clay-tensor-host 0.0.1

The native host that answers Clay's Tensor and Tokenizer abilities (eager CPU, f32). A child process driven by the runner; today the tensor_run example driver.

✓ Signature verified (DEV key)

Checked against the dev trust set only: a DEV ONLY key, not a trust root. It proves the manifest was not altered since signing, not that the code is safe.

What is this?

A small native program that does the heavy number crunching for machine learning. A Clay program describes a model as steps like multiply, add and normalise; this host performs those steps on real arrays of numbers and hands back the result. It also turns text into the numbers a model reads (and back) using a tokenizer.

Who uses it?

The GPT-2 port (packages/tensor-host/clay/gpt2.clay), run by the tensor-host tests and the reference-diff harness.
Later: the Clay Cloud runner (rung M3) starts it as a child process next to the job.

What does it need?

No network. It reads the model files you point it at (see gpt2-small-weights). A macOS arm64 build today; other platforms would be separate entries.

Facts

Kind: process
Determinism: may-vary-across-versions
Licence: unknown
Size: 4.9 MB (4940192 bytes)
Clay ability: Tensor #q42ehuofl9q3topqpd02bej7akdfue33istc8imchmgbt4gm7fkcagjjhog7e4af4o6ahhs27v2vnmdc1fbks1av4mtmpr00oibpjuo
Tokenizer #nn3cjmcgge5kmp38u16enbiqnfe9ovph0ku3314nintthrdfnuoga5fjni48ujlgfi8q030m7r8epjn05j1t6i1rk5bmu5f0pbaoed8
Hashes are the structural ability hashes the kernel computes for the declarations in packages/tensor-host/clay (tensor.clay, gpt2.clay), as the page and the host see them when a program runs. The Tensor hash covers `type tensor`, which is structural, so every store computes the same hash (wishlist W35). A new operation is a new ability version and a new hash.
Status: ok
Manifest id: b3492f85c32563d9237e531cd80cdb257fa7d46b90ba398fabd194681d567d78

Upstream pin

Project: https://github.com/robmurtagh/clay
Version: 0.0.1
Commit / revision: 8fa21355ecac656a2ea725f16b352130268a2720
Licence: unknown

Built from

Package	Version	Licence
matrixmultiply	0.3.11	MIT/Apache-2.0
tokenizers	0.22.2	Apache-2.0
onig	6.5.3	MIT

Artifact

SHA-256: 14dc1a3d31b16412ae7a432564772ca0c7426f89a6d27dc730e94326d69538ba
Format: file
Source path: packages/tensor-host/target/release/examples/tensor_run
Download: /v1/artifacts/14dc1a3d31b1...

Resources

Memory: unknown
CPU: aarch64; this build is a macOS arm64 Mach-O executable
Network: none

Provenance

Toolchain: repo-local Rust 1.98.1 (.tools), release profile opt-level 3, LTO, 1 codegen unit
Built by: repo maintainer
Built at: 2026-11-11T06:30:08Z
Notes: upstream.commit is the repository HEAD when this entry was made (the build is newer than the last change to packages/tensor-host/src). determinism is the closest class in this vocabulary: output bits are NOT promised equal across devices or builds (docs/ml-workloads.md section 6). The build is not reproducible byte for byte across machines. Licence null: the repo declares none yet. Dependencies listed are the direct ones plus the regex engine tokenizers links; the full transitive tree is not listed.

cd packages/tensor-host && cargo build --release --example tensor_run
Versions are those in packages/tensor-host/Cargo.lock; licences are read from each crate's Cargo.toml in the local cargo registry.

Signatures (checked when this page was built)

c5878029baeb4574 DEV ONLY signing key (not a trust root) DEV ONLY: valid (ok)

How a runner loads it

Fetch /v1/manifests/b3492f85c32563d9237e531cd80cdb257fa7d46b90ba398fabd194681d567d78.json, get the artifact bytes by hash, then call checkLoadable(manifest, bytes, policy). It checks the signature against the trusted keys, the revocation list, the fingerprints and sizes, and the policy's allowed kinds and determinism classes, and returns a reason when it refuses.