SZLHOLDINGS
/

YARQA-ATTN

@@ -1,107 +1,130 @@
----
-thumbnail: https://huggingface.co/SZLHOLDINGS/YARQA-ATTN/resolve/main/og-card.png
-language:
-- code
-license: apache-2.0
-library_name: kernels
-tags:
-  - kernel
-  - attention
-  - yarqa
-  - governed-ai
-  - szl-holdings
-  - kernel-lane
-szl:
-  owner: KERNEL
-  not_a_weight: true
-  not_an_alias: true
-  collection: none
-  python: present
-  import_live: true
-  gpu: UNAVAILABLE
----
-# YARQA-ATTN
-<p align="center">
-  <img src="og-card.png" alt="YARQA-ATTN" width="100%"/>
-</p>
-`KANCHAY` · Doctrine v11 · Lean `749/14/163` · Λ = Conjecture 1 (advisory) · [a-11-oy.com](https://a-11-oy.com)
-Python kernel is on this repo. CPU `get_kernel` import-LIVE MEASURED (`7e533ce`). GPU cubins **UNAVAILABLE** (not ROADMAP). Not an alias of `szl-receipt-attn`. Not a fourth Flash / Flex / paged stack.
-<!-- SZL-KERNEL-STATUS:import-LIVE:START -->
-## Status
-<!-- SZL-ATELIER-CUT:v1:START -->
-## The cut
-FlashAttention is faster. YARQA is accountable. We steal the kernel discipline from NVIDIA and spend it on provenance, not FLOPs.
-An attention op whose softmax support is reconstructable from a signed log.
-### Silhouette → leave → SZL
-| Leader | Take, then tweak |
-|---|---|
-| Anthropic | Interpretability as a runtime artifact. |
-| NVIDIA | cuDNN / FlashAttention silhouette — then we add the receipt. |
-| Unsloth | Unrelated. Don't wrap this in FastLanguageModel. |
-Nobody else ships this combination. That is the point of a one-of-one.
-## Intended use
-Drop-in attention with an audit tape.
-## Limitations
-- Not a checkpoint.
-- Performance vs FlashAttention is not claimed.
-Canonical GitHub: [`szl-holdings/szl-khipu`](https://github.com/szl-holdings/szl-khipu/blob/main/szl_khipu/yarqa.py)
-<!-- SZL-ATELIER-CUT:v1:END -->
-> **STATUS: import-LIVE** on CPU Kernel Hub `get_kernel` (kernels `0.16.1`). GPU cubins **UNAVAILABLE** this session (not ROADMAP).
-| Thing | Label | Method / N / date / what-NOT |
-|---|---|---|
-| Kernel Hub `get_kernel` | **import-LIVE** | MEASURED 2026-08-28 3:08pm ET on kernels `0.16.1`. Package HEAD [`7e533ce`](https://huggingface.co/kernels/SZLHOLDINGS/YARQA-ATTN/commit/7e533ce702029061bc68f9f9cafe88efdd7f5f00) (`7e533ce702029061bc68f9f9cafe88efdd7f5f00`). README at MEASURE [`2871b3c`](https://huggingface.co/kernels/SZLHOLDINGS/YARQA-ATTN/commit/2871b3cbd73ee05e9f1aa010b75b190683f0ecd3). Legal name `yarqa-attn` (Python module `yarqa_attn`). Variants: `build/torch-universal` (default `get_kernel`) and `build/torch-cpu` (`backend="cpu"`). Working calls: `get_kernel("SZLHOLDINGS/YARQA-ATTN", revision="main", trust_remote_code=True)` and the same with `backend="cpu"`. `selfcheck` **ok**. `max_abs_vs_compartment_ref=3.58e-07` (full `3.5762786865234375e-07`), `path=torch_compartment`. What-NOT: no tokens/s; no joules; not a fourth Flash / Flex / paged stack. Lambda = Conjecture 1 (advisory). |
-| GPU cubins | **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 shipped. No tokens/s. No joules. CPU import-LIVE unchanged. Not a fourth Flash / Flex / paged stack. Lab stays Khipu. |
-<!-- SZL-KERNEL-STATUS:import-LIVE:END -->
-KERNEL kernel card. Original SZL **compartment / plug-flow** attention cut. Receipt-aware. Honesty-labeled.
