diff options
Diffstat (limited to 'tests')
| -rw-r--r-- | tests/test_bf16_model.c | 20 | ||||
| -rw-r--r-- | tests/test_cli.c | 76 | ||||
| -rw-r--r-- | tests/test_kernels.c | 119 | ||||
| -rw-r--r-- | tests/test_late.c | 85 | ||||
| -rw-r--r-- | tests/test_qwen3.c | 25 | ||||
| -rw-r--r-- | tests/test_safetensors.c | 67 | ||||
| -rw-r--r-- | tests/test_server.c | 436 | ||||
| -rw-r--r-- | tests/test_tokenizer.c | 159 | ||||
| -rw-r--r-- | tests/test_workspace.c | 78 | ||||
| -rw-r--r-- | tests/tiny_model.h | 97 | ||||
| -rw-r--r-- | tests/tok_fixture.h | 36 |
11 files changed, 596 insertions, 602 deletions
diff --git a/tests/test_bf16_model.c b/tests/test_bf16_model.c index cc9b2fb..9a5d842 100644 --- a/tests/test_bf16_model.c +++ b/tests/test_bf16_model.c @@ -15,12 +15,14 @@ int main(void) { char root[1024], f32_dir[1088], bf16_dir[1088]; snprintf(root, sizeof(root), "%s/embed-bf16-test-XXXXXX", getenv("TMPDIR") ? getenv("TMPDIR") : "/tmp"); - if (!mkdtemp(root)) { perror("mkdtemp"); return 2; } + if (!mkdtemp(root)) { + perror("mkdtemp"); + return 2; + } snprintf(f32_dir, sizeof(f32_dir), "%s/f32", root); snprintf(bf16_dir, sizeof(bf16_dir), "%s/bf16", root); if (mkdir(f32_dir, 0755) != 0 || mkdir(bf16_dir, 0755) != 0 || - tm_write_model(f32_dir, "F32") != 0 || - tm_write_model(bf16_dir, "BF16") != 0) { + tm_write_model(f32_dir, "F32") != 0 || tm_write_model(bf16_dir, "BF16") != 0) { fprintf(stderr, "fixture creation failed\n"); return 2; } @@ -58,7 +60,8 @@ int main(void) { return 1; } float d = fabsf(a[i] - b[i]); - if (d > max_diff) max_diff = d; + if (d > max_diff) + max_diff = d; norm += b[i] * b[i]; } norm = sqrtf(norm); @@ -73,7 +76,8 @@ int main(void) { * (linear_nobias_weight); parity must hold there too. */ enum { LONG_N = 24 }; int long_ids[LONG_N]; - for (int i = 0; i < LONG_N; i++) long_ids[i] = (i % 14) + 1; + for (int i = 0; i < LONG_N; i++) + long_ids[i] = (i % 14) + 1; if (embed_model_encode_into(mf32, wf32, long_ids, LONG_N, a) != 0 || embed_model_encode_into(mbf16, wbf16, long_ids, LONG_N, b) != 0) { fprintf(stderr, "long-sequence embedding failed\n"); @@ -86,15 +90,15 @@ int main(void) { return 1; } float d = fabsf(a[i] - b[i]); - if (d > long_diff) long_diff = d; + if (d > long_diff) + long_diff = d; } if (long_diff > 2e-5f) { fprintf(stderr, "bad widen-path parity: max_diff=%.9g\n", long_diff); return 1; } - printf("ok: bf16 model parity dim=%d max_abs_diff=%.9g norm=%.9g\n", - dim, max_diff, norm); + printf("ok: bf16 model parity dim=%d max_abs_diff=%.9g norm=%.9g\n", dim, max_diff, norm); embed_workspace_free(wf32); embed_workspace_free(wbf16); diff --git a/tests/test_cli.c b/tests/test_cli.c index 58cbad3..8d56f09 100644 --- a/tests/test_cli.c +++ b/tests/test_cli.c @@ -13,68 +13,69 @@ #include <sys/wait.h> #include <unistd.h> -static char g_dir[1024]; /* fixture model dir */ -static char g_out[1100]; /* captured stdout */ -static char g_err[1100]; /* captured stderr */ -static char g_in[1100]; /* stdin source */ +static char g_dir[1024]; /* fixture model dir */ +static char g_out[1100]; /* captured stdout */ +static char g_err[1100]; /* captured stderr */ +static char g_in[1100]; /* stdin source */ static const char *g_cli; static int g_failures; /* Run the CLI with sh-formatted args; returns its exit status (-1 on a * harness error). stdin_text == NULL feeds /dev/null. */ -static int run_cli(const char *args, const char *stdin_text) -{ +static int run_cli(const char *args, const char *stdin_text) { if (stdin_text) { FILE *f = fopen(g_in, "w"); - if (!f) return -1; + if (!f) + return -1; fputs(stdin_text, f); fclose(f); } char cmd[4096]; - int n = snprintf(cmd, sizeof(cmd), "'%s' %s > '%s' 2> '%s' < '%s'", - g_cli, args, g_out, g_err, + int n = snprintf(cmd, sizeof(cmd), "'%s' %s > '%s' 2> '%s' < '%s'", g_cli, args, g_out, g_err, stdin_text ? g_in : "/dev/null"); - if (n < 0 || (size_t)n >= sizeof(cmd)) return -1; + if (n < 0 || (size_t)n >= sizeof(cmd)) + return -1; int st = system(cmd); - if (st == -1 || !WIFEXITED(st)) return -1; + if (st == -1 || !WIFEXITED(st)) + return -1; return WEXITSTATUS(st); } -static void expect(int cond, const char *what) -{ +static void expect(int cond, const char *what) { if (!cond) { fprintf(stderr, "FAIL: %s\n", what); g_failures++; } } -static int file_contains(const char *path, const char *needle) -{ +static int file_contains(const char *path, const char *needle) { static char buf[1 << 16]; FILE *f = fopen(path, "rb"); - if (!f) return 0; + if (!f) + return 0; size_t n = fread(buf, 1, sizeof(buf) - 1, f); fclose(f); buf[n] = '\0'; return strstr(buf, needle) != NULL; } -static int count_lines(const char *path) -{ +static int count_lines(const char *path) { FILE *f = fopen(path, "rb"); - if (!f) return -1; + if (!f) + return -1; int lines = 0, c; while ((c = fgetc(f)) != EOF) - if (c == '\n') lines++; + if (c == '\n') + lines++; fclose(f); return lines; } /* First stdout line must hold exactly `dim` finite floats. */ -static int stdout_is_embedding(int dim) -{ +static int stdout_is_embedding(int dim) { FILE *f = fopen(g_out, "r"); - if (!f) return 0; + if (!f) + return 0; int count = 0; float v; while (fscanf(f, "%f", &v) == 1) { @@ -88,8 +89,7 @@ static int stdout_is_embedding(int dim) return count == dim; } -int main(int argc, char **argv) -{ +int main(int argc, char **argv) { g_cli = argc > 1 ? argv[1] : "./embed"; tm_dims_t dims = {4, 2, 1, 2, 8, TF_VOCAB_SIZE}; @@ -109,13 +109,11 @@ int main(int argc, char **argv) /* Help and missing-argument handling. */ expect(run_cli("-h", NULL) == 0, "-h exits 0"); expect(run_cli("", NULL) == 1, "no arguments exits 1"); - expect(run_cli("-d /nonexistent-embed-dir hi", NULL) == 1, - "missing model dir exits 1"); + expect(run_cli("-d /nonexistent-embed-dir hi", NULL) == 1, "missing model dir exits 1"); /* -V reports the program's base name, not the path it was invoked by. */ expect(run_cli("-V", NULL) == 0, "-V exits 0"); - expect(!file_contains(g_out, "tests/"), - "-V strips the directory from the program name"); + expect(!file_contains(g_out, "tests/"), "-V strips the directory from the program name"); /* Mode 1: one text prints one embedding line. */ snprintf(args, sizeof(args), "-d '%s' 'hello world'", g_dir); @@ -123,11 +121,9 @@ int main(int argc, char **argv) expect(stdout_is_embedding(dims.hidden), "single text prints dim floats"); /* Thread count. */ - snprintf(args, sizeof(args), - "-d '%s' -t 2 -v 'hello world'", g_dir); + snprintf(args, sizeof(args), "-d '%s' -t 2 -v 'hello world'", g_dir); expect(run_cli(args, NULL) == 0, "-t 2 -v exits 0"); - expect(file_contains(g_err, "Using 2 CPU thread(s)"), - "-v reports the thread count"); + expect(file_contains(g_err, "Using 2 CPU thread(s)"), "-v reports the thread count"); /* Mode 2+: several texts print a bare numeric matrix, one row per line, * with no titles or labels. Two texts give a 2x2 with a 1.0 diagonal. */ @@ -152,8 +148,7 @@ int main(int argc, char **argv) /* -e appends the raw embeddings (label-free) after the matrix: three * matrix rows, a blank separator, then three embedding rows. */ - snprintf(args, sizeof(args), - "-d '%s' -e -b 1 'hello' 'world' 'held'", g_dir); + snprintf(args, sizeof(args), "-d '%s' -e -b 1 'hello' 'world' 'held'", g_dir); expect(run_cli(args, NULL) == 0, "-e -b 1 with three texts exits 0"); expect(count_lines(g_out) == 7, "-e prints matrix then raw embeddings"); @@ -167,16 +162,17 @@ int main(int argc, char **argv) snprintf(args, sizeof(args), "-d '%s' --stream", g_dir); expect(run_cli(args, "hello\nworld\n") == 0, "--stream exits 0"); expect(count_lines(g_out) == 2, "--stream writes one line per input"); - expect(file_contains(g_out, "{\"embedding\":["), - "--stream writes embedding JSON"); + expect(file_contains(g_out, "{\"embedding\":["), "--stream writes embedding JSON"); snprintf(args, sizeof(args), "-d '%s' --stream -b 2 -v", g_dir); - expect(run_cli(args, "hello\nworld\nheld\n") == 0, - "--stream -b 2 exits 0"); + expect(run_cli(args, "hello\nworld\nheld\n") == 0, "--stream -b 2 