diff options
| author | Kawrakow <iwankawrakow@gmail.com> | 2024-12-09 16:59:18 +0100 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2024-12-09 16:59:18 +0100 |
| commit | 3ec193b4856df8e5827b83a8c7686e8498c5e5b8 (patch) | |
| tree | 149666dbffdf1d443bb9ff8f2564ed9bb1959201 /ggml/src/ggml-common.h | |
| parent | 43e65a672a98d931998559785b58f1e980e87f54 (diff) | |
Q4_K_R4 (#129)
* Something is still wrong
* Simply don't see what is wrong
* q4_k_r4: finally works on Zen4
I had forgotten to prevent token_embd.weight being quantized
with q4_k_r4!
* q4_k_r4: AVX2
We get PP-512(LLaMA-3.1-8B) = 267 t/s on a Ryzen-5975WX.
This is ~30% better than Q4_K_S.
* q4_k_r4: NEON
We get PP-512(LLaMA-3.1-8B) = 110 t/s.
Not quite as good as q4_0_r4, but still a massive
improvement compared to he 69 t/s for q4_K.
* q4_k_r4: slightly better AVX2
PP-512 goes from 267 t/s to 282 t/s on Ryzen-5975WX
* Minor
* Minor
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Diffstat (limited to 'ggml/src/ggml-common.h')
| -rw-r--r-- | ggml/src/ggml-common.h | 8 |
1 files changed, 8 insertions, 0 deletions
diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h index 3f701873..784125fc 100644 --- a/ggml/src/ggml-common.h +++ b/ggml/src/ggml-common.h @@ -299,6 +299,14 @@ typedef struct { } block_q4_K; static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding"); +typedef struct { + ggml_half d[8]; + uint8_t scales_h[QK_K/16];// scales and mins, quantized with 6 bits + uint8_t scales_l[QK_K/8]; // scales and mins, quantized with 6 bits + uint8_t qs[QK_K*2]; // 4--bit quants +} block_q4_k_r4; +static_assert(sizeof(block_q4_k_r4) == 8*sizeof(ggml_half) + QK_K/16 + QK_K/8 + QK_K*2, "wrong q4_k_r4 block size/padding"); + // 5-bit quantization // 8 blocks of 32 elements each // weight is represented as x = a * q + b |