Release notes#

What is new, and where it has been proven.

0.1.0 (unreleased)#

First version.

Language

  • .vis with canvases, kernels, measures and a run block; clamp, wrap and zero edges.

  • 46 built-in functions implemented by Vision, identical on every tested device.

  • Loading .npy, .csv, .txt, .asc and .png, with explicit handling of missing data.

  • Measures with a fixed reduction order and about 44 bits of precision.

  • 3-D canvases (canvas name[w, h, d], the coordinate z, reads t[x, y, z], 2-D and 3-D canvases in one program), 3-D .npy, and .vti files (VTK ImageData, for ParaView) of a channel or of a whole canvas.

  • Derivatives of a program with respect to its params (vision.differentiate) and fitting params to data (vision.calibrate): the same fitted bits on every device.

Proof

  • Receipts, Ed25519 signatures, RFC 3161 time-stamps and self-contained HTML reports.

  • Verification by re-execution, also without a GPU (numpy).

  • Conformance kit with 15 cases and an HTML certificate.

Tensor library

  • Autodiff, GPT, AdamW and SGD; auditable, gpu-exact and fast modes.

  • Exact int8 inference.

  • Audited training with a chained log, and heterogeneous training across vendors.

  • Number formats for matrix products: float32, mx8, mx16 and mx16r, per call, per layer or per device, also on tensor cores; the same bits on every GPU and on numpy. mx16 trains the mini-GPT like float32 (validation loss 1.7855 against 1.7856).

  • Inference of pretrained models of the Llama family (Qwen2.5, Llama, Mistral, SmolLM2) from safetensors or GGUF files, with seeded sampling, batches and inference receipts — see the LLM example in the repository.

  • An optional native runtime that makes queueing work on the GPU faster.

Apple Silicon (Metal)

The first runs on a Mac (GitHub’s macOS arm64 runner) found three differences, now handled; the conformance kit and the tests then pass with the same bits as on NVIDIA, Intel, llvmpipe and numpy.

The macOS runner is a virtual machine with a paravirtualised GPU: the bits are meaningful, the speeds are not those of a real M-series Mac.

Tested on NVIDIA RTX 4050 Laptop (driver 595.99), Intel UHD RPL-P (Mesa 25.2), llvmpipe (LLVM 20.1) and numpy, on Linux; Apple Silicon via Metal (macOS 15.7, GitHub runner); Windows via DX12 (WARP, GitHub runner). AMD is not tested yet (the kit is ready: vision conformance on the machine).

On this page