exits 0"); expect(count_lines(g_out) == 3, "--stream -b 2 covers the tail batch"); /* Flag validation: missing model dir must exit 1. */ - static const struct { const char *args; const char *what; } bad[] = { + static const struct { + const char *args; + const char *what; + } bad[] = { {"-d '' t", "empty model dir"}, }; for (size_t i = 0; i < sizeof(bad) / sizeof(bad[0]); i++) diff --git a/tests/test_kernels.c b/tests/test_kernels.c index 4a2a57d..fb3d2df 100644 --- a/tests/test_kernels.c +++ b/tests/test_kernels.c @@ -18,8 +18,7 @@ static void expect_close(const char *name, float got, float want, float tol) { float diff = fabsf(got - want); float scale = fmaxf(1.0f, fmaxf(fabsf(got), fabsf(want))); if (diff > tol * scale) { - fprintf(stderr, "%s: got %.9g want %.9g diff %.9g\n", - name, got, want, diff); + fprintf(stderr, "%s: got %.9g want %.9g diff %.9g\n", name, got, want, diff); failures++; } } @@ -39,10 +38,7 @@ static float bf16_to_f32(uint16_t x) { static void test_rms_norm(void) { const int seq = 2, hidden = 4; - const float x[8] = { - 1.0f, 2.0f, -3.0f, 4.0f, - 0.5f, -1.0f, 2.0f, -0.5f - }; + const float x[8] = {1.0f, 2.0f, -3.0f, 4.0f, 0.5f, -1.0f, 2.0f, -0.5f}; const float w[4] = {1.0f, 0.5f, -1.0f, 2.0f}; float out[8]; @@ -66,10 +62,7 @@ static void test_rope_neox(void) { const int seq = 2, heads = 1, head_dim = 4; const int positions[2] = {0, 3}; float cosv[8], sinv[8]; - float x[8] = { - 1.0f, 2.0f, 3.0f, 4.0f, - -1.0f, 0.5f, 2.0f, -0.25f - }; + float x[8] = {1.0f, 2.0f, 3.0f, 4.0f, -1.0f, 0.5f, 2.0f, -0.25f}; float orig[8]; memcpy(orig, x, sizeof(x)); @@ -92,10 +85,17 @@ static void test_rope_neox(void) { } } -static void reference_packed_gqa(float *out, const float *Q, const float *K, - const float *V, const int *offsets, int batch, - int n_heads, int n_kv_heads, int head_dim, - float scale, int causal) { +static void reference_packed_gqa(float *out, + const float *Q, + const float *K, + const float *V, + const int *offsets, + int batch, + int n_heads, + int n_kv_heads, + int head_dim, + float scale, + int causal) { int q_hidden = n_heads * head_dim; int kv_hidden = n_kv_heads * head_dim; int heads_per_kv = n_heads / n_kv_heads; @@ -112,22 +112,28 @@ static void reference_packed_gqa(float *out, const float *Q, const float *K, for (int j = start; j < key_end; j++) { const float *k = K + j * kv_hidden + kv_h * head_dim; float score = 0.0f; - for (int d = 0; d < head_dim; d++) score += q[d] * k[d]; + for (int d = 0; d < head_dim; d++) + score += q[d] * k[d]; score *= scale; - if (score > max_score) max_score = score; + if (score > max_score) + max_score = score; } float denom = 0.0f; - for (int d = 0; d < head_dim; d++) o[d] = 0.0f; + for (int d = 0; d < head_dim; d++) + o[d] = 0.0f; for (int j = start; j < key_end; j++) { const float *k = K + j * kv_hidden + kv_h * head_dim; const float *v = V + j * kv_hidden + kv_h * head_dim; float score = 0.0f; - for (int d = 0; d < head_dim; d++) score += q[d] * k[d]; + for (int d = 0; d < head_dim; d++) + score += q[d] * k[d]; float wt = expf(score * scale - max_score); denom += wt; - for (int d = 0; d < head_dim; d++) o[d] += wt * v[d]; + for (int d = 0; d < head_dim; d++) + o[d] += wt * v[d]; } - for (int d = 0; d < head_dim; d++) o[d] /= denom; + for (int d = 0; d < head_dim; d++) + o[d] /= denom; } } } @@ -151,11 +157,9 @@ static void test_packed_gqa_attention(void) { V[i] = sinf((float)i * 0.07f) - cosf((float)i * 0.13f); } - qwen_bidirectional_gqa_attention_packed(got, Q, K, V, offsets, 2, - n_heads, n_kv_heads, head_dim, + qwen_bidirectional_gqa_attention_packed(got, Q, K, V, offsets, 2, n_heads, n_kv_heads, head_dim, 1.0f / sqrtf((float)head_dim)); - reference_packed_gqa(want, Q, K, V, offsets, 2, - n_heads, n_kv_heads, head_dim, + reference_packed_gqa(want, Q, K, V, offsets, 2, n_heads, n_kv_heads, head_dim, 1.0f / sqrtf((float)head_dim), 0); for (int i = 0; i < total * q_hidden; i++) @@ -172,7 +176,8 @@ static void test_packed_gqa_attention_long(void) { float *V = malloc((size_t)total * kv_hidden * sizeof(*V)); float *got = malloc((size_t)total * q_hidden * sizeof(*got)); float *want = malloc((size_t)total * q_hidden * sizeof(*want)); - if (!Q || !K || !V || !got || !want) exit(2); + if (!Q || !K || !V || !got || !want) + exit(2); for (int i = 0; i < total * q_hidden; i++) Q[i] = sinf((float)i * 0.017f) - 0.25f; @@ -182,11 +187,9 @@ static void test_packed_gqa_attention_long(void) { } qwen_set_threads(4); - qwen_bidirectional_gqa_attention_packed(got, Q, K, V, offsets, 2, - n_heads, n_kv_heads, head_dim, + qwen_bidirectional_gqa_attention_packed(got, Q, K, V, offsets, 2, n_heads, n_kv_heads, head_dim, 1.0f / sqrtf((float)head_dim)); - reference_packed_gqa(want, Q, K, V, offsets, 2, - n_heads, n_kv_heads, head_dim, + reference_packed_gqa(want, Q, K, V, offsets, 2, n_heads, n_kv_heads, head_dim, 1.0f / sqrtf((float)head_dim), 0); qwen_set_threads(1); @@ -209,7 +212,8 @@ static void test_packed_causal_gqa_attention(void) { float *V = malloc((size_t)total * kv_hidden * sizeof(*V)); float *got = malloc((size_t)total * q_hidden * sizeof(*got)); float *want = malloc((size_t)total * q_hidden * sizeof(*want)); - if (!Q || !K || !V || !got || !want) exit(2); + if (!Q || !K || !V || !got || !want) + exit(2); for (int i = 0; i < total * q_hidden; i++) Q[i] = sinf((float)i * 0.019f) - 0.2f; @@ -219,11 +223,9 @@ static void test_packed_causal_gqa_attention(void) { } qwen_set_threads(4); - qwen_causal_gqa_attention_packed(got, Q, K, V, offsets, 2, - n_heads, n_kv_heads, head_dim, + qwen_causal_gqa_attention_packed(got, Q, K, V, offsets, 2, n_heads, n_kv_heads, head_dim, 1.0f / sqrtf((float)head_dim)); - reference_packed_gqa(want, Q, K, V, offsets, 2, - n_heads, n_kv_heads, head_dim, + reference_packed_gqa(want, Q, K, V, offsets, 2, n_heads, n_kv_heads, head_dim, 1.0f / sqrtf((float)head_dim), 1); qwen_set_threads(1); @@ -238,18 +240,13 @@ static void test_packed_causal_gqa_attention(void) { static void test_bf16_linear(void) { enum { seq = 2, in_dim = 3, out_dim = 2 }; - const float x[6] = { - 1.0f, -2.0f, 0.5f, - -1.0f, 0.25f, 2.0f - }; - const float wf32[6] = { - 0.5f, -1.0f, 2.0f, - -0.25f, 0.75f, 1.5f - }; + const float x[6] = {1.0f, -2.0f, 0.5f, -1.0f, 0.25f, 2.0f}; + const float wf32[6] = {0.5f, -1.0f, 2.0f, -0.25f, 0.75f, 1.5f}; uint16_t wbf16[6]; float got[seq * out_dim]; - for (int i = 0; i < 6; i++) wbf16[i] = f32_to_bf16(wf32[i]); + for (int i = 0; i < 6; i++) + wbf16[i] = f32_to_bf16(wf32[i]); qwen_linear_nobias_bf16(got, x, wbf16, seq, in_dim, out_dim); for (int s = 0; s < seq; s++) { @@ -287,8 +284,7 @@ static void check_bf16_qkv(int n_threads) { } qwen_set_threads(n_threads); - qwen_linear_nobias_bf16_qkv(got_q, got_k, got_v, x, wq, wk, wv, - seq, in_dim, q_dim, kv_dim); + qwen_linear_nobias_bf16_qkv(got_q, got_k, got_v, x, wq, wk, wv, seq, in_dim, q_dim, kv_dim); for (int s = 0; s < seq; s++) { for (int o = 0; o < q_dim; o++) { @@ -349,7 +345,7 @@ static void test_bf16_matvec(void) { * rms_norm_worker; row-wise math is identical, so results must match the * single-thread run exactly. */ static void test_rms_norm_threaded(void) { - enum { seq = 96, hidden = 3072 }; /* 294912 elements */ + enum { seq = 96, hidden = 3072 }; /* 294912 elements */ size_t elems = (size_t)seq * hidden; float *x = (float *)malloc(elems * sizeof(float)); float *w = (float *)malloc((size_t)hidden * sizeof(float)); @@ -358,7 +354,10 @@ static void test_rms_norm_threaded(void) { if (!x || !w || !one || !many) { fprintf(stderr, "rms_norm_threaded: allocation failure\n"); failures++; - free(many); free(one); free(w); free(x); + free(many); + free(one); + free(w); + free(x); return; } unsigned s = 7u; @@ -377,13 +376,15 @@ static void test_rms_norm_threaded(void) { for (size_t i = 0; i < elems; i++) { if (one[i] != many[i]) { - fprintf(stderr, "rms_norm_threaded: mismatch at %zu: %.9g %.9g\n", - i, one[i], many[i]); + fprintf(stderr, "rms_norm_threaded: mismatch at %zu: %.9g %.9g\n", i, one[i], many[i]); failures++; break; } } - free(many); free(one); free(w); free(x); + free(many); + free(one); + free(w); + free(x); } /* bf16->f32 