prism-ml/Bonsai-8B-AWQ-4-bit

Bonsai-8B — AWQ 4-bit

Summary

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](https://huggingface.co/prism-ml/Bonsai-8B-mlx-1bit) — 1-bit MLX for Apple Silicon. - [Bonsai-8B-gguf](https://huggingface.co/prism-ml/Bonsai-8B-gguf) - 1-bit gguf supported by llama.cpp across many backends (GPU, Metal, CPU, Vulkan, etc) - [Bonsai-8B FP16](https://huggingface.co/prism-ml/Bonsai-8B-unpacked) — FP16 safetensors for stock HuggingFace tooling.

How It Works

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).

+d → scale=d, int4=9, zero=8 → d × (9-8) = +d
-d → scale=d, int4=7, zero=8 → d × (7-8) = -d

Serve

pip install sglang

python -m sglang.launch_server \
    --model /path/to/Bonsai-8B-awq/ \
    --port 8000 \
    --dtype bfloat16

Use

# Completion API
curl http://localhost:8000/v1/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"Bonsai-8B","prompt":"The capital of France is","max_tokens":20}'

# Chat API
curl 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}'

Multi-GPU (8× H100)

# Option A: DP=8 — 8 independent replicas, no inter-GPU comms
python -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 GPUs
python -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 each
python -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

Appendix

Launch time

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.

Known-good environment

Example of successful end-to-end serving environment:

• —sglang[all] == 0.5.9
• —torch == 2.9.1, transformers == 4.57.1, triton == 3.5.1
• —ninja == 1.13 on PATH
• —nvcc from CUDA 12.8 first on PATH (sglang's JIT Marlin needs -std=c++20; CUDA 11.x will fail)
• —Python 3.12