char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

Rohanjames1997tests: extend test-quantize-fns to test nrc=2 (i8mm) kernels (#16234)982937a33

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
10.9 KiB273 linesraw
1// Unit tests for quantization specific functions - quantize, dequantize and dot product
2
3#include "ggml.h"
4#include "ggml-cpu.h"
5
6#undef NDEBUG
7#include <assert.h>
8#include <algorithm>
9#include <cmath>
10#include <math.h>
11#include <stdio.h>
12#include <string>
13#include <vector>
14
15#if defined(_MSC_VER)
16#pragma warning(disable: 4244 4267) // possible loss of data
17#endif
18
19constexpr float MAX_QUANTIZATION_REFERENCE_ERROR = 0.0001f;
20constexpr float MAX_QUANTIZATION_TOTAL_ERROR = 0.002f;
21constexpr float MAX_QUANTIZATION_TOTAL_ERROR_BINARY = 0.025f;
22constexpr float MAX_QUANTIZATION_TOTAL_ERROR_TERNARY = 0.01f;
23constexpr float MAX_QUANTIZATION_TOTAL_ERROR_2BITS = 0.0075f;
24constexpr float MAX_QUANTIZATION_TOTAL_ERROR_3BITS = 0.0040f;
25constexpr float MAX_QUANTIZATION_TOTAL_ERROR_3BITS_XXS = 0.0050f;
26constexpr float MAX_QUANTIZATION_TOTAL_ERROR_FP4 = 0.0030f;
27constexpr float MAX_DOT_PRODUCT_ERROR = 0.02f;
28constexpr float MAX_DOT_PRODUCT_ERROR_LOWBIT = 0.04f;
29constexpr float MAX_DOT_PRODUCT_ERROR_FP4 = 0.03f;
30constexpr float MAX_DOT_PRODUCT_ERROR_BINARY = 0.40f;
31constexpr float MAX_DOT_PRODUCT_ERROR_TERNARY = 0.15f;
32
33static const char* RESULT_STR[] = {"ok", "FAILED"};
34
35
36// Generate synthetic data
37static void generate_data(float offset, size_t n, float * dst, float amplitude = 2.0f) {
38    for (size_t i = 0; i < n; i++) {
39        dst[i] = 0.1 + amplitude*cosf(i + offset);
40    }
41}
42
43// Calculate RMSE between two float arrays
44static float array_rmse(const float * a1, const float * a2, size_t n) {
45    double sum = 0;
46    for (size_t i = 0; i < n; i++) {
47        double diff = a1[i] - a2[i];
48        sum += diff * diff;
49    }
50    return sqrtf(sum) / n;
51}
52
53// Total quantization error on test data
54static float total_quantization_error(const ggml_type_traits * qfns, const ggml_type_traits_cpu * qfns_cpu, size_t test_size, const float * test_data) {
55    std::vector<uint8_t> tmp_q(2*test_size);
56    std::vector<float> tmp_out(test_size);
57
58    qfns_cpu->from_float(test_data, tmp_q.data(), test_size);
59    qfns->to_float(tmp_q.data(), tmp_out.data(), test_size);
60    return array_rmse(test_data, tmp_out.data(), test_size);
61}
62
63// Total quantization error on test data
64static float reference_quantization_error(const ggml_type_traits * qfns, const ggml_type_traits_cpu * qfns_cpu, size_t test_size, const float * test_data) {
65    std::vector<uint8_t> tmp_q(2*test_size);
66    std::vector<float> tmp_out(test_size);
67    std::vector<float> tmp_out_ref(test_size);
68
69    // FIXME: why is done twice?