widening is an exact bit shift; check the SIMD body and the @@ -407,8 +408,7 @@ static void test_bf16_widen_buf(void) { memcpy(&got_bits, &got[i], sizeof(got_bits)); memcpy(&want_bits, &want, sizeof(want_bits)); if (got_bits != want_bits) { - fprintf(stderr, "bf16_widen n=%d at %d: %.9g want %.9g\n", - n, i, got[i], want); + fprintf(stderr, "bf16_widen n=%d at %d: %.9g want %.9g\n", n, i, got[i], want); failures++; } } @@ -436,8 +436,7 @@ static void check_bf16_pair(int n_threads) { } qwen_set_threads(n_threads); - qwen_linear_nobias_bf16_pair(got_a, got_b, x, wa, wb, - seq, in_dim, a_dim, b_dim); + qwen_linear_nobias_bf16_pair(got_a, got_b, x, wa, wb, seq, in_dim, a_dim, b_dim); for (int s = 0; s < seq; s++) { for (int o = 0; o < a_dim; o++) { @@ -477,8 +476,10 @@ static void test_generic_vs_impl(void) { s = s * 1664525u + 1013904223u; b[i] = (float)(int32_t)s * 1e-9f; } - for (int i = 0; i < IN; i++) x[i] = a[i]; - for (int o = 0; o < OUT; o++) bias[o] = b[o]; + for (int i = 0; i < IN; i++) + x[i] = a[i]; + for (int o = 0; o < OUT; o++) + bias[o] = b[o]; for (size_t i = 0; i < sizeof(W) / sizeof(W[0]); i++) { s = s * 1664525u + 1013904223u; W[i] = f32_to_bf16((float)(int32_t)s * 1e-9f); @@ -497,8 +498,7 @@ static void test_generic_vs_impl(void) { const int sizes[] = {1, 7, 64, NMAX}; for (size_t t = 0; t < sizeof(sizes) / sizeof(sizes[0]); t++) { int n = sizes[t]; - expect_close("dot generic-vs-impl", - qwen_dot_f32_generic(a, b, n), + expect_close("dot generic-vs-impl", qwen_dot_f32_generic(a, b, n), qwen_dot_f32_impl(a, b, n), 1e-5f); memcpy(d0, a, (size_t)n * sizeof(float)); @@ -538,7 +538,8 @@ int main(void) { test_bf16_widen_buf(); test_bf16_pair(); test_generic_vs_impl(); - if (failures != 0) return 1; + if (failures != 0) + return 1; puts("ok: kernel golden tests passed"); return 0; } diff --git a/tests/test_late.c b/tests/test_late.c index 94bdfe2..e491596 100644 --- a/tests/test_late.c +++ b/tests/test_late.c @@ -15,50 +15,50 @@ enum { TOKEN_DIM = 6 }; -static float max_abs_diff(const float *a, const float *b, size_t n) -{ +static float max_abs_diff(const float *a, const float *b, size_t n) { float m = 0.0f; for (size_t i = 0; i < n; i++) { float d = fabsf(a[i] - b[i]); - if (d > m) m = d; + if (d > m) + m = d; } return m; } -static float row_norm(const float *v, int n) -{ +static float row_norm(const float *v, int n) { float s = 0.0f; - for (int i = 0; i < n; i++) s += v[i] * v[i]; + for (int i = 0; i < n; i++) + s += v[i] * v[i]; return sqrtf(s); } /* Parity between the grouped batch scorer and per-document MaxSim. */ -static int check_maxsim_batch_parity(const float *query, int query_tokens, - const float *docs, const int *offsets, - int n_docs, int dim, const char *what) -{ +static int check_maxsim_batch_parity(const float *query, + int query_tokens, + const float *docs, + const int *offsets, + int n_docs, + int dim, + const char *what) { float scores[8]; - if (n_docs > (int)(sizeof(scores) / sizeof(scores[0]))) return 1; - if (embed_late_maxsim_batch(query, query_tokens, docs, offsets, n_docs, - dim, scores) != 0) { + if (n_docs > (int)(sizeof(scores) / sizeof(scores[0]))) + return 1; + if (embed_late_maxsim_batch(query, query_tokens, docs, offsets, n_docs, dim, scores) != 0) { fprintf(stderr, "maxsim_batch (%s) failed\n", what); return 1; } for (int i = 0; i < n_docs; i++) { - float one = embed_late_maxsim(query, query_tokens, - docs + (size_t)offsets[i] * dim, - offsets[i + 1] - offsets[i], dim); + float one = embed_late_maxsim(query, query_tokens, docs + (size_t)offsets[i] * dim, + offsets[i + 1] - offsets[i], dim); if (fabsf(scores[i] - one) > 0.0001f) { - fprintf(stderr, "maxsim_batch (%s) doc %d: %g vs %g\n", - what, i, scores[i], one); + fprintf(stderr, "maxsim_batch (%s) doc %d: %g vs %g\n", what, i, scores[i], one); return 1; } } return 0; } -int main(void) -{ +int main(void) { int rc = 1; tm_dims_t dims = {4, 2, 1, 2, 8, TF_VOCAB_SIZE}; int hidden = dims.hidden; @@ -77,8 +77,7 @@ int main(void) char dir[1024]; snprintf(dir, sizeof(dir), "%s/embed-late-test-XXXXXX", getenv("TMPDIR") ? getenv("TMPDIR") : "/tmp"); - if (!mkdtemp(dir) || tf_write_vocab(dir) != 0 || - tm_write_model_dims(dir, "F32", &dims) != 0) { + if (!mkdtemp(dir) || tf_write_vocab(dir) != 0 || tm_write_model_dims(dir, "F32", &dims) != 0) { fprintf(stderr, "fixture creation failed\n"); return 2; } @@ -96,8 +95,7 @@ int main(void) fprintf(stderr, "late load succeeded without 1_Dense\n"); goto fail; } - if (embed_late_model_config(NULL) != NULL || - embed_late_model_token_dim(NULL) != 0 || + if (embed_late_model_config(NULL) != NULL || embed_late_model_token_dim(NULL) != 0 || embed_late_workspace_new(NULL) != NULL) { fprintf(stderr, "NULL-argument handling failed\n"); goto fail; @@ -106,8 +104,7 @@ int main(void) embed_late_model_free(NULL); /* A head whose tensor is not named linear.weight must be rejected. */ - if (tm_write_late_projection_named(dir, "F32", TOKEN_DIM, hidden, - "dense.weight") != 0) { + if (tm_write_late_projection_named(dir, "F32", TOKEN_DIM, hidden, "dense.weight") != 0) { fprintf(stderr, "failed to write malformed projection\n"); goto fail; } @@ -133,8 +130,7 @@ int main(void) } const embed_config_t *cfg = embed_late_model_config(late); - if (!cfg || cfg->hidden_size != hidden || - embed_late_model_token_dim(late) != TOKEN_DIM) { + if (!cfg || cfg->hidden_size != hidden || embed_late_model_token_dim(late) != TOKEN_DIM) { fprintf(stderr, "late config/token_dim mismatch\n"); goto fail; } @@ -253,33 +249,30 @@ int main(void) } offsets[i + 1] = offsets[i] + lens[i]; } - doc_vecs = (float *)calloc((size_t)offsets[3] * TOKEN_DIM, - sizeof(float)); + doc_vecs = (float *)calloc((size_t)offsets[3] * TOKEN_DIM, sizeof(float)); if (!doc_vecs) { fprintf(stderr, "allocation failure\n"); goto fail; } for (int i = 0; i < 3; i++) { - if (embed_late_model_encode_tokens( - late, ws, doc_ids[i], lens[i], 1, - doc_vecs + (size_t)offsets[i] * TOKEN_DIM) != 0) { + if (embed_late_model_encode_tokens(late, ws, doc_ids[i], lens[i], 1, + doc_vecs + (size_t)offsets[i] * TOKEN_DIM) != 0) { fprintf(stderr, "doc encode failed\n"); goto fail; } } - if (check_maxsim_batch_parity(vecs, n_tokens, doc_vecs, offsets, 3, - TOKEN_DIM, "encoded docs") != 0) + if (check_maxsim_batch_parity(vecs, n_tokens, doc_vecs, offsets, 3, TOKEN_DIM, + "encoded docs") != 0) goto fail; float scores[3]; int bad_first[4] = {1, 2, 3, 4}; int empty_doc[3] = {0, 2, 2}; - if (embed_late_maxsim_batch(vecs, n_tokens, doc_vecs, bad_first, 3, - TOKEN_DIM, scores) == 0 || - embed_late_maxsim_batch(vecs, n_tokens, doc_vecs, empty_doc, 2, - TOKEN_DIM, scores) == 0 || - embed_late_maxsim_batch(vecs, n_tokens, doc_vecs, offsets, 3, - TOKEN_DIM, NULL) == 0) { + if (embed_late_maxsim_batch(vecs, n_tokens, doc_vecs, bad_first, 3, TOKEN_DIM, scores) == + 0 || + embed_late_maxsim_batch(vecs, n_tokens, doc_vecs, empty_doc, 2, TOKEN_DIM, scores) == + 0 || + embed_late_maxsim_batch(vecs, n_tokens, doc_vecs, offsets, 3, TOKEN_DIM, NULL) == 0) { fprintf(stderr, "maxsim_batch accepted invalid arguments\n"); goto fail; } @@ -305,20 +298,20 @@ int main(void) s = s * 1664525u + 1013904223u; big_d[i] = (float)((s >> 8) & 0xFFFF) / 32768.0f - 1.0f; } - if (check_maxsim_batch_parity(big_q, BIG_Q, big_d, offsets, 2, - BIG_DIM, "oversized doc") != 0) + if (check_maxsim_batch_parity(big_q, BIG_Q, big_d, offsets, 2, BIG_DIM, "oversized doc") != + 0) goto fail; } - printf("ok: late API token_dim=%d tokens=%d hidden=%d\n", - TOKEN_DIM, n_tokens, hidden); + printf("ok: late API token_dim=%d tokens=%d hidden=%d\n", TOKEN_DIM, n_tokens, hidden); rc = 0; fail: free(big_d); free(big_q); free(doc_vecs); - for (int i = 0; i < 3; i++) free(doc_ids[i]); + for (int i = 0; i < 3; i++) + free(doc_ids[i]); free(expected); free(vecs); free(raw); diff --git a/tests/test_qwen3.c b/tests/test_qwen3.c index 4d63a35..d66e0f7 100644 --- a/tests/test_qwen3.c +++ b/tests/test_qwen3.c @@ -9,24 +9,22 @@ #include <string.h> #include <unistd.h> -static float max_abs_diff(const float *a, const float *b, int n) -{ +static float max_abs_diff(const float *a, const float *b, int n) { float out = 0.0f; for (int i = 0; i < n; i++) { float d = fabsf(a[i] - b[i]); - if (d > out) out = d; + if (d > out) + out = d; } return out; } -int main(void) -{ +int main(void) { char dir[1024]; snprintf(dir, sizeof(dir), "%s/embed-qwen3-test-XXXXXX", getenv("TMPDIR") ? getenv("TMPDIR") : "/tmp"); tm_dims_t dims = {4, 2, 1, 2, 8, 16}; - if (!mkdtemp(dir) || - tm_write_qwen3_model_dims(dir, "F32", &dims, 15) != 0) { + if (!mkdtemp(dir) || tm_write_qwen3_model_dims(dir, "F32", &dims, 15) != 0) { fprintf(stderr, "fixture creation failed\n"); return 2; } @@ -39,8 +37,8 @@ int main(void) return 1; } if (cfg->attention_mode != EMBED_ATTENTION_CAUSAL || - cfg->pooling_mode != EMBED_POOL_LAST_TOKEN || - !cfg->normalize_embeddings || !cfg->append_terminal_token) { + cfg->pooling_mode != EMBED_POOL_LAST_TOKEN || !cfg->normalize_embeddings || + !cfg->append_terminal_token) { fprintf(stderr, "Qwen3 config semantics were not detected\n"); return 1; } @@ -58,15 +56,15 @@ int main(void) float expected[4]; memcpy(expected, states + 3 * 4, sizeof(expected)); - if (embed_l2_normalize(expected, 4) != 0 || - max_abs_diff(embedding, expected, 4) > 1e-5f || + if (embed_l2_normalize(expected, 4) != 0 || max_abs_diff(embedding, expected, 4) > 1e-5f || max_abs_diff(pooled, expected, 4) > 1e-5f) { fprintf(stderr, "last-token pooling mismatch\n"); return 1; } float norm_sq = 0.0f; - for (int d = 0; d < 4; d++) norm_sq += embedding[d] * embedding[d]; + for (int d = 0; d < 4; d++) + norm_sq += embedding[d] * embedding[d]; if (fabsf(norm_sq - 1.0f) > 1e-5f) { fprintf(stderr, "Qwen3 embedding is not normalized: %.9g\n", norm_sq); return 1; @@ -78,8 +76,7 @@ int main(void) float single[4]; if (embed_model_encode_batch(model, ws, inputs, 2, batch) != 0 || embed_model_encode_into(model, ws, short_ids, 2, single) != 0 || - max_abs_diff(batch, embedding, 4) > 1e-5f || - max_abs_diff(batch + 4, single, 4) > 1e-5f) { + max_abs_diff(batch, embedding, 4) > 1e-5f || max_abs_diff(batch + 4, single, 4) > 1e-5f) { fprintf(stderr, "Qwen3 packed batch parity failed\n"); return 1; } diff --git a/tests/test_safetensors.c b/tests/test_safetensors.c index d527951..36ec8c5 100644 --- a/tests/test_safetensors.c +++ b/tests/test_safetensors.c @@ -15,29 +15,31 @@ static int failures = 0; static char tmproot[1024]; /* Write u64-LE header length + JSON header + payload (safetensors layout). */ -static void write_st(const char *path, const char *header, - const void *payload, size_t payload_len) -{ +static void +write_st(const char *path, const char *header, const void *payload, size_t payload_len) { FILE *f = fopen(path, "wb"); - if (!f) { perror(path); exit(2); } + if (!f) { + perror(path); + exit(2); + } uint64_t n = strlen(header); unsigned char len8[8]; - for (int i = 0; i < 8; i++) len8[i] = (unsigned char)(n >> (8 * i)); + for (int i = 0; i < 8; i++) + len8[i] = (unsigned char)(n >> (8 * i)); fwrite(len8, 1, 8, f); fwrite(header, 1, n, f); - if (payload_len) fwrite(payload, 1, payload_len, f); + if (payload_len) + fwrite(payload, 1, payload_len, f); fclose(f); } -static const char *path_in(const char *rel) -{ +static const char *path_in(const char *rel) { static char buf[2048]; snprintf(buf, sizeof(buf), "%s/%s", tmproot, rel); return buf; } -static void check(const char *mode, const char *path, int want_ok) -{ +static void check(const char *mode, const char *path, int want_ok) { int got_ok; if (strcmp(mode, "file") == 0) { safetensors_file_t *sf = safetensors_open(path); @@ -49,8 +51,8 @@ static void check(const char *mode, const char *path, int want_ok) multi_safetensors_close(ms); } if (got_ok != want_ok) { - fprintf(stderr, "%s %s: got %s want %s\n", mode, path, - got_ok ? "ok" : "fail", want_ok ? "ok" : "fail"); + fprintf(stderr, "%s %s: got %s want %s\n", mode, path, got_ok ? "ok" : "fail", + want_ok ? "ok" : "fail"); failures++; } } @@ -62,11 +64,13 @@ static const char ONE_TENSOR_B[] = static const char ONE_TENSOR_X[] = "{\"x\":{\"dtype\":\"F32\",\"shape\":[1],\"data_offsets\":[0,4]}}"; -int main(void) -{ +int main(void) { snprintf(tmproot, sizeof(tmproot), "%s/embed-st-test-XXXXXX", getenv("TMPDIR") ? getenv("TMPDIR") : "/tmp"); - if (!mkdtemp(tmproot)) { perror("mkdtemp"); return 2; } + if (!mkdtemp(tmproot)) { + perror("mkdtemp"); + return 2; + } static const char zeros[4 * 1025] = {0}; @@ -77,14 +81,15 @@ int main(void) { size_t cap = 1025 * 64 + 16; char *hdr = malloc(cap); - if (!hdr) return 2; + if (!hdr) + return 2; size_t off = 0; off += (size_t)snprintf(hdr + off, cap - off, "{"); for (int i = 0; i < 1025; i++) { off += (size_t)snprintf(hdr + off, cap - off, - "%s\"t%d\":{\"dtype\":\"F32\",\"shape\":[1]," - "\"data_offsets\":[%d,%d]}", - i ? "," : "", i, i * 4, i * 4 + 4); + "%s\"t%d\":{\"dtype\":\"F32\",\"shape\":[1]," + "\"data_offsets\":[%d,%d]}", + i ? "," : "", i, i * 4, i * 4 + 4); } snprintf(hdr + off, cap - off, "}"); write_st(path_in("too_many.safetensors"), hdr, zeros, sizeof(zeros)); @@ -94,27 +99,26 @@ int main(void) write_st(path_in("too_many_dims.safetensors"), "{\"x\":{\"dtype\":\"F32\",\"shape\":[1,1,1,1,1,1,1,1,1]," - "\"data_offsets\":[0,4]}}", zeros, 4); + "\"data_offsets\":[0,4]}}", + zeros, 4); check("file", path_in("too_many_dims.safetensors"), 0); write_st(path_in("bad_offsets.safetensors"), "{\"x\":{\"dtype\":\"F32\",\"shape\":[1]," - "\"data_offsets\":[4,0]}}", zeros, 4); + "\"data_offsets\":[4,0]}}", + zeros, 4); check("file", path_in("bad_offsets.safetensors"), 0); mkdir(path_in("valid_multi"), 0755); - write_st(path_in("valid_multi/model-00001-of-00002.safetensors"), - ONE_TENSOR_A, zeros, 4); - write_st(path_in("valid_multi/model-00002-of-00002.safetensors"), - ONE_TENSOR_B, zeros, 4); + write_st(path_in("valid_multi/model-00001-of-00002.safetensors"), ONE_TENSOR_A, zeros, 4); + write_st(path_in("valid_multi/model-00002-of-00002.safetensors"), ONE_TENSOR_B, zeros, 4); check("multi", path_in("valid_multi"), 1); /* Payload accounting across shards: two 4-byte F32 tensors. */ { multi_safetensors_t *ms = multi_safetensors_open(path_in("valid_multi")); size_t nbytes = 0; - if (!ms || multi_safetensors_data_nbytes(ms, &nbytes) != 0 || - nbytes != 8) { + if (!ms || multi_safetensors_data_nbytes(ms, &nbytes) != 0 || nbytes != 8) { fprintf(stderr, "multi data_nbytes: got %zu want 8\n", nbytes); failures++; } @@ -122,18 +126,15 @@ int main(void) } mkdir(path_in("missing"), 0755); - write_st(path_in("missing/model-00001-of-00002.safetensors"), - ONE_TENSOR_X, zeros, 4); + write_st(path_in("missing/model-00001-of-00002.safetensors"), ONE_TENSOR_X, zeros, 4); check("multi", path_in("missing"), 0); mkdir(path_in("too_many_shards"), 0755); - write_st(path_in("too_many_shards/model-00001-of-00009.safetensors"), - ONE_TENSOR_X, zeros, 4); + write_st(path_in("too_many_shards/model-00001-of-00009.safetensors"), ONE_TENSOR_X, zeros, 4); check("multi", path_in("too_many_shards"), 0); mkdir(path_in("bad_name"), 0755); - write_st(path_in("bad_name/model-foo.safetensors"), - ONE_TENSOR_X, zeros, 4); + write_st(path_in("bad_name/model-foo.safetensors"), ONE_TENSOR_X, zeros, 4); check("multi", path_in("bad_name"), 0); if (failures) { diff --git a/tests/test_server.c b/tests/test_server.c index a55462c..d799107 100644 --- a/tests/test_server.c +++ b/tests/test_server.c @@ -16,30 +16,35 @@ static int test_failures = 0; -#define TEST_ASSERT(cond) do { \ - if (!(cond)) { \ - fprintf(stderr, "FAIL %s:%d: %s\n", __FILE__, __LINE__, #cond); \ - test_failures++; \ - } \ -} while (0) +#define TEST_ASSERT(cond) \ + do { \ + if (!