turboquant_plus · Issues· 47 open

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)

uv pip install turboquant also required

Can it be used together with the MTP draft model?

Upstream: TurboQuant discussion + contribution requirements for llama.cpp

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

Feature request: prebuilt binary for Windows CPU only

Diagnostic: [RTX 3090]

Question about QJL abliation study

time and performance overhead of quantization and dequantization

Reproducibility of QJL resurrection issue

it run

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

Error when building from source

polar coordinate transformation operations

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