No official torrent yet — requesting it tells the flock what to host next. You cannot seed a catalog-only page.
Cataloged from prism-ml/Bonsai-8B-AWQ-4-bit
All models by prism-ml →This repo provides an AWQ 4-bit checkpoint so you can run Bonsai-8B on sglang (or vLLM) until native 1-bit support lands in those engines. The repack from 1-bit to AWQ 4-bit is lossless. Both formats use group size 128, and Bonsai's binary weights (±d) fit exactly inside AWQ INT4 by an exact conversion formula.
For the best Bonsai experience on edge or consumer-grade hardware, use the native 1-bit releases. The 1-bit format is where Bonsai's memory and energy wins come from.
- Bonsai-8B MLX 1-bit — 1-bit MLX for Apple Silicon.
- Bonsai-8B-gguf - 1-bit gguf supported by llama.cpp across many backends (GPU, Metal, CPU, Vulkan, etc)
- Bonsai-8B FP16 — FP16 safetensors for stock HuggingFace tooling.
Bonsai weights are ±d (binary) with a shared scale across a group size of 128. INT4 can represent these values exactly. Embedding and lm_head stay FP16 due to sglang limitations.
AWQ INT4 dequantization: weight = scale × (int4 − zero).
With pre-built sgl-kernel wheels: 15 s. First launch on an arch without pre-built wheels (e.g. L40S / sm_89) takes 3–5 min while sglang JIT-compiles Marlin GEMM + FlashInfer kernels; artifacts cache to ~/.cache/tvm-ffi/ and ~/.cache/flashinfer/, so subsequent launches drop back to 15 s.
Example of successful end-to-end serving environment:
sglang[all] == 0.5.9torch == 2.9.1, transformers == 4.57.1, triton == 3.5.1ninja == 1.13 on PATHnvcc from CUDA 12.8 first on PATH (sglang's JIT Marlin needs -std=c++20; CUDA 11.x will fail)+d → scale=d, int4=9, zero=8 → d × (9-8) = +d-d → scale=d, int4=7, zero=8 → d × (7-8) = -dpip install sglang
python -m sglang.launch_server \ --model /path/to/Bonsai-8B-awq/ \ --port 8000 \ --dtype bfloat16# Completion APIcurl http://localhost:8000/v1/completions \ -H "Content-Type: application/json" \ -d '{"model":"Bonsai-8B","prompt":"The capital of France is","max_tokens":20}'
# Chat APIcurl http://localhost:8000/v1/chat/completions \ -H "Content-Type: application/json" \ -d '{"model":"Bonsai-8B","messages":[{"role":"user","content":"Who are you?"}],"max_tokens":100}'# Option A: DP=8 — 8 independent replicas, no inter-GPU commspython -m sglang.launch_server \ --model /path/to/Bonsai-8B-awq/ \ --dp-size 8 \ --load-balance-method total_tokens \ --port 8000 --dtype bfloat16
# Option B: TP=2 DP=4 — 4 replicas, each split across 2 GPUspython -m sglang.launch_server \ --model /path/to/Bonsai-8B-awq/ \ --tp-size 2 --dp-size 4 \ --load-balance-method total_tokens \ --port 8000 --dtype bfloat16
# Option C: TP=4 DP=2 — 2 replicas across 4 GPUs eachpython -m sglang.launch_server \ --model /path/to/Bonsai-8B-awq/ \ --tp-size 4 --dp-size 2 \ --load-balance-method total_tokens \ --port 8000 --dtype bfloat16