-**Not a Fall 2026 ATELIER weight.** No tensors in this repo. Not an alias of [`szl-receipt-attn`](https://github.com/szl-holdings/szl-receipt-attn). Not a pointer at the Triton trio (`szl-receipt-attn`, `szl-maskmod`, `szl-block-kv`). Those three stay separate. a11oy-net does not list this as a fourth Flash / Flex / paged stack.
-GitHub is source of truth: [`szl-holdings/YARQA-ATTN`](https://github.com/szl-holdings/YARQA-ATTN). KERNEL binds Hub bytes from that tree. Do not PUT an empty card.
-| | |
-|---|---|
-| **Owner** | KERNEL |
-| **Artifact** | kernel (Python present; no weights; GPU cubins not claimed) |
-| **Status** | **import-LIVE** CPU · GPU cubins **UNAVAILABLE** |
-| **License** | Apache-2.0 |
-| **Λ** | Conjecture 1 (advisory, never a theorem) |
-| **Path** | `torch_compartment` (CPU) |
-| **Serve studio** | not this repo. Live CPU lab is [`szl-model-inference-lab`](https://huggingface.co/spaces/SZLHOLDINGS/szl-model-inference-lab) (Khipu GGUF only) |
-Silhouette: partition a sequence into canals (contiguous compartments), attend within a canal, emit SHA3-256 of the partition and of the attention output. We do not copy Dao hopper, Sage `csrc`, vLLM paged `.cu`, cuDNN FMHA, TRT cubins, CuTeDSL, or `flex_attention.py`. Metaphor only vs [`szl-holdings/yarqa`](https://github.com/szl-holdings/yarqa) (CFD; different product). Throughput is MEASURED only from a timed run on named hardware. Until then every speed claim is unstamped. No tokens/s. No joules.
-Do not list this next to Chaski, Qantu, Waman, Chakana, or Tinku.
-## Load
-```python
-from kernels import get_kernel
-attn = get_kernel("SZLHOLDINGS/YARQA-ATTN", revision="main", trust_remote_code=True)
-```
-Fashion GO 2026-08-28 3:10pm ET. import-LIVE CPU stays. GPU cubins stamped **UNAVAILABLE** 2026-08-28 7:01pm ET (no CUDA device this session). Not a fourth Flash / Flex / paged stack.
-Apache-2.0. Copyright 2026 SZL Holdings.

+---
+thumbnail: https://huggingface.co/SZLHOLDINGS/YARQA-ATTN/resolve/main/og-card.png
+language:
+- code
+license: apache-2.0
+library_name: kernels
+tags:
+  - kernel
+  - attention
+  - yarqa
+  - governed-ai
+  - szl-holdings
+  - kernel-lane
+szl:
+  owner: KERNEL
+  not_a_weight: true
+  not_an_alias: true
+  collection: none
+  python: present
+  import_live: true
+  gpu: UNAVAILABLE
+---
+<!-- SZL-KERNEL-OPERATIONAL:START -->
+## Operational (MEASURED laptop-Blackwell)
+> **STATUS:** tests **PASS**. `get_kernel` **import-LIVE**. Unsloth/LoRA is the wrong tool. Receipted kernels, not silent CUDA.