70    qfns_cpu->from_float(test_data, tmp_q.data(), test_size);
71    qfns->to_float(tmp_q.data(), tmp_out.data(), test_size);
72
73    qfns->from_float_ref(test_data, tmp_q.data(), test_size);
74    qfns->to_float(tmp_q.data(), tmp_out_ref.data(), test_size);
75
76    return array_rmse(tmp_out.data(), tmp_out_ref.data(), test_size);
77}
78
79static float dot_product(const float * a1, const float * a2, size_t test_size) {
80    double sum = 0;
81    for (size_t i = 0; i < test_size; i++) {
82        sum += a1[i] * a2[i];
83    }
84    return sum;
85}
86
87// Total dot product error
88static float dot_product_error(const ggml_type_traits_cpu * qfns_cpu, ggml_type src0_type, size_t test_size,
89                               const float * test_data1, const float * test_data2,
90                               const float * test_data3, const float * test_data4,
91                               const int nrc) {
92    const auto * vdot = ggml_get_type_traits_cpu(qfns_cpu->vec_dot_type);
93    const size_t pad  = 64;
94    const size_t bx   = ggml_row_size(src0_type, test_size) + pad;
95    const size_t by   = ggml_row_size(qfns_cpu->vec_dot_type, test_size) + pad;
96
97    std::vector<uint8_t> tmp_q1(bx * nrc);
98    std::vector<uint8_t> tmp_q2(by * nrc);
99
100    qfns_cpu->from_float(test_data1, tmp_q1.data(), test_size);
101    vdot->from_float(test_data2, tmp_q2.data(), test_size);
102
103    if (nrc == 1) {
104        float result = INFINITY;
105        qfns_cpu->vec_dot(test_size, &result, 0, tmp_q1.data(), 0, tmp_q2.data(), 0, 1);
106
107        const float dot_ref = dot_product(test_data1, test_data2, test_size);
108        return fabsf(result - dot_ref) / test_size;
109    }
110
111    // nrc == 2: kernel computes a 2x2 dot product matrix
112    // Output layout: s[0]=dot(vx0,vy0), s[1]=dot(vx1,vy0), s[bs]=dot(vx0,vy1), s[bs+1]=dot(vx1,vy1)
113    // row and output strides are padded, same as in the mul_mat path
114    qfns_cpu->from_float(test_data3, tmp_q1.data() + bx, test_size);
115    vdot->from_float(test_data4, tmp_q2.data() + by, test_size);
116
117    const size_t bs = 16;
118    std::vector<float> result(bs + 2, INFINITY);
119    qfns_cpu->vec_dot(test_size, result.data(), bs, tmp_q1.data(), bx, tmp_q2.data(), by, 2);
120
121    const float ref00 = dot_product(test_data1, test_data2, test_size);
122    const float ref10 = dot_product(test_data3, test_data2, test_size);
123    const float ref01 = dot_product(test_data1, test_data4, test_size);
124    const float ref11 = dot_product(test_data3, test_data4, test_size);
125
126    const auto err = [test_size](float val, float ref) {
127        const float e = fabsf(val - ref) / test_size;
128        return std::isfinite(e) ? e : INFINITY;
129    };
130
131    return std::max({err(result[0], ref00), err(result[1], ref10), err(result[bs], ref01), err(result[bs + 1], ref11)});
132}
133
134static int test_vec_dot_f32(bool verbose) {
135    const auto * f32 = ggml_get_type_traits_cpu(GGML_TYPE_F32);
136    int num_failed = 0;
137    for (int n : {1, 2, 3, 5, 7, 8, 15, 16, 17, 31, 33, 63, 67, 127, 129, 193, 255, 1023}) {
138        std::vector<float> a(n);
139        std::vector<float> b(n);
140        generate_data(0.0, n, a.data());
141        generate_data(1.0, n, b.data());
142
143        float result = 0.0f;
144        f32->vec_dot(n, &result, 0, a.data(), 0, b.data(), 0, 1);
145        const float ref = dot_product(a.data(), b.data(), n);
146        const float error = fabsf(result - ref) / n;
147
148        const bool failed = !(error < MAX_QUANTIZATION_REFERENCE_ERROR);
149        num_failed += failed;
150        if (failed || verbose) {
151            printf(" f32 vec_dot n=%4d:                 %s (ref=%f got=%f err=%f)\n",
152                   n, RESULT_STR[failed], ref, result, error);
153        }
154    }
155    return num_failed;
156}
157
158static int test_vec_dot_q(bool verbose) {
159    int num_failed = 0;
160
161    const size_t test_size = 32 * 128;
162
163    std::vector<float> test_data(test_size);
164    std::vector<float> test_data2(test_size);
165    std::vector<float> test_data3(test_size);
166    std::vector<float> test_data4(test_size);
167
168    generate_data(0.0, test_data.size(), test_data.data());
169    generate_data(1.0, test_data2.size(), test_data2.data());
170    generate_data(3.0, test_data3.size(), test_data3.data(), 1.0f);
171    generate_data(4.0, test_data4.size(), test_data4.data(), 1.5f);
172
173    for (int i = 0; i < GGML_TYPE_COUNT; i++) {
174        ggml_type type = (ggml_type) i;
175        const auto * qfns = ggml_get_type_traits(type);
