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/src/ggml-cuda.cu | |
| 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/src/ggml-cuda.cu')
| -rw-r--r-- | ggml/src/ggml-cuda.cu | 3 |
1 files changed, 2 insertions, 1 deletions
diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu index ba9d89aa..d34aa386 100644 --- a/ggml/src/ggml-cuda.cu +++ b/ggml/src/ggml-cuda.cu @@ -2753,9 +2753,10 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons case GGML_TYPE_IQ3_XXS: case GGML_TYPE_IQ4_NL: case GGML_TYPE_IQ4_XS: + case GGML_TYPE_IQ2_K: + case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ4_K: case GGML_TYPE_IQ5_K: - case GGML_TYPE_IQ2_K: case GGML_TYPE_IQ1_BN: case GGML_TYPE_IQ2_BN: return true; |