diff options
| author | Kawrakow <iwankawrakow@gmail.com> | 2024-12-15 09:54:21 +0100 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2024-12-15 09:54:21 +0100 |
| commit | 85c5a1a99569ccc00c280835fe3a69b4af02c43b (patch) | |
| tree | da421487d5ddd0467b2bfd6cbbfb2666406c46f1 /ggml/src/ggml.c | |
| parent | 20758edcae65213b2f575b6d23dfea67ad9dd0e0 (diff) | |
BF16_R16 - 16 interleaved bf16 rows (#142)
* Not working bf16_r4
* Adding bf16_r8
Small performance gain compared to bf16 - 258 t/s vs 234 t/s.
I guess, this is still sub-obtimal.
* bf16_rx: Very slightly faster by interleaving 16 rows
258 t/s -> 263 t/s
* Rename bf16_r4 to bf16_r16
We are interleaving 16 rows now.
* Cleanup unused stuff
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Diffstat (limited to 'ggml/src/ggml.c')
| -rw-r--r-- | ggml/src/ggml.c | 20 |
1 files changed, 20 insertions, 0 deletions
diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index 772c70c4..51ef6eb2 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -1231,6 +1231,19 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { .nrows = 1, .row_meta_size = 0, }, + [GGML_TYPE_BF16_R16] = { + .type_name = "bf16_r16", + .blck_size = 1, + .type_size = sizeof(ggml_bf16_t), + .is_quantized = false, + //.to_float = (ggml_to_float_t) ggml_bf16_to_fp32_row, + //.from_float = (ggml_from_float_t) ggml_fp32_to_bf16_row, + //.from_float_ref = (ggml_from_float_t) ggml_fp32_to_bf16_row_ref, + //.vec_dot = (ggml_vec_dot_t) ggml_vec_dot_bf16, + .vec_dot_type = GGML_TYPE_BF16, + .nrows = 1, + .row_meta_size = 0, + }, [GGML_TYPE_Q4_0_4_4] = { .type_name = "q4_0_4x4", .blck_size = QK4_0, @@ -4110,6 +4123,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) { case GGML_FTYPE_ALL_F32: wtype = GGML_TYPE_F32; break; case GGML_FTYPE_MOSTLY_F16: wtype = GGML_TYPE_F16; break; case GGML_FTYPE_MOSTLY_BF16: wtype = GGML_TYPE_BF16; break; + case GGML_FTYPE_MOSTLY_BF16_R16: wtype = GGML_TYPE_BF16_R16;break; case GGML_FTYPE_MOSTLY_Q4_0: wtype = GGML_TYPE_Q4_0; break; case GGML_FTYPE_MOSTLY_Q4_1: wtype = GGML_TYPE_Q4_1; break; case GGML_FTYPE_MOSTLY_Q5_0: wtype = GGML_TYPE_Q5_0; break; @@ -15748,6 +15762,7 @@ static void ggml_compute_forward_clamp( } break; case GGML_TYPE_F16: case GGML_TYPE_BF16: + case GGML_TYPE_BF16_R16: case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: case GGML_TYPE_Q5_0: @@ -22651,6 +22666,11 @@ size_t ggml_quantize_chunk( ggml_fp32_to_bf16_row_ref(src + start, (ggml_bf16_t *)dst + start, n); result = n * elemsize; } break; + case GGML_TYPE_BF16_R16: + { + repack_f32_bf16_r16(src + start, (char *) dst + start_row * row_size, nrows, n_per_row); + result = nrows * row_size; + } break; case GGML_TYPE_F32: { size_t elemsize = sizeof(float); |