Repository Issue Activity (beta)

#99 CUDA vec FA kernel: turbo4 V-cache branch consumes 4 of 8 elements per iteration — half of every head's V output is zero on batch-1 decode (patch attached)

#97 uv pip install turboquant also required

#92 Can it be used together with the MTP draft model?

#89 Consider proper attribution for EDEN quantization

#87 Use scale factor for improvements

#27 Upstream: TurboQuant discussion + contribution requirements for llama.cpp

#88 attn-rotation for native llama quants

#86 [ROCm] Scale operation fails with "invalid device function" during Gemma 4 loading

#85 Feature request: prebuilt binary for Windows CPU only

#84 Diagnostic: [RTX 3090]

#77 TurboFlash attention kernel: 13 GB excess Metal compute buffer allocation (commit 0d6b38aad)

#82 Question about QJL abliation study

#80 time and performance overhead of quantization and dequantization

#79 Reproducibility of QJL resurrection issue

#74 it run

#72 End-to-end inference test fails with --cache-type turbo4 on CPU

#70 Error when building from source

#69 polar coordinate transformation operations

#68 llama-quantize crashes with ios_base::failbit when using --tensor-type-file or --tensor-type (regression from #20503)

#42 Enable sparse V dequant on pre-M5 hardware after verification

#60 [Bug] Numerical instability (NaN/Infinite Loop) with Qwen 2.5/3.5 models using turbo4 on RTX 4070Ti

#59 Greetings from TurboQuant-MLX & Some Insights to Share! (System Prompt vs Boundary Layers)

#57 turbo - cant compile in Windows

#58 Unsupported gpu architecture 'compute_120a'

#48 No Difference in tokens/sec - Ministral3 8B Q5_K_M