+| Thing | Label | Method / N / date / what-NOT |
+|---|---|---|
+| tests (`PYTHONPATH=torch-ext`) | **PASS** | MEASURED 2026-08-29T15:54:08Z host `betterwithage` Windows-10-10.0.26200-SP0. torch `2.10.0+cu128`. GPU `NVIDIA GeForce RTX 5050 Laptop GPU` arch `Blackwell`. pytest `25 passed, 1 skipped in 0.33s`. Failed nodes: `none`. What-NOT: not a leaderboard. torch.compile fullgraph failures on Windows Blackwell (`cl is not found`) are MEASURED, not hidden. |
+| Kernel Hub `get_kernel` | **import-LIVE** | kernels `0.16.1`. Default: `get_kernel("SZLHOLDINGS/YARQA-ATTN", revision="main", trust_remote_code=True)` → `True`. `backend="cpu"` → `True`. trust_remote_code=False → `ValueError` (SZLHOLDINGS is not a trusted publisher). repo_type=kernel required (kernels 0.16). What-NOT: not a weight load; do not pickle/joblib.load. |
+| formula-tax | **ADVISORY** | locked-8 `F1 F4 F7 F11 F12 F18 F19 F22`. registry_count=21. Λ geomean `1.0`. uniqueness **Conjecture 1** (never a theorem). |
+| I1–I8 | **catalog** | `I1 receipt-chain-continuity; I2 ledger-failure-shape; I3 served-run-has-model; I4 signed-columns-atomic; I5 loop-steps-positive; I6 receipt-ed25519-verify; I7 receipt-columns-consistent; I8 flywheel-lineage`. Executed by `SZLHOLDINGS/szl-invariants`. Statuses never coerced. Λ untouched. |
+| CUDA speedup / tokens/s / joules | **UNAVAILABLE** | Not claimed. Receipted kernels, not silent CUDA. |
+GitHub source: [`szl-holdings/YARQA-ATTN`](https://github.com/szl-holdings/YARQA-ATTN) @ `160640bd8ed138e0170838c5a0de470ba8539367`. Artifacts: [`BENCH.laptop-blackwell.json`](./BENCH.laptop-blackwell.json), [`OPERATIONAL.json`](./OPERATIONAL.json).
+```python
+from kernels import get_kernel
+k = get_kernel("SZLHOLDINGS/YARQA-ATTN", revision="main", trust_remote_code=True)
+```
+<!-- SZL-KERNEL-OPERATIONAL:END -->
+# YARQA-ATTN
+<p align="center">
+  <img src="og-card.png" alt="YARQA-ATTN" width="100%"/>
+</p>
+`KANCHAY` · Doctrine v11 · Lean `749/14/163` · Λ = Conjecture 1 (advisory) · [a-11-oy.com](https://a-11-oy.com)
+Python kernel is on this repo. CPU `get_kernel` import-LIVE MEASURED (`7e533ce`). GPU cubins **UNAVAILABLE** (not ROADMAP). Not an alias of `szl-receipt-attn`. Not a fourth Flash / Flex / paged stack.
+<!-- SZL-KERNEL-STATUS:import-LIVE:START -->
+## Status
+<!-- SZL-ATELIER-CUT:v1:START -->
+## The cut
+FlashAttention is faster. YARQA is accountable. We steal the kernel discipline from NVIDIA and spend it on provenance, not FLOPs.
+An attention op whose softmax support is reconstructable from a signed log.
+### Silhouette → leave → SZL
+| Leader | Take, then tweak |
+|---|---|
+| Anthropic | Interpretability as a runtime artifact. |
+| NVIDIA | cuDNN / FlashAttention silhouette — then we add the receipt. |
+| Unsloth | Unrelated. Don't wrap this in FastLanguageModel. |
+Nobody else ships this combination. That is the point of a one-of-one.
+## Intended use
+Drop-in attention with an audit tape.
+## Limitations
+- Not a checkpoint.
+- Performance vs FlashAttention is not claimed.
+Canonical GitHub: [`szl-holdings/szl-khipu`](https://github.com/szl-holdings/szl-khipu/blob/main/szl_khipu/yarqa.py)
+<!-- SZL-ATELIER-CUT:v1:END -->
+> **STATUS: import-LIVE** on CPU Kernel Hub `get_kernel` (kernels `0.16.1`). GPU cubins **UNAVAILABLE** this session (not ROADMAP).