(cond)) { \ + fprintf(stderr, "FAIL %s:%d: %s\n", __FILE__, __LINE__, #cond); \ + test_failures++; \ + } \ + } while (0) -static void test_base64_rfc4648(void) -{ +static void test_base64_rfc4648(void) { /* RFC 4648 test vectors. */ - static const struct { const char *in, *out; } cases[] = { - {"", ""}, {"f", "Zg=="}, {"fo", "Zm8="}, {"foo", "Zm9v"}, - {"foob", "Zm9vYg=="}, {"fooba", "Zm9vYmE="}, {"foobar", "Zm9vYmFy"}, + static const struct { + const char *in, *out; + } cases[] = { + {"", ""}, + {"f", "Zg=="}, + {"fo", "Zm8="}, + {"foo", "Zm9v"}, + {"foob", "Zm9vYg=="}, + {"fooba", "Zm9vYmE="}, + {"foobar", "Zm9vYmFy"}, }; for (size_t i = 0; i < sizeof(cases) / sizeof(cases[0]); i++) { - char *got = base64_encode((const unsigned char *)cases[i].in, - strlen(cases[i].in)); + char *got = base64_encode((const unsigned char *)cases[i].in, strlen(cases[i].in)); TEST_ASSERT(strcmp(got, cases[i].out) == 0); free(got); } } -static void test_quantize_int8_tanh(void) -{ +static void test_quantize_int8_tanh(void) { TEST_ASSERT(quantize_int8_tanh(0.0f) == 0); /* tanh saturates to +-1, scaled by 127. */ TEST_ASSERT(quantize_int8_tanh(100.0f) == 127); @@ -50,8 +55,7 @@ static void test_quantize_int8_tanh(void) TEST_ASSERT(quantize_int8_tanh(0.5f) == 59); } -static void test_encode_embedding_int8(void) -{ +static void test_encode_embedding_int8(void) { float emb[4] = {0.0f, 0.5f, -0.5f, 100.0f}; signed char expect[4] = {0, 59, -59, 127}; char *got = encode_embedding(emb, 4, "base64_int8"); @@ -61,8 +65,7 @@ static void test_encode_embedding_int8(void) free(want); } -static void test_encode_embedding_binary(void) -{ +static void test_encode_embedding_binary(void) { /* Sign bits pack LSB-first: bit i of byte i/8 is 1 when emb[i] >= 0. */ float emb[8] = {1.0f, -1.0f, 1.0f, -1.0f, -1.0f, -1.0f, -1.0f, 1.0f}; unsigned char expect = 0x85; /* bits 0, 2, 7 set */ @@ -82,8 +85,7 @@ static void test_encode_embedding_binary(void) free(want); } -static void test_encode_embedding_base64_float32(void) -{ +static void test_encode_embedding_base64_float32(void) { /* OpenAI/DashScope "base64" is base64 of the raw little-endian float32 * vector - lossless, unlike the int8 formats. */ float emb[4] = {0.0f, 0.5f, -0.5f, 100.0f}; @@ -98,8 +100,7 @@ static void test_encode_embedding_base64_float32(void) free(want); } -static void test_append_embedding_value_float(void) -{ +static void test_append_embedding_value_float(void) { float emb[1] = {0.5f}; /* OpenAI/DashScope (Qwen3): the true float32 value. */ @@ -115,16 +116,14 @@ static void test_append_embedding_value_float(void) sbuf_free(&b2); } -static void test_encoding_from_root_family(void) -{ +static void test_encoding_from_root_family(void) { cJSON *detail = cJSON_CreateArray(); /* Default encoding differs by family when the field is absent. */ cJSON *empty = cJSON_CreateObject(); - TEST_ASSERT(strcmp(encoding_from_root(empty, detail, EMBED_API_OPENAI), - "float") == 0); - TEST_ASSERT(strcmp(encoding_from_root(empty, detail, EMBED_API_PERPLEXITY), - "base64_int8") == 0); + TEST_ASSERT(strcmp(encoding_from_root(empty, detail, EMBED_API_OPENAI), "float") == 0); + TEST_ASSERT(strcmp(encoding_from_root(empty, detail, EMBED_API_PERPLEXITY), "base64_int8") == + 0); cJSON_Delete(empty); TEST_ASSERT(cJSON_GetArraySize(detail) == 0); @@ -144,8 +143,7 @@ static void test_encoding_from_root_family(void) cJSON_Delete(detail); } -static void test_collect_job_batch_deadline(void) -{ +static void test_collect_job_batch_deadline(void) { http_server s; memset(&s, 0, sizeof(s)); s.batch_size = 32; @@ -158,10 +156,8 @@ static void test_collect_job_batch_deadline(void) job before_deadline = {.srv = &s, .created_ns = first_ns + 500000u}; job after_deadline = {.srv = &s, .created_ns = first_ns + 2000000u}; snprintf(first.path, sizeof(first.path), "%s", "/v1/embeddings"); - snprintf(before_deadline.path, sizeof(before_deadline.path), "%s", - "/v1/embeddings"); - snprintf(after_deadline.path, sizeof(after_deadline.path), "%s", - "/v1/embeddings"); + snprintf(before_deadline.path, sizeof(before_deadline.path), "%s", "/v1/embeddings"); + snprintf(after_deadline.path, sizeof(after_deadline.path), "%s", "/v1/embeddings"); enqueue_job(&first); enqueue_job(&before_deadline); enqueue_job(&after_deadline); @@ -179,8 +175,7 @@ static void test_collect_job_batch_deadline(void) pthread_mutex_destroy(&s.mu); } -static void test_collect_job_batch_zero_wait(void) -{ +static void test_collect_job_batch_zero_wait(void) { http_server s; memset(&s, 0, sizeof(s)); s.batch_size = 32; @@ -211,10 +206,10 @@ static void test_collect_job_batch_zero_wait(void) #define TEST_API_KEY "tt-test-key" -static int find_free_port(void) -{ +static int find_free_port(void) { int fd = socket(AF_INET, SOCK_STREAM, 0); - if (fd < 0) return -1; + if (fd < 0) + return -1; struct sockaddr_in a; memset(&a, 0, sizeof(a)); a.sin_family = AF_INET; @@ -230,10 +225,10 @@ static int find_free_port(void) return port; } -static int tcp_connect(int port) -{ +static int tcp_connect(int port) { int fd = socket(AF_INET, SOCK_STREAM, 0); - if (fd < 0) return -1; + if (fd < 0) + return -1; struct sockaddr_in a; memset(&a, 0, sizeof(a)); a.sin_family = AF_INET; @@ -246,13 +241,13 @@ static int tcp_connect(int port) return fd; } -static int send_all(int fd, const char *buf, size_t n) -{ +static int send_all(int fd, const char *buf, size_t n) { size_t off = 0; while (off < n) { ssize_t w = write(fd, buf + off, n - off); if (w < 0) { - if (errno == EINTR) continue; + if (errno == EINTR) + continue; return -1; } off += (size_t)w; @@ -262,10 +257,14 @@ static int send_all(int fd, const char *buf, size_t n) /* Send one request; raw == NULL builds a normal request, otherwise `raw` * bytes go on the wire verbatim (for malformed-HTTP checks). */ -static int http_send(int fd, const char *method, const char *path, - const char *auth, const char *body, const char *raw) -{ - if (raw) return send_all(fd, raw, strlen(raw)); +static int http_send(int fd, + const char *method, + const char *path, + const char *auth, + const char *body, + const char *raw) { + if (raw) + return send_all(fd, raw, strlen(raw)); char head[512]; size_t blen = body ? strlen(body) : 0; int n = snprintf(head, sizeof(head), @@ -275,55 +274,71 @@ static int http_send(int fd, const char *method, const char *path, "Content-Type: application/json\r\n" "Content-Length: %zu\r\n" "\r\n", - method, path, - auth ? "Authorization: Bearer " : "", - auth ? auth : "", auth ? "\r\n" : "", - blen); - if (n < 0 || (size_t)n >= sizeof(head)) return -1; - if (send_all(fd, head, (size_t)n) != 0) return -1; + method, path, auth ? "Authorization: Bearer " : "", auth ? auth : "", + auth ? "\r\n" : "", blen); + if (n < 0 || (size_t)n >= sizeof(head)) + return -1; + if (send_all(fd, head, (size_t)n) != 0) + return -1; return blen ? send_all(fd, body, blen) : 0; } /* Read one response: returns the status code (or -1) and the body, which * the caller frees. Reads exactly Content-Length body bytes so it works * with keep-alive connections. */ -static int http_recv(int fd, char **out_body) -{ +static int http_recv(int fd, char **out_body) { *out_body = NULL; size_t cap = 8192, len = 0; char *buf = malloc(cap); - if (!buf) return -1; + if (!buf) + return -1; char *hdr_end = NULL; while (!hdr_end) { if (len + 1 >= cap) { cap *= 2; char *nb = realloc(buf, cap); - if (!nb) { free(buf); return -1; } + if (!nb) { + free(buf); + return -1; + } buf = nb; } ssize_t r = read(fd, buf + len, cap - len - 1); - if (r <= 0) { free(buf); return -1; } + if (r <= 0) { + free(buf); + return -1; + } len += (size_t)r; buf[len] = '\0'; hdr_end = strstr(buf, "\r\n\r\n"); } int status = -1; - if (sscanf(buf, "HTTP/1.1 %d", &status) != 1) { free(buf); return -1; } + if (sscanf(buf, "HTTP/1.1 %d", &status) != 1) { + free(buf); + return -1; + } size_t content_length = 0; const char *cl = strstr(buf, "Content-Length:"); - if (cl && cl < hdr_end) content_length = strtoul(cl + 15, NULL, 10); + if (cl && cl < hdr_end) + content_length = strtoul(cl + 15, NULL, 10); size_t body_off = (size_t)(hdr_end - buf) + 4; size_t need = body_off + content_length; if (need + 1 > cap) { char *nb = realloc(buf, need + 1); - if (!nb) { free(buf); return -1; } + if (!nb) { + free(buf); + return -1; + } buf = nb; } while (len < need) { ssize_t r = read(fd, buf + len, need - len); - if (r <= 