llama.cpp/tests/test-quantize-fns.cpp

167 lines
5.6 KiB
C++
Raw Normal View History

// Unit tests for quantization specific functions - quantize, dequantize and dot product
#include "ggml.h"
#undef NDEBUG
#include <assert.h>
#include <math.h>
#include <stdio.h>
#include <string>
#include <vector>
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
constexpr float MAX_QUANTIZATION_REFERENCE_ERROR = 0.0001f;
constexpr float MAX_QUANTIZATION_TOTAL_ERROR = 0.002f;
constexpr float MAX_QUANTIZATION_TOTAL_ERROR_2BITS = 0.0075f;
constexpr float MAX_QUANTIZATION_TOTAL_ERROR_3BITS = 0.0040f;
constexpr float MAX_DOT_PRODUCT_ERROR = 0.02f;
static const char* RESULT_STR[] = {"ok", "FAILED"};
// Generate synthetic data
static void generate_data(float offset, size_t n, float * dst) {
for (size_t i = 0; i < n; i++) {
dst[i] = 0.1 + 2*cosf(i + offset);
}
}
// Calculate RMSE between two float arrays
static float array_rmse(const float * a1, const float * a2, size_t n) {
double sum = 0;
for (size_t i = 0; i < n; i++) {
double diff = a1[i] - a2[i];
sum += diff * diff;
}
return sqrtf(sum) / n;
}
// Total quantization error on test data
static float total_quantization_error(ggml_type_traits_t & qfns, size_t test_size, const float * test_data) {
std::vector<uint8_t> tmp_q(2*test_size);
std::vector<float> tmp_out(test_size);
qfns.from_float(test_data, tmp_q.data(), test_size);
qfns.to_float(tmp_q.data(), tmp_out.data(), test_size);
return array_rmse(test_data, tmp_out.data(), test_size);
}
// Total quantization error on test data
static float reference_quantization_error(ggml_type_traits_t & qfns, size_t test_size, const float * test_data) {
std::vector<uint8_t> tmp_q(2*test_size);
std::vector<float> tmp_out(test_size);
std::vector<float> tmp_out_ref(test_size);
qfns.from_float(test_data, tmp_q.data(), test_size);
qfns.to_float(tmp_q.data(), tmp_out.data(), test_size);
qfns.from_float_reference(test_data, tmp_q.data(), test_size);
qfns.to_float(tmp_q.data(), tmp_out_ref.data(), test_size);
return array_rmse(tmp_out.data(), tmp_out_ref.data(), test_size);
}
static float dot_product(const float * a1, const float * a2, size_t test_size) {
double sum = 0;
for (size_t i = 0; i < test_size; i++) {
sum += a1[i] * a2[i];
}
return sum;
}
// Total dot product error
static float dot_product_error(
ggml_type_traits_t & qfns, size_t test_size, const float * test_data1, const float *test_data2
) {
std::vector<uint8_t> tmp_q1(2*test_size);
std::vector<uint8_t> tmp_q2(2*test_size);
auto vdot = ggml_internal_get_type_traits(qfns.vec_dot_type);
qfns.from_float(test_data1, tmp_q1.data(), test_size);
vdot.from_float(test_data2, tmp_q2.data(), test_size);
float result = INFINITY;
qfns.vec_dot(test_size, &result, tmp_q1.data(), tmp_q2.data());
const float dot_ref = dot_product(test_data1, test_data2, test_size);
return fabsf(result - dot_ref) / test_size;
}
int main(int argc, char * argv[]) {
bool verbose = false;
const size_t test_size = 32 * 128;
std::string arg;
for (int i = 1; i < argc; i++) {
arg = argv[i];
if (arg == "-v") {
verbose = true;
} else {
fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
return 1;
}
}
std::vector<float> test_data(test_size);
std::vector<float> test_data2(test_size);
generate_data(0.0, test_data.size(), test_data.data());
generate_data(1.0, test_data2.size(), test_data2.data());
// Initialize GGML, ensures float conversion tables are initialized
struct ggml_init_params ggml_params = {
/* .mem_size = */ 1*1024,
/* .mem_buffer = */ NULL,
/* .no_alloc = */ true,
};
struct ggml_context * ctx = ggml_init(ggml_params);
int num_failed = 0;
bool failed = false;
for (int i = 0; i < GGML_TYPE_COUNT; i++) {
ggml_type type = (ggml_type) i;
ggml_type_traits_t qfns = ggml_internal_get_type_traits(type);
if (qfns.from_float && qfns.to_float) {
const float total_error = total_quantization_error(qfns, test_size, test_data.data());
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
const float max_quantization_error =
type == GGML_TYPE_Q2_K ? MAX_QUANTIZATION_TOTAL_ERROR_2BITS :
type == GGML_TYPE_Q3_K ? MAX_QUANTIZATION_TOTAL_ERROR_3BITS : MAX_QUANTIZATION_TOTAL_ERROR;
failed = !(total_error < max_quantization_error);
num_failed += failed;
if (failed || verbose) {
printf("%5s absolute quantization error: %s (%f)\n", ggml_type_name(type), RESULT_STR[failed], total_error);
}
const float reference_error = reference_quantization_error(qfns, test_size, test_data.data());
failed = !(reference_error < MAX_QUANTIZATION_REFERENCE_ERROR);
num_failed += failed;
if (failed || verbose) {
printf("%5s reference implementation error: %s (%f)\n", ggml_type_name(type), RESULT_STR[failed], reference_error);
}
const float vec_dot_error = dot_product_error(qfns, test_size, test_data.data(), test_data2.data());
failed = !(vec_dot_error < MAX_DOT_PRODUCT_ERROR);
num_failed += failed;
if (failed || verbose) {
printf("%5s dot product error: %s (%f)\n", ggml_type_name(type), RESULT_STR[failed], vec_dot_error);
}
}
}
if (num_failed || verbose) {
printf("%d tests failed\n", num_failed);
}
ggml_free(ctx);
return num_failed > 0;
}