Windows test

Author: ivanstepanovftw

.github/workflows/build.yml

@@ -190,7 +190,7 @@ jobs:
         if: ${{ matrix.build != 'avx512' || env.HAS_AVX512F == '1' }} # Test AVX-512 only when possible
         run: |
           cd build
-          ctest -C Release --verbose
+          ctest -C Release --verbose --output-on-failure
 
       - name: Get commit hash
         id: commit

tests/test-sampling.cpp

@@ -1,14 +1,12 @@
 #include "llama.h"
-#include <assert.h>
-#include <math.h>
+#include <cassert>
+#include <cmath>
 #include <numeric>
 #include <cassert>
 #include <iostream>
 #include <vector>
 #include <algorithm>
 
-#undef assert
-#define assert(__expr) do { if (!(__expr)) { printf("%s:%d (%s) %s\n", __FILE__, __LINE__, __func__, #__expr); exit(1); } } while(0)
 
 void dump(const llama_token_data_array * candidates) {
     for (size_t i = 0; i < candidates->size; i++) {
@@ -22,19 +20,28 @@ void dump(const llama_token_data_array * candidates) {
 void test_top_k(const std::vector<float> & probs,
                 const std::vector<float> & expected_probs,
                 int k) {
+    printf("%s:%d (%s)\n", __FILE__, __LINE__, __func__);
+    fflush(stdout);
     size_t n_vocab = probs.size();
     std::vector<llama_token_data> candidates;
     candidates.reserve(n_vocab);
     for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
-        float logit = log(probs[token_id]);
+        printf("%s:%d (%s) token_id: %d\n", __FILE__, __LINE__, __func__, token_id);
+        fflush(stdout);
+        float logit = std::log(probs[token_id]);
         candidates.emplace_back(llama_token_data{token_id, logit, 0.0f});
     }
+    printf("%s:%d (%s)\n", __FILE__, __LINE__, __func__);
+    fflush(stdout);
 
     llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
+    printf("%s:%d (%s)\n", __FILE__, __LINE__, __func__);
+    fflush(stdout);
     llama_sample_softmax(nullptr, &candidates_p);
-    // DUMP(&candidates_p);
+    DUMP(&candidates_p);
     llama_sample_top_k(nullptr, &candidates_p, k);
-    // DUMP(&candidates_p);
+    DUMP(&candidates_p);
+    fflush(stdout);
 
     assert(candidates_p.size == expected_probs.size());
     for (size_t i = 0; i < candidates_p.size; i++) {
@@ -57,9 +64,9 @@ void test_top_p(const std::vector<float> & probs,
 
     llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
     llama_sample_softmax(nullptr, &candidates_p);
-    // DUMP(&candidates_p);
+    DUMP(&candidates_p);
     llama_sample_top_p(nullptr, &candidates_p, p);
-    // DUMP(&candidates_p);
+    DUMP(&candidates_p);
 
     assert(candidates_p.size == expected_probs.size());
     for (size_t i = 0; i < candidates_p.size; i++) {
@@ -80,9 +87,9 @@ void test_tfs(const std::vector<float> & probs,
     }
 
     llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-    // DUMP(&candidates_p);
+    DUMP(&candidates_p);
     llama_sample_tail_free(nullptr, &candidates_p, z);
-    // DUMP(&candidates_p);
+    DUMP(&candidates_p);
 
     assert(candidates_p.size == expected_probs.size());
     for (size_t i = 0; i < candidates_p.size; i++) {
@@ -103,9 +110,9 @@ void test_typical(const std::vector<float> & probs,
     }
 
     llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-    // DUMP(&candidates_p);
+    DUMP(&candidates_p);
     llama_sample_typical(nullptr, &candidates_p, p);
-    // DUMP(&candidates_p);
+    DUMP(&candidates_p);
 
     assert(candidates_p.size == expected_probs.size());
     for (size_t i = 0; i < candidates_p.size; i++) {
@@ -172,6 +179,8 @@ void test_frequency_presence_penalty(
 }
 
 int main(void) {
+    printf("main\n");
+    fflush(stdout);
     test_top_k({0.1, 0.2, 0.3, 0.4}, {0.4}, 1);
     test_top_k({0.1, 0.2, 0.3, 0.4}, {0.4, 0.3, 0.2}, 3);