0) { free(buf); return -1; } + if (r <= 0) { + free(buf); + return -1; + } len += (size_t)r; } buf[need] = '\0'; @@ -334,19 +349,25 @@ static int http_recv(int fd, char **out_body) /* One-shot request helper; returns status, body via out_body (caller frees, * may be NULL when only the status matters). */ -static int http_req(int port, const char *method, const char *path, - const char *auth, const char *body, const char *raw, - char **out_body) -{ +static int http_req(int port, + const char *method, + const char *path, + const char *auth, + const char *body, + const char *raw, + char **out_body) { int fd = tcp_connect(port); - if (fd < 0) return -1; + if (fd < 0) + return -1; char *resp = NULL; int status = -1; if (http_send(fd, method, path, auth, body, raw) == 0) status = http_recv(fd, &resp); close(fd); - if (out_body) *out_body = resp; - else free(resp); + if (out_body) + *out_body = resp; + else + free(resp); return status; } @@ -356,18 +377,17 @@ typedef struct { int rc; } srv_ctx; -static void *srv_thread_main(void *arg) -{ +static void *srv_thread_main(void *arg) { srv_ctx *s = (srv_ctx *)arg; s->rc = embed_run_server(&s->cfg); return NULL; } -static cJSON *parse_embeddings_body(const char *body, int want_items) -{ +static cJSON *parse_embeddings_body(const char *body, int want_items) { cJSON *root = cJSON_Parse(body); TEST_ASSERT(root != NULL); - if (!root) return NULL; + if (!root) + return NULL; cJSON *data = cJSON_GetObjectItemCaseSensitive(root, "data"); TEST_ASSERT(cJSON_IsArray(data)); TEST_ASSERT(cJSON_GetArraySize(data) == want_items); @@ -378,71 +398,63 @@ static cJSON *parse_embeddings_body(const char *body, int want_items) return root; } -static void test_http_embeddings(int port) -{ +static void test_http_embeddings(int port) { char *body = NULL; /* Routing. */ - TEST_ASSERT(http_req(port, "OPTIONS", "/v1/embeddings", NULL, NULL, NULL, - NULL) == 204); - TEST_ASSERT(http_req(port, "GET", "/v1/embeddings", TEST_API_KEY, NULL, - NULL, NULL) == 404); - TEST_ASSERT(http_req(port, "POST", "/nope", TEST_API_KEY, "{}", NULL, - NULL) == 404); + TEST_ASSERT(http_req(port, "OPTIONS", "/v1/embeddings", NULL, NULL, NULL, NULL) == 204); + TEST_ASSERT(http_req(port, "GET", "/v1/embeddings", TEST_API_KEY, NULL, NULL, NULL) == 404); + TEST_ASSERT(http_req(port, "POST", "/nope", TEST_API_KEY, "{}", NULL, NULL) == 404); /* Auth: required, and the exact key. */ - const char *ok_req = - "{\"model\":\"pplx-embed-v1-0.6b\",\"input\":\"hello world\"," - "\"encoding_format\":\"float\"}"; - TEST_ASSERT(http_req(port, "POST", "/v1/embeddings", NULL, ok_req, NULL, - NULL) == 401); - TEST_ASSERT(http_req(port, "POST", "/v1/embeddings", "wrong-key", ok_req, - NULL, NULL) == 401); + const char *ok_req = "{\"model\":\"pplx-embed-v1-0.6b\",\"input\":\"hello world\"," + "\"encoding_format\":\"float\"}"; + TEST_ASSERT(http_req(port, "POST", "/v1/embeddings", NULL, ok_req, NULL, NULL) == 401); + TEST_ASSERT(http_req(port, "POST", "/v1/embeddings", "wrong-key", ok_req, NULL, NULL) == 401); /* Malformed HTTP (400) and malformed JSON (422 detail). */ - TEST_ASSERT(http_req(port, NULL, NULL, NULL, NULL, - "GARBAGE\r\n\r\n", NULL) == 400); - TEST_ASSERT(http_req(port, "POST", "/v1/embeddings", TEST_API_KEY, - "{nope", NULL, NULL) == 422); + TEST_ASSERT(http_req(port, NULL, NULL, NULL, NULL, "GARBAGE\r\n\r\n", NULL) == 400); + TEST_ASSERT(http_req(port, "POST", "/v1/embeddings", TEST_API_KEY, "{nope", NULL, NULL) == 422); /* Validation: missing/unknown model, wrong endpoint family, valid but * unloaded id, bad encoding enum, dimensions out of range; the * API answers validation errors with 422 detail arrays. */ - TEST_ASSERT(http_req(port, "POST", "/v1/embeddings", TEST_API_KEY, - "{\"input\":\"hi\"}", NULL, NULL) == 422); - TEST_ASSERT(http_req(port, "POST", "/v1/embeddings", TEST_API_KEY, - "{\"model\":\"foo\",\"input\":\"hi\"}", NULL, + TEST_ASSERT(http_req(port, "POST", "/v1/embeddings", TEST_API_KEY, "{\"input\":\"hi\"}", NULL, NULL) == 422); TEST_ASSERT(http_req(port, "POST", "/v1/embeddings", TEST_API_KEY, + "{\"model\":\"foo\",\"input\":\"hi\"}", NULL, NULL) == 422); + TEST_ASSERT(http_req(port, "POST", "/v1/embeddings", TEST_API_KEY, "{\"model\":\"pplx-embed-context-v1-0.6b\"," - "\"input\":\"hi\"}", NULL, NULL) == 422); + "\"input\":\"hi\"}", + NULL, NULL) == 422); TEST_ASSERT(http_req(port, "POST", "/v1/embeddings", TEST_API_KEY, - "{\"model\":\"pplx-embed-v1-4b\",\"input\":\"hi\"}", - NULL, NULL) == 503); + "{\"model\":\"pplx-embed-v1-4b\",\"input\":\"hi\"}", NULL, NULL) == 503); TEST_ASSERT(http_req(port, "POST", "/v1/embeddings", TEST_API_KEY, "{\"model\":\"pplx-embed-v1-0.6b\",\"input\":\"hi\"," - "\"encoding_format\":\"yaml\"}", NULL, NULL) == 422); + "\"encoding_format\":\"yaml\"}", + NULL, NULL) == 422); TEST_ASSERT(http_req(port, "POST", "/v1/embeddings", TEST_API_KEY, "{\"model\":\"pplx-embed-v1-0.6b\",\"input\":\"hi\"," - "\"dimensions\":64}", NULL, NULL) == 422); + "\"dimensions\":64}", + NULL, NULL) == 422); TEST_ASSERT(http_req(port, "POST", "/v1/embeddings", TEST_API_KEY, "{\"model\":\"pplx-embed-v1-0.6b\"," - "\"input\":[]}", NULL, NULL) == 422); + "\"input\":[]}", + NULL, NULL) == 422); /* Contextual endpoint: standard id is the wrong enum, the 4B contextual * id is valid but not loaded. */ - TEST_ASSERT(http_req(port, "POST", "/v1/contextualizedembeddings", - TEST_API_KEY, + TEST_ASSERT(http_req(port, "POST", "/v1/contextualizedembeddings", TEST_API_KEY, "{\"model\":\"pplx-embed-v1-0.6b\"," - "\"input\":[[\"a\"]]}", NULL, NULL) == 422); - TEST_ASSERT(http_req(port, "POST", "/v1/contextualizedembeddings", - TEST_API_KEY, + "\"input\":[[\"a\"]]}", + NULL, NULL) == 422); + TEST_ASSERT(http_req(port, "POST", "/v1/contextualizedembeddings", TEST_API_KEY, "{\"model\":\"pplx-embed-context-v1-4b\"," - "\"input\":[[\"a\"]]}", NULL, NULL) == 503); + "\"input\":[[\"a\"]]}", + NULL, NULL) == 503); /* Happy path, float encoding, single string input. */ - TEST_ASSERT(http_req(port, "POST", "/v1/embeddings", TEST_API_KEY, ok_req, - NULL, &body) == 200); + TEST_ASSERT(http_req(port, "POST", "/v1/embeddings", TEST_API_KEY, ok_req, NULL, &body) == 200); cJSON *root = body ? parse_embeddings_body(body, 1) : NULL; if (root) { cJSON *data = cJSON_GetObjectItemCaseSensitive(root, "data"); @@ -456,12 +468,12 @@ static void test_http_embeddings(int port) cJSON *v; cJSON_ArrayForEach(v, emb) { TEST_ASSERT(cJSON_IsNumber(v)); - if (v->valuedouble != 0.0) nonzero = 1; + if (v->valuedouble != 0.0) + nonzero = 1; } TEST_ASSERT(nonzero); cJSON *usage = cJSON_GetObjectItemCaseSensitive(root, "usage"); - cJSON *total = usage - ? cJSON_GetObjectItemCaseSensitive(usage, "total_tokens") : NULL; + cJSON *total = usage ? cJSON_GetObjectItemCaseSensitive(usage, "total_tokens") : NULL; TEST_ASSERT(total && cJSON_IsNumber(total) && total->valueint >= 1); cJSON_Delete(root); } @@ -478,19 +490,15 @@ static void test_http_embeddings(int port) root = body ? parse_embeddings_body(body, 3) : NULL; if (root) { cJSON *data = cJSON_GetObjectItemCaseSensitive(root, "data"); - cJSON *e0 = cJSON_GetObjectItemCaseSensitive( - cJSON_GetArrayItem(data, 0), "embedding"); - cJSON *e1 = cJSON_GetObjectItemCaseSensitive( - cJSON_GetArrayItem(data, 1), "embedding"); + cJSON *e0 = cJSON_GetObjectItemCaseSensitive(cJSON_GetArrayItem(data, 0), "embedding"); + cJSON *e1 = cJSON_GetObjectItemCaseSensitive(cJSON_GetArrayItem(data, 1), "embedding"); for (int i = 0; i < 3; i++) { - cJSON *idx = cJSON_GetObjectItemCaseSensitive( - cJSON_GetArrayItem(data, i), "index"); + cJSON *idx = cJSON_GetObjectItemCaseSensitive(cJSON_GetArrayItem(data, i), "index"); TEST_ASSERT(cJSON_IsNumber(idx) && idx->valueint == i); } int differ = 0; for (int i = 0; i < 1024 && !differ; i++) { - if (cJSON_GetArrayItem(e0, i)->valuedouble != - cJSON_GetArrayItem(e1, i)->valuedouble) + if (cJSON_GetArrayItem(e0, i)->valuedouble != cJSON_GetArrayItem(e1, i)->valuedouble) differ = 1; } TEST_ASSERT(differ); @@ -502,12 +510,12 @@ static void test_http_embeddings(int