+| Thing | Label | Method / N / date / what-NOT |
+|---|---|---|
+| Kernel Hub `get_kernel` | **import-LIVE** | MEASURED 2026-08-28 3:08pm ET on kernels `0.16.1`. Package HEAD [`7e533ce`](https://huggingface.co/kernels/SZLHOLDINGS/YARQA-ATTN/commit/7e533ce702029061bc68f9f9cafe88efdd7f5f00) (`7e533ce702029061bc68f9f9cafe88efdd7f5f00`). README at MEASURE [`2871b3c`](https://huggingface.co/kernels/SZLHOLDINGS/YARQA-ATTN/commit/2871b3cbd73ee05e9f1aa010b75b190683f0ecd3). Legal name `yarqa-attn` (Python module `yarqa_attn`). Variants: `build/torch-universal` (default `get_kernel`) and `build/torch-cpu` (`backend="cpu"`). Working calls: `get_kernel("SZLHOLDINGS/YARQA-ATTN", revision="main", trust_remote_code=True)` and the same with `backend="cpu"`. `selfcheck` **ok**. `max_abs_vs_compartment_ref=3.58e-07` (full `3.5762786865234375e-07`), `path=torch_compartment`. What-NOT: no tokens/s; no joules; not a fourth Flash / Flex / paged stack. Lambda = Conjecture 1 (advisory). |
+| GPU cubins | **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 shipped. No tokens/s. No joules. CPU import-LIVE unchanged. Not a fourth Flash / Flex / paged stack. Lab stays Khipu. |
+<!-- SZL-KERNEL-STATUS:import-LIVE:END -->
+KERNEL kernel card. Original SZL **compartment / plug-flow** attention cut. Receipt-aware. Honesty-labeled.
+**Not a Fall 2026 ATELIER weight.** No tensors in this repo. Not an alias of [`szl-receipt-attn`](https://github.com/szl-holdings/szl-receipt-attn). Not a pointer at the Triton trio (`szl-receipt-attn`, `szl-maskmod`, `szl-block-kv`). Those three stay separate. a11oy-net does not list this as a fourth Flash / Flex / paged stack.
+GitHub is source of truth: [`szl-holdings/YARQA-ATTN`](https://github.com/szl-holdings/YARQA-ATTN). KERNEL binds Hub bytes from that tree. Do not PUT an empty card.
+| | |
+|---|---|
+| **Owner** | KERNEL |
+| **Artifact** | kernel (Python present; no weights; GPU cubins not claimed) |
+| **Status** | **import-LIVE** CPU · GPU cubins **UNAVAILABLE** |
+| **License** | Apache-2.0 |
+| **Λ** | Conjecture 1 (advisory, never a theorem) |
+| **Path** | `torch_compartment` (CPU) |
+| **Serve studio** | not this repo. Live CPU lab is [`szl-model-inference-lab`](https://huggingface.co/spaces/SZLHOLDINGS/szl-model-inference-lab) (Khipu GGUF only) |
+Silhouette: partition a sequence into canals (contiguous compartments), attend within a canal, emit SHA3-256 of the partition and of the attention output. We do not copy Dao hopper, Sage `csrc`, vLLM paged `.cu`, cuDNN FMHA, TRT cubins, CuTeDSL, or `flex_attention.py`. Metaphor only vs [`szl-holdings/yarqa`](https://github.com/szl-holdings/yarqa) (CFD; different product). Throughput is MEASURED only from a timed run on named hardware. Until then every speed claim is unstamped. No tokens/s. No joules.
+Do not list this next to Chaski, Qantu, Waman, Chakana, or Tinku.
+## Load
+```python
+from kernels import get_kernel
+attn = get_kernel("SZLHOLDINGS/YARQA-ATTN", revision="main", trust_remote_code=True)
+```
+Fashion GO 2026-08-28 3:10pm ET. import-LIVE CPU stays. GPU cubins stamped **UNAVAILABLE** 2026-08-28 7:01pm ET (no CUDA device this session). Not a fourth Flash / Flex / paged stack.
+Apache-2.0. Copyright 2026 SZL Holdings.