diff options
| author | Iwan Kawrakow <iwan.kawrakow@gmail.com> | 2024-07-30 16:11:25 +0300 |
|---|---|---|
| committer | Kawrakow <48489457+ikawrakow@users.noreply.github.com> | 2024-08-01 09:38:06 +0200 |
| commit | 4f237d44f6d75afbb5cef39d4d6b0b35b2a517c7 (patch) | |
| tree | 097eb7ee5be6ec08d3958f747775f9225db3c7a4 /ggml/include | |
| parent | 36204c4ec7815f1b0f8f38c8e1cd3e6b3fef2eaa (diff) | |
iq3_k: Basics
Quantize/dequantize, CUDA dequantize.
PPL of LLaMA-3.1-8B is better than iq3_s and iq3_m.
Diffstat (limited to 'ggml/include')
| -rw-r--r-- | ggml/include/ggml.h | 10 |
1 files changed, 6 insertions, 4 deletions
diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h index b7585ad6..94ffae7e 100644 --- a/ggml/include/ggml.h +++ b/ggml/include/ggml.h @@ -390,8 +390,9 @@ extern "C" { GGML_TYPE_IQ2_BN = 35, GGML_TYPE_Q8_K64 = 36, GGML_TYPE_IQ2_K = 37, - GGML_TYPE_IQ4_K = 38, - GGML_TYPE_IQ5_K = 39, + GGML_TYPE_IQ3_K = 38, + GGML_TYPE_IQ4_K = 39, + GGML_TYPE_IQ5_K = 40, GGML_TYPE_COUNT, }; @@ -439,8 +440,9 @@ extern "C" { GGML_FTYPE_MOSTLY_IQ1_BN = 28, // except 1d tensors GGML_FTYPE_MOSTLY_IQ2_BN = 29, // except 1d tensors GGML_FTYPE_MOSTLY_IQ2_K = 30, // except 1d tensors - GGML_FTYPE_MOSTLY_IQ4_K = 31, // except 1d tensors - GGML_FTYPE_MOSTLY_IQ5_K = 32, // except 1d tensors + GGML_FTYPE_MOSTLY_IQ3_K = 31, // except 1d tensors + GGML_FTYPE_MOSTLY_IQ4_K = 32, // except 1d tensors + GGML_FTYPE_MOSTLY_IQ5_K = 33, // except 1d tensors }; // available tensor operations: |