176        const auto * qfns_cpu = ggml_get_type_traits_cpu(type);
177
178        // deprecated - skip
179        if (qfns->blck_size == 0) {
180            continue;
181        }
182
183        const ggml_type ei = (ggml_type)i;
184
185        printf("Testing %s\n", ggml_type_name((ggml_type) i));
186        ggml_quantize_init(ei);
187
188        if (qfns_cpu->from_float && qfns->to_float) {
189            const float total_error = total_quantization_error(qfns, qfns_cpu, test_size, test_data.data());
190            const float max_quantization_error =
191                type == GGML_TYPE_Q1_0    ? MAX_QUANTIZATION_TOTAL_ERROR_BINARY :
192                type == GGML_TYPE_TQ1_0   ? MAX_QUANTIZATION_TOTAL_ERROR_TERNARY :
193                type == GGML_TYPE_TQ2_0   ? MAX_QUANTIZATION_TOTAL_ERROR_TERNARY :
194                type == GGML_TYPE_Q2_0    ? MAX_QUANTIZATION_TOTAL_ERROR_TERNARY :
195                type == GGML_TYPE_Q2_K    ? MAX_QUANTIZATION_TOTAL_ERROR_2BITS :
196                type == GGML_TYPE_IQ2_S   ? MAX_QUANTIZATION_TOTAL_ERROR_2BITS :
197                type == GGML_TYPE_Q3_K    ? MAX_QUANTIZATION_TOTAL_ERROR_3BITS :
198                type == GGML_TYPE_IQ3_S   ? MAX_QUANTIZATION_TOTAL_ERROR_3BITS :
199                type == GGML_TYPE_IQ3_XXS ? MAX_QUANTIZATION_TOTAL_ERROR_3BITS_XXS :
200                type == GGML_TYPE_NVFP4   ? MAX_QUANTIZATION_TOTAL_ERROR_FP4 : MAX_QUANTIZATION_TOTAL_ERROR;
201            bool failed = !(total_error < max_quantization_error);
202            num_failed += failed;
203            if (failed || verbose) {
204                printf("%5s absolute quantization error:    %s (%f)\n", ggml_type_name(type), RESULT_STR[failed], total_error);
205            }
206
207            const float reference_error = reference_quantization_error(qfns, qfns_cpu, test_size, test_data.data());
208            failed = !(reference_error < MAX_QUANTIZATION_REFERENCE_ERROR);
209            num_failed += failed;
210            if (failed || verbose) {
211                printf("%5s reference implementation error: %s (%f)\n", ggml_type_name(type), RESULT_STR[failed], reference_error);
212            }
213
214            const float vec_dot_error = dot_product_error(qfns_cpu, type, test_size, test_data.data(), test_data2.data(), nullptr, nullptr, 1);
215            const float max_allowed_error = type == GGML_TYPE_Q2_K || type == GGML_TYPE_IQ2_XS || type == GGML_TYPE_IQ2_XXS ||
216                type == GGML_TYPE_IQ3_XXS || type == GGML_TYPE_IQ3_S || type == GGML_TYPE_IQ2_S
217                ? MAX_DOT_PRODUCT_ERROR_LOWBIT
218                : type == GGML_TYPE_Q1_0
219                ? MAX_DOT_PRODUCT_ERROR_BINARY
220                : type == GGML_TYPE_TQ1_0 || type == GGML_TYPE_TQ2_0 || type == GGML_TYPE_Q2_0
221                ? MAX_DOT_PRODUCT_ERROR_TERNARY
222                : type == GGML_TYPE_NVFP4
223                ? MAX_DOT_PRODUCT_ERROR_FP4
224                : MAX_DOT_PRODUCT_ERROR;
225            failed = !(vec_dot_error < max_allowed_error);
226            num_failed += failed;
227            if (failed || verbose) {
228                printf("%5s dot product error:              %s (%f)\n", ggml_type_name(type), RESULT_STR[failed], vec_dot_error);
229            }
230
231            // Test nrc=2 path for types that support it
232            if (qfns_cpu->nrows == 2) {
233                const float vec_dot_error_nrc2 = dot_product_error(qfns_cpu, type, test_size, test_data.data(), test_data2.data(), test_data3.data(), test_data4.data(), 2);
234                failed = !(vec_dot_error_nrc2 < max_allowed_error);
235                num_failed += failed;
236                if (failed || verbose) {
237                    printf("%5s dot product error (nrc=2):    %s (%f)\n", ggml_type_name(type), RESULT_STR[failed], vec_dot_error_nrc2);
238                }
239            }
240        }
241    }
242
243    return num_failed;
244}
245
246int main(int argc, char * argv[]) {
247    bool verbose = false;
248
249    std::string arg;
250    for (int i = 1; i < argc; i++) {
251        arg = argv[i];
252
253        if (arg == "-v") {
254            verbose = true;
255        } else {
256            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
257            return 1;
258        }
259    }
260
261    ggml_cpu_init();
262
263    int num_failed = 0;
264
265    num_failed += test_vec_dot_f32(verbose);
266    num_failed += test_vec_dot_q(verbose);
267
268    if (num_failed || verbose) {
269        printf("%d tests failed\n", num_failed);
270    }
271
272    return num_failed > 0;
273}