@fugood/llama.node 0.3.1 → 0.3.3

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

@@ -1,6 +1,7 @@
 // Unit tests for quantization specific functions - quantize, dequantize and dot product
 
 #include "ggml.h"
+#include "ggml-cpu.h"
 
 #undef NDEBUG
 #include <assert.h>
@@ -15,11 +16,13 @@
 
 constexpr float MAX_QUANTIZATION_REFERENCE_ERROR = 0.0001f;
 constexpr float MAX_QUANTIZATION_TOTAL_ERROR = 0.002f;
+constexpr float MAX_QUANTIZATION_TOTAL_ERROR_TERNARY = 0.01f;
 constexpr float MAX_QUANTIZATION_TOTAL_ERROR_2BITS = 0.0075f;
 constexpr float MAX_QUANTIZATION_TOTAL_ERROR_3BITS = 0.0040f;
 constexpr float MAX_QUANTIZATION_TOTAL_ERROR_3BITS_XXS = 0.0050f;
 constexpr float MAX_DOT_PRODUCT_ERROR = 0.02f;
 constexpr float MAX_DOT_PRODUCT_ERROR_LOWBIT = 0.04f;
+constexpr float MAX_DOT_PRODUCT_ERROR_TERNARY = 0.15f;
 
 static const char* RESULT_STR[] = {"ok", "FAILED"};
 
@@ -42,26 +45,27 @@ static float array_rmse(const float * a1, const float * a2, size_t 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) {
+static float total_quantization_error(const ggml_type_traits * qfns, const ggml_type_traits_cpu * qfns_cpu, 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);
+    qfns_cpu->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) {
+static float reference_quantization_error(const ggml_type_traits * qfns, const ggml_type_traits_cpu * qfns_cpu, 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);
+    // FIXME: why is done twice?
+    qfns_cpu->from_float(test_data, tmp_q.data(), test_size);
+    qfns->to_float(tmp_q.data(), tmp_out.data(), test_size);
 
-    qfns.from_float_ref(test_data, tmp_q.data(), test_size);
-    qfns.to_float(tmp_q.data(), tmp_out_ref.data(), test_size);
+    qfns->from_float_ref(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);
 }
@@ -76,18 +80,18 @@ static float dot_product(const float * a1, const float * a2, size_t test_size) {
 
 // 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
+    const ggml_type_traits * qfns, const ggml_type_traits_cpu * qfns_cpu, 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);
+    const auto * vdot = ggml_get_type_traits_cpu(qfns_cpu->vec_dot_type);
 
-    qfns.from_float(test_data1, tmp_q1.data(), test_size);
-    vdot.from_float(test_data2, tmp_q2.data(), test_size);
+    qfns_cpu->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, 0, tmp_q1.data(), 0, tmp_q2.data(), 0, 1);
+    qfns_cpu->vec_dot(test_size, &result, 0, tmp_q1.data(), 0, tmp_q2.data(), 0, 1);
 
     const float dot_ref = dot_product(test_data1, test_data2, test_size);
 
@@ -129,10 +133,11 @@ int main(int argc, char * argv[]) {
 
     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);
+        const auto * qfns = ggml_get_type_traits(type);
+        const auto * qfns_cpu = ggml_get_type_traits_cpu(type);
 
         // deprecated - skip
-        if (qfns.blck_size == 0) {
+        if (qfns->blck_size == 0) {
             continue;
         }
 
@@ -141,9 +146,11 @@ int main(int argc, char * argv[]) {
         printf("Testing %s\n", ggml_type_name((ggml_type) i));
         ggml_quantize_init(ei);
 
-        if (qfns.from_float && qfns.to_float) {
-            const float total_error = total_quantization_error(qfns, test_size, test_data.data());
+        if (qfns_cpu->from_float && qfns->to_float) {
+            const float total_error = total_quantization_error(qfns, qfns_cpu, test_size, test_data.data());
             const float max_quantization_error =
+                type == GGML_TYPE_TQ1_0   ? MAX_QUANTIZATION_TOTAL_ERROR_TERNARY :
+                type == GGML_TYPE_TQ2_0   ? MAX_QUANTIZATION_TOTAL_ERROR_TERNARY :
                 type == GGML_TYPE_Q2_K    ? MAX_QUANTIZATION_TOTAL_ERROR_2BITS :
                 type == GGML_TYPE_IQ2_S   ? MAX_QUANTIZATION_TOTAL_ERROR_2BITS :
                 type == GGML_TYPE_Q3_K    ? MAX_QUANTIZATION_TOTAL_ERROR_3BITS :
@@ -155,17 +162,19 @@ int main(int argc, char * argv[]) {
                 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());
+            const float reference_error = reference_quantization_error(qfns, qfns_cpu, 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());
+            const float vec_dot_error = dot_product_error(qfns, qfns_cpu, test_size, test_data.data(), test_data2.data());
             const float max_allowed_error = type == GGML_TYPE_Q2_K || type == GGML_TYPE_IQ2_XS || type == GGML_TYPE_IQ2_XXS ||
                                             type == GGML_TYPE_IQ3_XXS || type == GGML_TYPE_IQ3_S || type == GGML_TYPE_IQ2_S
                                           ? MAX_DOT_PRODUCT_ERROR_LOWBIT
+                                          : type == GGML_TYPE_TQ1_0 || type == GGML_TYPE_TQ2_0
+                                          ? MAX_DOT_PRODUCT_ERROR_TERNARY
                                           : MAX_DOT_PRODUCT_ERROR;
             failed = !(vec_dot_error < max_allowed_error);
             num_failed += failed;

package/src/llama.cpp/tests/test-quantize-perf.cpp

@@ -1,12 +1,12 @@
 // Benchmark quantization specific functions on synthetic data
 
