diff options
Diffstat (limited to 'examples')
| -rw-r--r-- | examples/benchmark/benchmark-matmult.cpp | 6 | ||||
| -rw-r--r-- | examples/llava/clip.cpp | 55 |
2 files changed, 3 insertions, 58 deletions
diff --git a/examples/benchmark/benchmark-matmult.cpp b/examples/benchmark/benchmark-matmult.cpp index e89f3de2..47cb16c6 100644 --- a/examples/benchmark/benchmark-matmult.cpp +++ b/examples/benchmark/benchmark-matmult.cpp @@ -189,12 +189,10 @@ int main(int argc, char ** argv) { int32_t nelements = sizex*sizey; - std::vector<int64_t> hist_cur(1 << 4, 0); - // Set up a the benchmark matrices // printf("Creating new tensor q11 & Running quantize\n"); struct ggml_tensor * q11 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey); - ggml_quantize_chunk(qtype, (const float *) m11->data, q11->data, 0, nelements/m11->ne[0], m11->ne[0], hist_cur.data(), nullptr); + ggml_quantize_chunk(qtype, (const float *) m11->data, q11->data, 0, nelements/m11->ne[0], m11->ne[0], nullptr); // Set up a the compute graph // printf("Creating new tensor q31\n"); @@ -207,7 +205,7 @@ int main(int argc, char ** argv) { // Set up a second graph computation to make sure we override the CPU cache lines // printf("Creating new tensor q12 & Running quantize\n"); struct ggml_tensor * q12 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey); - ggml_quantize_chunk(qtype, (const float *) m12->data, q12->data, 0, nelements/m12->ne[0], m12->ne[0], hist_cur.data(), nullptr); + ggml_quantize_chunk(qtype, (const float *) m12->data, q12->data, 0, nelements/m12->ne[0], m12->ne[0], nullptr); // printf("Creating new tensor q32\n"); struct ggml_tensor * q32 = ggml_mul_mat(ctx, q12, m2); diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp index ef9e4ba7..6653b815 100644 --- a/examples/llava/clip.cpp +++ b/examples/llava/clip.cpp @@ -1862,7 +1862,6 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i std::vector<uint8_t> work(512); std::vector<float> conv_buf(512); - std::vector<int64_t> hist_all(1 << 4, 0); size_t total_size_org = 0; size_t total_size_new = 0; @@ -1917,48 +1916,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i } new_data = work.data(); - std::vector<int64_t> hist_cur(1 << 4, 0); - - switch (new_type) { - case GGML_TYPE_Q4_0: { - new_size = ggml_quantize_q4_0(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data()); - } break; - case GGML_TYPE_Q4_1: { - new_size = ggml_quantize_q4_1(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data()); - } break; - case GGML_TYPE_Q5_0: { - new_size = ggml_quantize_q5_0(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data()); - } break; - case GGML_TYPE_Q5_1: { - new_size = ggml_quantize_q5_1(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data()); - } break; - case GGML_TYPE_Q8_0: { - new_size = ggml_quantize_q8_0(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data()); - } break; - case GGML_TYPE_Q2_K: { - new_size = ggml_quantize_q2_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data()); - } break; - case GGML_TYPE_Q3_K: { - new_size = ggml_quantize_q3_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data()); - } break; - case GGML_TYPE_Q4_K: { - new_size = ggml_quantize_q4_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data()); - } break; - case GGML_TYPE_Q5_K: { - new_size = ggml_quantize_q5_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data()); - } break; - case GGML_TYPE_Q6_K: { - new_size = ggml_quantize_q6_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data()); - } break; - default: { - fprintf(stderr, "%s: unsupported quantization type %d\n", __func__, new_type); - return false; - } - } - - for (size_t j = 0; j < hist_cur.size(); ++j) { - hist_all[j] += hist_cur[j]; - } + new_size = ggml_quantize_chunk(new_type, f32_data, new_data, 0, n_elms/cur->ne[0], cur->ne[0], nullptr); } else { new_type = cur->type; new_data = cur->data; @@ -1993,17 +1951,6 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i { printf("%s: original size = %8.2f MB\n", __func__, total_size_org / 1024.0 / 1024.0); printf("%s: quantized size = %8.2f MB\n", __func__, total_size_new / 1024.0 / 1024.0); - - int64_t sum_all = 0; - for (size_t i = 0; i < hist_all.size(); ++i) { - sum_all += hist_all[i]; - } - - printf("%s: hist: ", __func__); - for (size_t i = 0; i < hist_all.size(); ++i) { - printf("%5.3f ", hist_all[i] / (float)sum_all); - } - printf("\n"); } return true; |