BotPsy-0.8B in CMF: runs on cortiq (CUDA, Vulkan, Metal, CPU)

#2
by infosave - opened

Hi! I added support for BotPsy-0.8B to cortiq, a Rust inference engine with no Python, and converted it into its single-file format (CMF): infosave/BotPsy-0.8B-cmf.

Two files, both quantized from your BF16 weights (text only):

  • botpsy-0.8b-q8_2f.cmf, 8-bit, 783 MB: perplexity 12.20 against 12.10 for BF16 (450 tokens);
  • botpsy-0.8b-q4tp.cmf, 4-bit, 422 MB: perplexity 13.88.

It runs on NVIDIA cards through CUDA or Vulkan, AMD and Intel through Vulkan, Apple silicon through Metal, and on the CPU. Decode, one stream: RTX A5000 on CUDA 339 tok/s (q4tp), on Vulkan 212-222 tok/s; Apple M4 on Metal 80-132 tok/s, on the CPU 78-88. A 4096-token prompt on the A5000 (q4tp, CUDA) is read at 20 100 tok/s, the first token after 0.20 s.

Tool calling works through the OpenAI-compatible server (cortiq serve): a get_weather request comes back as message.tool_calls with finish_reason: "tool_calls".

Your Apache-2.0 license ships with the files (LICENSE, ATTRIBUTIONS.md). If something on the page should be worded differently, tell me and I will fix it. Thank you for the model and for publishing the weights.

Sign up or log in to comment