 #include "ggml.h"
+#include "ggml-cpu.h"
 
 #undef NDEBUG
 #include <algorithm>
 #include <assert.h>
 #include <functional>
-#include <inttypes.h>
 #include <math.h>
 #include <memory>
 #include <stdio.h>
@@ -122,9 +122,10 @@ static void usage(char * argv[]) {
     printf("  --type TYPE           set test type as");
     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);
+        const auto * qfns     = ggml_get_type_traits(type);
+        const auto * qfns_cpu = ggml_get_type_traits_cpu(type);
         if (ggml_type_name(type) != NULL) {
-            if (qfns.from_float && qfns.to_float) {
+            if (qfns_cpu->from_float && qfns->to_float) {
                 printf(" %s", ggml_type_name(type));
             }
         }
@@ -270,12 +271,13 @@ int main(int argc, char * argv[]) {
 
     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);
+        const auto * qfns = ggml_get_type_traits(type);
+        const auto * qfns_cpu = ggml_get_type_traits_cpu(type);
         if (!params.include_types.empty() && ggml_type_name(type) && std::find(params.include_types.begin(), params.include_types.end(), ggml_type_name(type)) == params.include_types.end()) {
             continue;
         }
 
-        if (qfns.from_float && qfns.to_float) {
+        if (qfns_cpu->from_float && qfns->to_float) {
             printf("%s\n", ggml_type_name(type));
 
             ggml_quantize_init(type);
@@ -285,7 +287,7 @@ int main(int argc, char * argv[]) {
                 for (size_t size : params.test_sizes) {
                     printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
                     auto quantize_fn = [&](void) -> float {
-                        qfns.from_float_ref(test_data1, test_q1, size);
+                        qfns->from_float_ref(test_data1, test_q1, size);
                         return test_q1[0];
                     };
                     size_t quantized_size = ggml_row_size(type, size);
@@ -299,7 +301,7 @@ int main(int argc, char * argv[]) {
                 for (size_t size : params.test_sizes) {
                     printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
                     auto quantize_fn = [&](void) -> float {
-                        qfns.from_float(test_data1, test_q1, size);
+                        qfns_cpu->from_float(test_data1, test_q1, size);
                         return test_q1[0];
                     };
                     size_t quantized_size = ggml_row_size(type, size);
@@ -310,11 +312,11 @@ int main(int argc, char * argv[]) {
 
             if (params.op_dequantize_row_q) {
                 printf("  dequantize_row_q\n");
-                qfns.from_float(test_data1, test_q1, largest);
+                qfns_cpu->from_float(test_data1, test_q1, largest);
                 for (size_t size : params.test_sizes) {
                     printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
                     auto quantize_fn = [&](void) -> float {
-                        qfns.to_float(test_q1, test_out, size);
+                        qfns->to_float(test_q1, test_out, size);
                         return test_out[0];
                     };
                     size_t quantized_size = ggml_row_size(type, size);
@@ -328,8 +330,8 @@ int main(int argc, char * argv[]) {
                 for (size_t size : params.test_sizes) {
                     printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
                     auto quantize_fn = [&](void) -> float {
-                        auto vdot = ggml_internal_get_type_traits(qfns.vec_dot_type);
-                        vdot.from_float(test_data1, test_q1, size);
+                        const auto * vdot = ggml_get_type_traits_cpu(qfns_cpu->vec_dot_type);
+                        vdot->from_float(test_data1, test_q1, size);
                         return test_q1[0];
                     };
                     size_t quantized_size = ggml_row_size(type, size);
@@ -340,13 +342,13 @@ int main(int argc, char * argv[]) {
 
             if (params.op_vec_dot_q) {
                 printf("  vec_dot_q\n");
-                qfns.from_float(test_data1, test_q1, largest);
-                qfns.from_float(test_data2, test_q2, largest);
+                qfns_cpu->from_float(test_data1, test_q1, largest);
+                qfns_cpu->from_float(test_data2, test_q2, largest);
                 for (size_t size : params.test_sizes) {
                     printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
                     auto quantize_fn = [&](void) -> float {
                         float result;
-                        qfns.vec_dot(size, &result, 0, test_q1, 0, test_q2, 0, 1);
+                        qfns_cpu->vec_dot(size, &result, 0, test_q1, 0, test_q2, 0, 1);
                         return result;
                     };
                     size_t quantized_size = ggml_row_size(type, size);

package/src/llama.cpp/tests/test-rope.cpp

@@ -1,4 +1,5 @@
 #include "ggml.h"
+#include "ggml-cpu.h"
 
 #include <cmath>
 #include <cstdio>
@@ -113,7 +114,7 @@ static struct ggml_tensor * get_random_tensor_f32(
 }
 
 static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
-    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
+    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads, nullptr);
 
     if (plan.work_size > 0) {
         buf.resize(plan.work_size);