Repository Issue Activity (beta)

#197 Bug: Support consumer RDNA2 (gfx1030) under ROCm/HIP backend for Ternary-Bonsai-27B

#206 cuda version bug in download_binaries.sh?

#196 Three CPU/AVX2 + recurrent-state PRs for Ternary Bonsai-2 27B (3.9 → 8.3 tok/s bit-exact on a laptop CPU)

#174 LM Studio MLX: Failed to load the model

#204 Try to plant laya inside your project for better routing

#182 Bonsai-2 27B crashes on load — ggml-cpu AVX-512 access violation on Zen5 (Ryzen 9 9950X, Windows); Intel Arc A770 unused → CPU-only ~2 tok/s

#194 prism-ml/Ternary-Bonsai-2-27B-mlx-2bit model is not part of the supported models within the mlx_lm package.

#193 scripts: context tier sizes by system RAM only - default -c doesn't fit on mid-size GPUs

#176 Windows setup script not detecting CUDA correctly

#191 ptq1_0 Vulkan matvec kernel ~6x slower than tq2_0 at n=1 (195 vs 1080 GFLOPS, RX 6600)

#184 Setup.ps1 only downloads PQ2_0 for Bonsai-2 27B; PTQ1_0 (which has working Vulkan kernels) is never offered

#183 [27B / agentic] Harness works, code quality unusable — feedback wanted on setup and approach (Apple M5 Max)

#181 community-benchmarks: Bonsai 2 27B on RTX 3090 (Windows, CUDA)

#177 failed to compile on macOS 27

#130 Bonsai Alternatives

#137 Please consider making a mainline compatible dspark draft model or publish some advice on that

#139 Combination of BONSAI_KV4 and BONSAI_SPECULATIVE on 27B Ternary reports issue with KV cache tpye

#142 Just comment: bonsai 1.7B on haswell i5 iMac

#82 Ternary-Bonsai-27B Q2 produces gibberish; Ternary-Bonsai-8B Q2 works on same host/build

#72 POC: From-scratch Bonsai Q1_0 inference engine running on WebGPU compute shaders

#149 Qwen3.8 27b?

#145 [CUDA/mainline] Q4 KV causes a 3x-67x prompt-processing slowdown on Ternary-Bonsai-27B

#148 Request: Bonsai conversion of ThinkingCap-Qwen3.6-27B (halves thinking tokens, composes with ternary)

#147 [27B multimodal] Repeated 14K prompts get no prefix-cache reuse

#146 [27B/OpenAI API] Clarify reasoning defaults and per-request opt-out for agentic tool calls