Issues

Issues

• contact

  quantumaikr/quant.cpp#97 · opened on Jun 14, 2026

• Gemma 4 forward pass reference points from a working CUDA implementation

  quantumaikr/quant.cpp#96 · opened on Apr 14, 2026

• Research: 32K-vocab English-optimized small model for quant.cpp

  quantumaikr/quant.cpp#92 · opened on Apr 12, 2026

• Port Phi-3 fused QKV/FFN from quant.h to libturboquant (enables Metal GPU)

  quantumaikr/quant.cpp#91 · opened on Apr 12, 2026

• RLV Locator: add lightweight semantic search as third RRF signal

  quantumaikr/quant.cpp#89 · opened on Apr 12, 2026

• Single Source of Truth: auto-generate split sources from quant.h

  quantumaikr/quant.cpp#85 · opened on Apr 12, 2026

• KV cache pre-build for RLV: question latency 35s → 5s

  quantumaikr/quant.cpp#83 · opened on Apr 12, 2026
  1

• Qwen3.5-4B DeltaNet layers: FP32 dequant bottleneck causes 0.7 tok/s

  quantumaikr/quant.cpp#70 · opened on Apr 12, 2026

• Qwen3.5-4B: quant_generate works but quant_ask produces empty/garbage output

  quantumaikr/quant.cpp#69 · opened on Apr 12, 2026
  2

• Metal backend is currently slower than CPU-only on all tested models

  quantumaikr/quant.cpp#16 · opened on Apr 8, 2026
  1

• TurboQuant follow-ups: per-channel outliers, Llama 3.1 8B reproduction, 5-bit codebook

  quantumaikr/quant.cpp#15 · opened on Apr 7, 2026
  1