Age | Commit message (Expand) | Author |
2022-09-13 | Fixed cpu blockwise quantization for small input tensors. | Tim Dettmers |
2022-09-11 | Fixed 2^31 max size issue for cpu blockwise quant. | Tim Dettmers |
2022-08-23 | Fixed issue where Pascal was not displaying proper error. | Tim Dettmers |
2022-08-16 | Enhanced error handling in CUDA SETUP failures. | Tim Dettmers |
2022-08-16 | Added fused bias in dequant_mm. | Tim Dettmers |
2022-08-16 | Removed storage() from get_ptr; added boilerplate for bias dequant_mm. | Tim Dettmers |
2022-08-06 | Removed faulty asserts. | Tim Dettmers |
2022-08-04 | Merge branch 'extract_outliers' into debug | Tim Dettmers |
2022-08-03 | Added fixes for the case that matmullt dim A is zero, e.g. [0, 768]. | Tim Dettmers |
2022-08-03 | Added CUDA block assert and is_on_gpu check. | Tim Dettmers |
2022-07-26 | Merge branch 'patch_merge' into extract_outliers | Tim Dettmers |
2022-07-26 | Added col_ampere outlier extraction kernel. | Tim Dettmers |
2022-07-26 | Working outlier extraction for Turing. | Tim Dettmers |
2022-07-26 | Boilerplate and test for extract_outliers. | Tim Dettmers |
2022-07-26 | Fixed cpuonly build. | Tim Dettmers |
2022-07-25 | Some progress on build script; added multi-cuda install script. | Tim Dettmers |
2022-07-25 | Fixed makefile; fixed Ampere igemmlt_8 bug. | Tim Dettmers |
2022-07-22 | Fixed rowcol synchronization bug. | Tim Dettmers |
2022-07-22 | Most tests passing. | Tim Dettmers |
2022-07-01 | Reduce diff | Max Ryabinin |
2022-07-01 | Reduce diff | Max Ryabinin |
2022-07-01 | Reduce diff | Max Ryabinin |
2022-07-01 | Reduce diff | Max Ryabinin |
2022-07-01 | Add a CPU-only build option | Max Ryabinin |
2021-11-28 | Added AdamW. #10 #13 | Tim Dettmers |
2021-11-10 | Added adagrad with tests (no clipping). | Tim Dettmers |
2021-10-21 | Added compilation from source instructions; easier compilation. | Tim Dettmers |
2021-10-20 | Added skip_zeros; tests are passing. | Tim Dettmers |
2021-10-20 | Initial plumbing for skip_zeros. | Tim Dettmers |
2021-10-05 | Initial commit | Tim Dettmers |