summaryrefslogtreecommitdiff
AgeCommit message (Expand)Author
2024-09-04Performance improvements for legacy quants on ARM_NEON (#37)Kawrakow
2024-09-04Zen4 Flash Attnetion 2 (#36)Kawrakow
2024-09-02Fix Zen4 Flash Attention (#35)Kawrakow
2024-09-02Do not process prompts containing binary data for escapes (#33)Kawrakow
2024-09-01Zen4 Flash Attention (#32)Kawrakow
2024-08-31Fix build when iqk_mul_mat is disabled (#31)Kawrakow
2024-08-27Faster Gemma2 (#27)Kawrakow
2024-08-21softcap: minor improvement (#24)Kawrakow
2024-08-20Fused soft cap and SIMD-ified GeLU (#9)Kawrakow
2024-08-20iq4_k: use iq5_k also when n_gqa = 2 (#23)Kawrakow
2024-08-19AVX2 quantization for Q8_K (#22)Kawrakow
2024-08-19quantize_stats: print rmse and max error as fraction of <x> (#21)Kawrakow
2024-08-19iq2_k: slightly better bpw - accuracy compromise (#20)Kawrakow
2024-08-14Skip barriers of noops (#19)Kawrakow
2024-08-12Update README.mdKawrakow
2024-08-12Merge mainline - Aug 12 2024 (#17)Kawrakow
2024-08-09Fix MakefileIwan Kawrakow
2024-08-09Fix Zen4 implementation of iq3_k, iq4_k, iq5_kIwan Kawrakow
2024-08-09iq6_k: AVX2Iwan Kawrakow
2024-08-09iq6_k: MetalIwan Kawrakow
2024-08-09iq6_k: NEONIwan Kawrakow
2024-08-09iq6_k: slightly better Zen4 iqk_mul_matIwan Kawrakow
2024-08-09iq6_k: Zen4 iqk_mul_matIwan Kawrakow
2024-08-09iq6_k: CUDA dot productIwan Kawrakow
2024-08-09iq6_k: CUDA dequantizeIwan Kawrakow
2024-08-09iq6_k: WIP (quantize/dequantize)Iwan Kawrakow
2024-08-09iq6_k: WIP (nothing works)Iwan Kawrakow
2024-08-07Adding IQ2_TN for use with ternary models (#13)Kawrakow
2024-08-05q2_K: allow it to detect ternary nets and quantize accordinglyIwan Kawrakow
2024-08-05Update README.mdKawrakow
2024-08-05iq3_k, iq5_k: faster quantizationIwan Kawrakow
2024-08-03iq4_k: speedup quantization by a factor of ~2Iwan Kawrakow
2024-08-01Add copyright noticeIwan Kawrakow
2024-08-01iq2/3_k: tiny bit faster Metal dot productsIwan Kawrakow
2024-08-01iq3_k: slightly faster Metal dequantize kernelIwan Kawrakow
2024-08-01iq3_k: Metal dot productIwan Kawrakow
2024-08-01iq2_k: Metal dot product finally worksIwan Kawrakow
2024-08-01iq3_k: Metal dequantizeIwan Kawrakow
2024-08-01iq3_k: NEONIwan Kawrakow
2024-08-01iq3_k: AVX2 iqk_mul_matIwan Kawrakow
2024-08-01iq3_k: AVX512 iqk_mul_matIwan Kawrakow
2024-08-01iq3_k: faster CUDA dot productIwan Kawrakow
2024-08-01iq3_k: CUDA dot productIwan Kawrakow
2024-08-01iq3_k: BasicsIwan Kawrakow
2024-08-01iq2_k: very slightly better CUDA dot productIwan Kawrakow
2024-08-01iq2_k: better CUDA dot productIwan Kawrakow
2024-08-01iq2_k: CUDA dot product finally worksIwan Kawrakow
2024-08-01iq5_k: CUDA dot product finally worksIwan Kawrakow
2024-08-01Factor out iqk CUDA dot productsIwan Kawrakow
2024-08-01iq5_k: CUDA dot product still not workingIwan Kawrakow