Bonsai 2 PQ2 for vLLM

Packed PQ2 safetensors conversion of Prism ML's Bonsai 2 27B. Requires the Bonsai 2 vLLM plugin and vLLM 0.29.0.

Validated: A100 40GB, BF16 text, 2,048-token context, one executing request. The fast profile uses a separate DFlash2 drafter and about 34 GiB of GPU memory. Vision, multi-GPU and longer contexts are not validated.

Quick start

Requires Linux, Python 3.12, Git, uv, a CUDA 13-compatible driver, CUDA toolkit (nvcc) and a C++ compiler.

git clone --branch v0.1.0 https://github.com/wonder-dot-ai/bonsai2-vllm-plugin.git
cd bonsai2-vllm-plugin
uv sync --locked
uv run --locked python tools/download_models.py
uv run --locked bonsai2-serve

OpenAI-compatible API: http://127.0.0.1:8000/v1, model name: bonsai2. First startup may take several minutes to compile kernels.

See the plugin README for benchmarks, configuration and conversion details.

Apache-2.0. Created using Bonsai by Prism ML; upstream notices are included in NOTICE.txt.

Downloads last month
149
Safetensors
Model size
7B params
Tensor type
F32
F16
U8
Inference Providers NEW
This model isn't deployed by any Inference Provider. 馃檵 Ask for provider support

Model tree for the-sweater-cat/bonsai-2-pq2

Base model

Qwen/Qwen3.8-27B
Quantized
(21)
this model