{
"mcpServers": {
"hf-mcp-server": {
"url": "https://huggingface.co/mcp",
"headers": {
"Authorization": "Bearer <YOUR_HF_TOKEN>"
}
}
}
}
backend="onnx" or backend="openvino" when initializing a SparseEncoder to get started, but I also included utility functions for optimization, dynamic quantization, and static quantization, plus benchmarks.n-tuple-scores output format from mine_hard_negativesgather_across_devices=True to load in-batch negatives from the other devices too! Essentially a free lunch, pretty big impact potential in my evals.transformers, and you install trackio with pip install trackio, then your experiments will also automatically be tracked locally with trackio. Just open up localhost and have a look at your losses/evals, no logins, no metric uploading.
I'm using https://artificialanalysis.ai/ just because it puts everything in one place! It's not the best resource but these days I'm all about saving time.
@ThomasTheMaker if you make an issue on the repo, I'll look into it!
@ThomasTheMaker it's just the raw attention and transformer architecture in golang designed for serverless so performance will definitely be less than ggml and llama.cpp since it's not accelerated by GPU's but if you're into edge AI CPU only, this is the first, only and best way to compute attention.
Quantization can definitely be supported as it's just a math model!
This is a pretty big update for sure. The models have improved significantly which is great for everyone involved, especially the end user. Those datasets look very promising as well!
Sounds interesting, I’ll check it out!