port) /* Default encoding is base64_int8: a string of 1024 quantized bytes. */ TEST_ASSERT(http_req(port, "POST", "/v1/embeddings", TEST_API_KEY, "{\"model\":\"pplx-embed-v1-0.6b\"," - "\"input\":\"hello\"}", NULL, &body) == 200); + "\"input\":\"hello\"}", + NULL, &body) == 200); root = body ? parse_embeddings_body(body, 1) : NULL; if (root) { cJSON *emb = cJSON_GetObjectItemCaseSensitive( - cJSON_GetArrayItem(cJSON_GetObjectItemCaseSensitive(root, "data"), - 0), "embedding"); + cJSON_GetArrayItem(cJSON_GetObjectItemCaseSensitive(root, "data"), 0), "embedding"); TEST_ASSERT(cJSON_IsString(emb)); TEST_ASSERT(strlen(emb->valuestring) == ((1024 + 2) / 3) * 4); cJSON_Delete(root); @@ -524,8 +532,7 @@ static void test_http_embeddings(int port) root = body ? parse_embeddings_body(body, 1) : NULL; if (root) { cJSON *emb = cJSON_GetObjectItemCaseSensitive( - cJSON_GetArrayItem(cJSON_GetObjectItemCaseSensitive(root, "data"), - 0), "embedding"); + cJSON_GetArrayItem(cJSON_GetObjectItemCaseSensitive(root, "data"), 0), "embedding"); TEST_ASSERT(cJSON_IsArray(emb) && cJSON_GetArraySize(emb) == 128); cJSON_Delete(root); } @@ -548,28 +555,24 @@ static void test_http_embeddings(int port) NULL, &body) == 200); cJSON *qwen_mrl = body ? parse_embeddings_body(body, 1) : NULL; if (qwen_full && qwen_mrl) { - cJSON *full_data = - cJSON_GetObjectItemCaseSensitive(qwen_full, "data"); - cJSON *mrl_data = - cJSON_GetObjectItemCaseSensitive(qwen_mrl, "data"); - cJSON *full_emb = cJSON_GetObjectItemCaseSensitive( - cJSON_GetArrayItem(full_data, 0), "embedding"); - cJSON *mrl_emb = cJSON_GetObjectItemCaseSensitive( - cJSON_GetArrayItem(mrl_data, 0), "embedding"); - TEST_ASSERT(cJSON_IsArray(full_emb) && - cJSON_GetArraySize(full_emb) == 1024); - TEST_ASSERT(cJSON_IsArray(mrl_emb) && - cJSON_GetArraySize(mrl_emb) == 32); + cJSON *full_data = cJSON_GetObjectItemCaseSensitive(qwen_full, "data"); + cJSON *mrl_data = cJSON_GetObjectItemCaseSensitive(qwen_mrl, "data"); + cJSON *full_emb = + cJSON_GetObjectItemCaseSensitive(cJSON_GetArrayItem(full_data, 0), "embedding"); + cJSON *mrl_emb = + cJSON_GetObjectItemCaseSensitive(cJSON_GetArrayItem(mrl_data, 0), "embedding"); + TEST_ASSERT(cJSON_IsArray(full_emb) && cJSON_GetArraySize(full_emb) == 1024); + TEST_ASSERT(cJSON_IsArray(mrl_emb) && cJSON_GetArraySize(mrl_emb) == 32); double prefix_diff = 0.0; for (int i = 0; i < 32; i++) { double d = fabs(cJSON_GetArrayItem(full_emb, i)->valuedouble - cJSON_GetArrayItem(mrl_emb, i)->valuedouble); - if (d > prefix_diff) prefix_diff = d; + if (d > prefix_diff) + prefix_diff = d; } TEST_ASSERT(prefix_diff > 0.0); cJSON *usage = cJSON_GetObjectItemCaseSensitive(qwen_mrl, "usage"); - cJSON *total = usage - ? cJSON_GetObjectItemCaseSensitive(usage, "total_tokens") : NULL; + cJSON *total = usage ? cJSON_GetObjectItemCaseSensitive(usage, "total_tokens") : NULL; TEST_ASSERT(total && total->valueint == 2); } cJSON_Delete(qwen_full); @@ -585,11 +588,12 @@ static void test_http_embeddings(int port) fds[i] = tcp_connect(port); TEST_ASSERT(fds[i] >= 0); if (fds[i] >= 0) - TEST_ASSERT(http_send(fds[i], "POST", "/v1/embeddings", - TEST_API_KEY, ok_req, NULL) == 0); + TEST_ASSERT(http_send(fds[i], "POST", "/v1/embeddings", TEST_API_KEY, ok_req, NULL) == + 0); } for (int i = 0; i < N_CONC; i++) { - if (fds[i] < 0) continue; + if (fds[i] < 0) + continue; char *resp = NULL; TEST_ASSERT(http_recv(fds[i], &resp) == 200); free(resp); @@ -598,8 +602,7 @@ static void test_http_embeddings(int port) } /* data[di].data[ci].embedding from a contextual response. */ -static cJSON *ctx_chunk_emb(cJSON *root, int di, int ci) -{ +static cJSON *ctx_chunk_emb(cJSON *root, int di, int ci) { cJSON *docs = cJSON_GetObjectItemCaseSensitive(root, "data"); cJSON *doc = cJSON_GetArrayItem(docs, di); cJSON *chunks = cJSON_GetObjectItemCaseSensitive(doc, "data"); @@ -607,40 +610,40 @@ static cJSON *ctx_chunk_emb(cJSON *root, int di, int ci) return cJSON_GetObjectItemCaseSensitive(item, "embedding"); } -static double emb_max_absdiff(cJSON *a, cJSON *b) -{ - if (!cJSON_IsArray(a) || !cJSON_IsArray(b)) return 1e9; +static double emb_max_absdiff(cJSON *a, cJSON *b) { + if (!cJSON_IsArray(a) || !cJSON_IsArray(b)) + return 1e9; int n = cJSON_GetArraySize(a); - if (n != cJSON_GetArraySize(b)) return 1e9; + if (n != cJSON_GetArraySize(b)) + return 1e9; double m = 0.0; for (int i = 0; i < n; i++) { - double d = fabs(cJSON_GetArrayItem(a, i)->valuedouble - - cJSON_GetArrayItem(b, i)->valuedouble); - if (d > m) m = d; + double d = + fabs(cJSON_GetArrayItem(a, i)->valuedouble - cJSON_GetArrayItem(b, i)->valuedouble); + if (d > m) + m = d; } return m; } -static void test_http_contextual(int port) -{ +static void test_http_contextual(int port) { char *body = NULL; /* Input must be an array of chunk arrays; chunks must be non-empty. */ - TEST_ASSERT(http_req(port, "POST", "/v1/contextualizedembeddings", - TEST_API_KEY, + TEST_ASSERT(http_req(port, "POST", "/v1/contextualizedembeddings", TEST_API_KEY, "{\"model\":\"pplx-embed-context-v1-0.6b\"," - "\"input\":\"hi\"}", NULL, NULL) == 422); - TEST_ASSERT(http_req(port, "POST", "/v1/contextualizedembeddings", - TEST_API_KEY, + "\"input\":\"hi\"}", + NULL, NULL) == 422); + TEST_ASSERT(http_req(port, "POST", "/v1/contextualizedembeddings", TEST_API_KEY, "{\"model\":\"pplx-embed-context-v1-0.6b\"," - "\"input\":[[\"\"]]}", NULL, NULL) == 422); + "\"input\":[[\"\"]]}", + NULL, NULL) == 422); /* Happy path: chunks concatenate per document with one separator token * between them, one embedding per chunk in document order. Fixture * token counts: hello=1, world=3, held=2, one separator in doc 0, so * usage must report 7. */ - TEST_ASSERT(http_req(port, "POST", "/v1/contextualizedembeddings", - TEST_API_KEY, + TEST_ASSERT(http_req(port, "POST", "/v1/contextualizedembeddings", TEST_API_KEY, "{\"model\":\"pplx-embed-context-v1-0.6b\"," "\"input\":[[\"hello\",\"world\"],[\"held\"]]," "\"encoding_format\":\"float\"}", @@ -664,21 +667,18 @@ static void test_http_contextual(int port) cJSON *cidx = cJSON_GetObjectItemCaseSensitive(item, "index"); cJSON *emb = ctx_chunk_emb(multi, di, ci); TEST_ASSERT(cidx && cidx->valueint == ci); - TEST_ASSERT(cJSON_IsArray(emb) && - cJSON_GetArraySize(emb) == 1024); + TEST_ASSERT(cJSON_IsArray(emb) && cJSON_GetArraySize(emb) == 1024); } } cJSON *usage = cJSON_GetObjectItemCaseSensitive(multi, "usage"); - cJSON *total = usage - ? cJSON_GetObjectItemCaseSensitive(usage, "total_tokens") : NULL; + cJSON *total = usage ? cJSON_GetObjectItemCaseSensitive(usage, "total_tokens") : NULL; TEST_ASSERT(total && total->valueint == 7); } /* A single-chunk document has no separators and its pooling span covers * the whole sequence, so it must match the standard embedding of the * same text from the same weights. */ - TEST_ASSERT(http_req(port, "POST", "/v1/contextualizedembeddings", - TEST_API_KEY, + TEST_ASSERT(http_req(port, "POST", "/v1/contextualizedembeddings", TEST_API_KEY, "{\"model\":\"pplx-embed-context-v1-0.6b\"," "\"input\":[[\"hello\"]]," "\"encoding_format\":\"float\"}", @@ -699,8 +699,7 @@ static void test_http_contextual(int port) if (multi && single && std) { cJSON *std_emb = cJSON_GetObjectItemCaseSensitive( - cJSON_GetArrayItem(cJSON_GetObjectItemCaseSensitive(std, "data"), - 0), "embedding"); + cJSON_GetArrayItem(cJSON_GetObjectItemCaseSensitive(std, "data"), 0), "embedding"); cJSON *alone = ctx_chunk_emb(single, 0, 0); cJSON *in_doc = ctx_chunk_emb(multi, 0, 0); TEST_ASSERT(emb_max_absdiff(std_emb, alone) < 1e-4); @@ -713,8 +712,7 @@ static void test_http_contextual(int port) cJSON_Delete(std); } -static void test_http_rerank(int port) -{ +static void test_http_rerank(int port) { char *body = NULL; TEST_ASSERT(http_req(port, "POST", "/v1/rerank", TEST_API_KEY, @@ -735,55 +733,45 @@ static void test_http_rerank(int port) "\"top_n\":1,\"top_k\":1}", NULL, NULL) == 422); - const char *request = - "{\"model\":\"pplx-embed-v1-late-0.6b\"," - "\"query\":\"hello\",\"documents\":[" - "\"hello!