. |

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

## Intended use

Import the ops. Do not pretender-load as transformers.

## Limitations

- Not a neural model.

Canonical GitHub: [`szl-holdings/szl-khipu`](https://github.com/szl-holdings/szl-khipu/blob/main/szl_khipu/)
<!-- SZL-ATELIER-CUT:v1:END -->

)
# CPU numpy path is LIVE.
# CUDA is UNAVAILABLE this session — that is honesty, not a missing bench.

print(k.CONJECTURE_1)
gate = k.lambda_gate([0.95] * 13)
print(gate["score"], gate["passed"], gate["proven_trust"])  # advisory; Conjecture 1 OPEN

import numpy as np
q = kv = np.random.default_rng(7).standard_normal((12, 4))
_out, _probs, leaked = k.yarqa_attn(q, kv, kv, n_canals=3)
print(leaked)  # bound ≤ 1e-9

CPU fallback when Kernel Hub is not in the environment (labeled; not a Hub load):

from szl_khipu import YUYAY_FLOORS, evaluate_lambda, yarqa_attn, lambda_gate

GitHub source: szl-holdings/szl-khipu.

Path	Status
CPU NumPy	LIVE
CUDA / Triton cubin	UNAVAILABLE
Energy	UNAVAILABLE — never a fabricated joule
proven_trust	false
Conjecture 1	OPEN
Not	FlashAttention, SageAttention, FlexAttention, vLLM

Apache-2.0. Copyright 2026 SZL Holdings. Doctrine v11 LOCKED.