v0.1.13

pip install -U flide

What's new

• New models: FlexDM-LoRA-v2.1, Sieve-31B-v3 and JDI-Decision-V1, which reads the text in images with PaddleOCR as its release does.
• Precompiled kernels for RTX 40 series GPUs, and --quantize fp8 now tested on them.
• flide serve reports what it has done on GET /stats (#12).
• flide run takes the engine and family options that flide serve takes (#10).
• A request whose text spells an image placeholder now gets a 422 instead of a server error.
• Pull requests now start from an accepted issue, as CONTRIBUTING.md describes.

The wheel under Assets is the same file as on PyPI, for installing without PyPI. It carries precompiled kernels for sm_86 (A40, A10, RTX 30 series), sm_89 (RTX 40 series), sm_90 (H100, H200) and sm_120 (RTX 50 series, RTX PRO 6000) with torch 2.14 and 2.13; other GPUs compile kernels on first start.

Full changelog: v0.1.12...v0.1.13

v0.1.12

pip install -U flide

What's new

• New models: GLiDE 0.8B and 2B, Nimble-v4-max and clef-flash.
• Precompiled kernels for RTX 50 series and RTX PRO 6000 GPUs, and Gemma 4 and the batch-invariant mode now run on them (#5).
• Flide says at start when no precompiled kernels fit the GPU, so a first start that compiles no longer looks hung (#2).
• flide run warms up before it starts timing (#6).
• flash-linear-attention 0.6 support. Thanks to @reneleonhardt.
• OneJev and Open-Jev plan long requests faster, with the same answers as before.

The wheel under Assets is the same file as on PyPI, for installing without PyPI. It carries precompiled kernels for sm_86 (A40, A10, RTX 30 series), sm_90 (H100, H200) and sm_120 (RTX 50 series, RTX PRO 6000) with torch 2.14 and 2.13; other GPUs compile kernels on first start.

New contributors

• @Lumea-art made their first contribution in #3.

Full changelog: v0.1.11...v0.1.12

Flide 0.1.11

pip install -U flide

New example: the email triage demo, where Flide tags 10,000 emails with eight questions each in 28 seconds on one H200. The folder has the code to run it yourself.

When the installed torch does not match the GPU driver, the error message now gives the command that installs a matching one. Thanks to Rene Leonhardt for this change.

The repository now has a contributing guide, issue and pull request templates, and CPU checks on every pull request.

Flide 0.1.10

pip install -U flide

New models: Ainvest-Decision, Darwin-27B-RSI-Decision and KnowLine-31B-Gen1.5.

The KnowLine family now reads images, as image parts of a chat state, as the request's images, and as Jev assets for KnowLine-31B-Gen1.5.

Each plugin in plugins/ is now one file, or a folder when a family has several. Imports and entry points are unchanged.

Speeds for every model are in the docs.