Skip to content

Paddock 0.1.13

Choose a tag to compare

View all tags
@panterlo panterlo released this 03 Oct 18:44
v0.1.13
70dd87a

A decisions and speech release: Cloudflare's Clef and Clef Flash answer
fixed questions about a text or a picture, NVIDIA's Nemotron 3 Diarization
tells who spoke when, and Reads gains a Camera mode that asks your questions
of the webcam live. The DGX Spark now reports its memory. Windows x64,
Linux x64 and the NVIDIA DGX Spark. NVIDIA GPUs, driver 580 or newer. The
macOS pre-release for Apple Silicon is built from the same commit.

New

  • Clef and Clef Flash (Cloudflare). Decision models: give them a text, a
    picture or both, and a set of yes/no, choice and score questions, and each
    answer comes back with a calibrated probability, every option scored in one
    pass. Clef Flash is built on a 9B model and Clef on a 27B one; both serve
    in 8-bit with their own vision, on NVIDIA GPUs and on the Mac, behind the
    same /v1/systemone call the Reads page uses. Clef Flash needs about 16 GB
    of GPU memory, Clef about 37 GB.

  • Nemotron 3 Diarization (NVIDIA). Who spoke when, for a recording or
    live audio, on NVIDIA GPUs and on the Mac. The Studio shows a speaker
    timeline beside the transcript, with each word attributed to its speaker,
    and keeps it with the conversation.

  • Camera mode in Reads. Switch Reads to Camera and your questions are
    asked of the webcam frame after frame. The answers ride on the picture as
    large cards - yes in green, no in red, a choice in white, each with its
    confidence - and a card flashes when its answer flips. Snapshot keeps a
    frame and its answers in the read's history. In the web Studio and the Mac
    app, with any reader that takes pictures.

Improved

  • The start form proposes the weights paddock serve would start, and a
    full-precision build is labelled "Full precision - no quantization" instead
    of being offered as a smaller one.

  • paddock ps names decision models such as Laya and Clef instead of
    showing "-".

  • Models from Cloudflare show Cloudflare's logo.

Fixed

  • The DGX Spark reports its memory. The Studio said "No GPU" beside every
    model on a Spark, and starting a model there was never checked against the
    memory it needs. Both now read the Spark's unified memory.

  • A cloud model refuses a malformed output cap. A max_output_tokens
    that is not a whole number used to be forwarded, dropped or replaced
    depending on the provider; it is now refused with a 400 before anything is
    sent. Thanks to @DevChiniwala (#35).

macOS (pre-release)

  • Clef and Clef Flash run natively on Metal, in 8-bit, with vision.
  • Reads has a native camera, with snapshots saved in the read's history.
  • Nemotron 3 Diarization runs on Metal, with speaker timelines in the app.

Known

  • Clef reads text and pictures; videos are refused.

  • Nemotron 3 Diarization is qualified against NVIDIA's reference on the
    pyannote and AMI test sets;     other recordings are not measured yet.

  • Laya reads text only, and its confidence on choices with more than ten
    options is not calibrated: the English checkpoint ships an out-of-range
    temperature for that case, which Paddock clamps, as Laya's own server does.

  • Kumo Tabular predicts up to ten classes.

  • On Qwen 3.8 Flash-Next's NVFP4 versions a greedy speculative reply can
    part from plain decoding at a near tie     - the same model with slightly
    different numerics, not a wrong answer.

  • On-demand loading covers Whisper only.

  • The fp8 KV cache's paged attention on RTX 50-series cards shows a small
    numeric deviation in one split configuration (#5).

Paddock 0.1.13, built from 70dd87ad on the maintainers' machines (Windows natively, Linux x64 in a Docker builder, the DGX Spark natively on a Spark), each archive booted on a GPU and asked a question before it was uploaded.

Requirements

  • An NVIDIA card the kernel pack covers - x64: RTX 30/40/50 series, the RTX A / PRO line, B200; aarch64: the DGX Spark (GB10). Other cards get an honest refusal, never a slow fallback. This build: sm_86 sm_89 sm_100 sm_100a sm_120 sm_120a.
  • A driver from the r580 branch or newer. No CUDA toolkit: the kernels and the runtime are inside paddock-runner.
  • Windows 11 x64; Linux x64 with glibc 2.35 or newer (Ubuntu 22.04+, Debian 12+); or a DGX Spark on DGX OS 7 (Ubuntu 24.04, glibc 2.39).

Run

  • Unzip or untar anywhere. paddock starts the manager and the Studio on http://localhost:11500. paddock-runner --model <file.gguf> serves a bare OpenAI/Anthropic-compatible endpoint. README.txt inside says the rest.

sha256

25b6ca762e43ce6011bac556515440152e05bebdb2cf13587cb7b4207d4a83c9  paddock-0.1.13-x86_64-windows.zip
d179c5bca8e0f9dc5bca8f806991645a344e4eda90857e5043f4c8c3cad5b701  paddock-0.1.13-x86_64-linux.tar.gz
9fa72d5e28825992d94a5c7cf55336f4f1aacafc90c4015873aed50efd1eeef6  paddock-0.1.13-aarch64-linux.tar.gz

Contributors

DevChiniwala
Assets 14
Loading