vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

use higher eps only for the quants that need it

Browse source 
ggml-ci
This commit is contained in:
slaren 2024-05-18 01:15:34 +02:00
parent f59edeeae9
commit f07e570c03
2 changed files with 21 additions and 8 deletions
Download patch file Download diff file Expand all files Collapse all files

18
ggml-quants.c
View file

@ -14,7 +14,11 @@
#include <stdlib.h> // for qsort
#include <stdio.h> // for GGML_ASSERT
#define GROUP_MAX_EPS 1e-7f
#define GROUP_MAX_EPS 1e-15f
#define GROUP_MAX_EPS_IQ3_XXS 1e-8f
#define GROUP_MAX_EPS_IQ2_S 1e-8f
#define GROUP_MAX_EPS_IQ1_M 1e-7f
#define GROUP_MAX_EPS_IQ1_S 1e-12f
#if defined(_MSC_VER)
// disable "possible loss of data" to avoid warnings for hundreds of casts
@ -1648,7 +1652,7 @@ static float make_qp_quants(int n, int nmax, const float * restrict x, uint8_t *
break;
}
}
return sumlx / suml2;
return sumlx/suml2;
}
static void quantize_row_q2_K_impl(const float * restrict x, block_q2_K * restrict y, int k, const float * restrict quant_weights) {
@ -12598,7 +12602,7 @@ static void quantize_row_iq2_xxs_impl(const float * restrict x, void * restrict
}
float max = xval[0];
for (int i = 1; i < 32; ++i) max = MAX(max, xval[i]);
if (!max) {
if (max < GROUP_MAX_EPS) {
scales[ib] = 0;
memset(L, 0, 32);
continue;
@ -13215,7 +13219,7 @@ static void quantize_row_iq3_xxs_impl(int grid_size, const float * restrict x, v
}
float max = xval[0];
for (int i = 1; i < 32; ++i) max = MAX(max, xval[i]);
if (max < GROUP_MAX_EPS) {
if (max < GROUP_MAX_EPS_IQ3_XXS) {
scales[ib] = 0;
memset(L, 0, 32);
continue;
@ -13755,7 +13759,7 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy
for (int i = 0; i < block_size; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
float max = fabsf(xb[0]);
for (int i = 1; i < block_size; ++i) max = MAX(max, fabsf(xb[i]));
if (max < GROUP_MAX_EPS) {
if (max < GROUP_MAX_EPS_IQ1_S) {
scales[ib] = 0;
memset(L, 1, block_size);
continue;
@ -13943,7 +13947,7 @@ static void quantize_row_iq1_m_impl(const float * restrict x, void * restrict vy
}
float max = fabsf(xb[0]);
for (int i = 1; i < block_size; ++i) max = MAX(max, fabsf(xb[i]));
if (max < GROUP_MAX_EPS) {
if (max < GROUP_MAX_EPS_IQ1_M) {
scales[ib] = 0;
memset(L, 1, block_size);
continue;
@ -14428,7 +14432,7 @@ static void quantize_row_iq2_s_impl(const float * restrict x, void * restrict vy
}
float max = xval[0];
for (int i = 1; i < 16; ++i) max = MAX(max, xval[i]);
if (max < GROUP_MAX_EPS) {
if (max < GROUP_MAX_EPS_IQ2_S) {
scales[ib] = 0;
continue;
}

11
tests/test-backend-ops.cpp
View file

@ -16,6 +16,7 @@
#include <thread>
#include <vector>
static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float max = 1.0f) {
// static RNG initialization (revisit if n_threads stops being constant)
static const size_t n_threads = std::thread::hardware_concurrency();
@ -50,10 +51,17 @@ static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float m
}
#if 0
const char * val_str = getenv("GGML_TEST_EPS");
float val = 1e-9f;
if (val_str != nullptr) {
val = std::stof(val_str);
printf("GGML_TEST_EPS=%e\n", val);
}
// test quantization with very small values that may result in nan scales due to division by zero
if (ggml_is_quantized(tensor->type)) {
for (int i = 0; i < 256; i++) {
data[i] = 1e-7f;
data[i] = val;
}
}
#endif
@ -73,6 +81,7 @@ static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float m
}
}
ggml_quantize_chunk(tensor->type, data.data(), dataq.data(), 0, size/tensor->ne[0], tensor->ne[0], im);
GGML_ASSERT(ggml_validate_row_data(tensor->type, dataq.data(), dataq.size()));
ggml_backend_tensor_set(tensor, dataq.data(), 0, dataq.size());
} else if (tensor->type == GGML_TYPE_I8 || tensor->type == GGML_TYPE_I16 || tensor->type == GGML_TYPE_I32) {
// This is going to create some weird integers though.