\",\"world\",\"quoted \\\"document\\\"\"]," - "\"top_n\":2,\"return_documents\":true}"; - TEST_ASSERT(http_req(port, "POST", "/v1/rerank", TEST_API_KEY, request, - NULL, &body) == 200); + const char *request = "{\"model\":\"pplx-embed-v1-late-0.6b\"," + "\"query\":\"hello\",\"documents\":[" + "\"hello!\",\"world\",\"quoted \\\"document\\\"\"]," + "\"top_n\":2,\"return_documents\":true}"; + TEST_ASSERT(http_req(port, "POST", "/v1/rerank", TEST_API_KEY, request, NULL, &body) == 200); cJSON *root = body ? cJSON_Parse(body) : NULL; TEST_ASSERT(root != NULL); if (root) { cJSON *object = cJSON_GetObjectItemCaseSensitive(root, "object"); cJSON *model = cJSON_GetObjectItemCaseSensitive(root, "model"); cJSON *results = cJSON_GetObjectItemCaseSensitive(root, "results"); - TEST_ASSERT(cJSON_IsString(object) && - strcmp(object->valuestring, "list") == 0); + TEST_ASSERT(cJSON_IsString(object) && strcmp(object->valuestring, "list") == 0); TEST_ASSERT(cJSON_IsString(model) && - strcmp(model->valuestring, - "pplx-embed-v1-late-0.6b") == 0); - TEST_ASSERT(cJSON_IsArray(results) && - cJSON_GetArraySize(results) == 2); + strcmp(model->valuestring, "pplx-embed-v1-late-0.6b") == 0); + TEST_ASSERT(cJSON_IsArray(results) && cJSON_GetArraySize(results) == 2); double previous = INFINITY; int seen[3] = {0}; cJSON *result; cJSON_ArrayForEach(result, results) { - cJSON *index = - cJSON_GetObjectItemCaseSensitive(result, "index"); - cJSON *score = - cJSON_GetObjectItemCaseSensitive(result, "relevance_score"); - cJSON *document = - cJSON_GetObjectItemCaseSensitive(result, "document"); - TEST_ASSERT(cJSON_IsNumber(index) && - index->valueint >= 0 && index->valueint < 3); - TEST_ASSERT(cJSON_IsNumber(score) && - score->valuedouble <= previous); + cJSON *index = cJSON_GetObjectItemCaseSensitive(result, "index"); + cJSON *score = cJSON_GetObjectItemCaseSensitive(result, "relevance_score"); + cJSON *document = cJSON_GetObjectItemCaseSensitive(result, "document"); + TEST_ASSERT(cJSON_IsNumber(index) && index->valueint >= 0 && index->valueint < 3); + TEST_ASSERT(cJSON_IsNumber(score) && score->valuedouble <= previous); TEST_ASSERT(cJSON_IsString(document)); - if (cJSON_IsNumber(index) && index->valueint >= 0 && - index->valueint < 3) + if (cJSON_IsNumber(index) && index->valueint >= 0 && index->valueint < 3) seen[index->valueint]++; - if (cJSON_IsNumber(score)) previous = score->valuedouble; + if (cJSON_IsNumber(score)) + previous = score->valuedouble; } TEST_ASSERT(seen[0] + seen[1] + seen[2] == 2); cJSON *usage = cJSON_GetObjectItemCaseSensitive(root, "usage"); - cJSON *query_tokens = usage - ? cJSON_GetObjectItemCaseSensitive(usage, "query_tokens") : NULL; - cJSON *document_tokens = usage - ? cJSON_GetObjectItemCaseSensitive(usage, "document_tokens") : NULL; - cJSON *total_tokens = usage - ? cJSON_GetObjectItemCaseSensitive(usage, "total_tokens") : NULL; + cJSON *query_tokens = + usage ? cJSON_GetObjectItemCaseSensitive(usage, "query_tokens") : NULL; + cJSON *document_tokens = + usage ? cJSON_GetObjectItemCaseSensitive(usage, "document_tokens") : NULL; + cJSON *total_tokens = + usage ? cJSON_GetObjectItemCaseSensitive(usage, "total_tokens") : NULL; TEST_ASSERT(query_tokens && document_tokens && total_tokens); if (query_tokens && document_tokens && total_tokens) { TEST_ASSERT(cJSON_IsNumber(query_tokens)); @@ -808,18 +796,16 @@ static void test_http_rerank(int port) TEST_ASSERT(root != NULL); if (root) { cJSON *results = cJSON_GetObjectItemCaseSensitive(root, "results"); - TEST_ASSERT(cJSON_IsArray(results) && - cJSON_GetArraySize(results) == 1); - cJSON *document = cJSON_GetObjectItemCaseSensitive( - cJSON_GetArrayItem(results, 0), "document"); + TEST_ASSERT(cJSON_IsArray(results) && cJSON_GetArraySize(results) == 1); + cJSON *document = + cJSON_GetObjectItemCaseSensitive(cJSON_GetArrayItem(results, 0), "document"); TEST_ASSERT(document == NULL); cJSON_Delete(root); } free(body); } -static void test_http_server(void) -{ +static void test_http_server(void) { char dir[1024], late_dir[1024], qwen_dir[1024]; snprintf(dir, sizeof(dir), "%s/embed-srv-test-XXXXXX", getenv("TMPDIR") ? getenv("TMPDIR") : "/tmp"); @@ -845,14 +831,11 @@ static void test_http_server(void) /* The 0.6b server slot requires hidden_size 1024; everything else stays * tiny. The model vocab covers every tokenizer fixture id. */ tm_dims_t dims = {1024, 2, 1, 64, 8, TF_VOCAB_SIZE}; - if (tf_write_vocab(dir) != 0 || - tm_write_model_dims(dir, "F32", &dims) != 0 || - tf_write_vocab(late_dir) != 0 || - tm_write_model_dims(late_dir, "F32", &dims) != 0 || + if (tf_write_vocab(dir) != 0 || tm_write_model_dims(dir, "F32", &dims) != 0 || + tf_write_vocab(late_dir) != 0 || tm_write_model_dims(late_dir, "F32", &dims) != 0 || tm_write_late_projection(late_dir, "F32", 128, 1024) != 0 || tf_write_vocab(qwen_dir) != 0 || - tm_write_qwen3_model_dims(qwen_dir, "F32", &dims, - TF_EOT_ID) != 0) { + tm_write_qwen3_model_dims(qwen_dir, "F32", &dims, TF_EOT_ID) != 0) { fprintf(stderr, "FAIL: fixture write\n"); test_failures++; return; @@ -888,9 +871,13 @@ static void test_http_server(void) } int up = 0; - for (int i = 0; i < 200; i++) { /* up to ~10 s */ + for (int i = 0; i < 200; i++) { /* up to ~10 s */ int fd = tcp_connect(port); - if (fd >= 0) { close(fd); up = 1; break; } + if (fd >= 0) { + close(fd); + up = 1; + break; + } usleep(50 * 1000); } TEST_ASSERT(up); @@ -905,8 +892,7 @@ static void test_http_server(void) TEST_ASSERT(ctx.rc == 0); } -int main(void) -{ +int main(void) { test_base64_rfc4648(); test_quantize_int8_tanh(); test_encode_embedding_int8(); diff --git a/tests/test_tokenizer.c b/tests/test_tokenizer.c index bed5b7d..5920977 100644 --- a/tests/test_tokenizer.c +++ b/tests/test_tokenizer.c @@ -16,31 +16,33 @@ static int failures = 0; int qwen_verbose = 0; -#define TEST_ASSERT(cond) do { \ - if (!(cond)) { \ - fprintf(stderr, "FAIL %s:%d: %s\n", __FILE__, __LINE__, #cond); \ - failures++; \ - } \ -} while (0) +#define TEST_ASSERT(cond) \ + do { \ + if (!(cond)) { \ + fprintf(stderr, "FAIL %s:%d: %s\n", __FILE__, __LINE__, #cond); \ + failures++; \ + } \ + } while (0) static void check_ids(const qwen_tokenizer_t *tok, - qwen_tokenizer_workspace_t *ws, const char *text, - const int *want, int n_want, const char *label) -{ + qwen_tokenizer_workspace_t *ws, + const char *text, + const int *want, + int n_want, + const char *label) { int n = 0; - int *ids = qwen_tokenizer_encode_with_workspace( - (qwen_tokenizer_t *)tok, ws, text, &n); + int *ids = qwen_tokenizer_encode_with_workspace((qwen_tokenizer_t *)tok, ws, text, &n); TEST_ASSERT(ids != NULL); - if (!ids) return; + if (!ids) + return; if (n != n_want) { - fprintf(stderr, "FAIL %s \"%s\": got %d ids want %d\n", - label, text, n, n_want); + fprintf(stderr, "FAIL %s \"%s\": got %d ids want %d\n", label, text, n, n_want); failures++; } else { for (int i = 0; i < n; i++) { if (ids[i] != want[i]) { - fprintf(stderr, "FAIL %s \"%s\": ids[%d]=%d want %d\n", - label, text, i, ids[i], want[i]); + fprintf(stderr, "FAIL %s \"%s\": ids[%d]=%d want %d\n", label, text, i, ids[i], + want[i]); failures++; } } @@ -51,25 +53,30 @@ static void check_ids(const qwen_tokenizer_t *tok, /* Concatenated decode of an encoding must reproduce the (NFC-normalized) * input - byte-level BPE is lossless. */ static void check_roundtrip(const qwen_tokenizer_t *tok, - qwen_tokenizer_workspace_t *ws, const char *text, - const char *want, const char *label) -{ + qwen_tokenizer_workspace_t *ws, + const char *text, + const char *want, + const char *label) { int n = 0; - int *ids = qwen_tokenizer_encode_with_workspace( - (qwen_tokenizer_t *)tok, ws, text, &n); + int *ids = qwen_tokenizer_encode_with_workspace((qwen_tokenizer_t *)tok, ws, text, &n); TEST_ASSERT(ids != NULL && n > 0); - if (!ids) return; + if (!ids) + return; size_t cap = 1; for (int i = 0; i < n; i++) cap += strlen(qwen_tokenizer_decode(tok, ids[i])); char *got = (char *)malloc(cap); - if (!got) { free(ids); TEST_ASSERT(0); return; } + if (!got) { + free(ids); + TEST_ASSERT(0); + return; + } got[0] = '\0'; for (int i = 0; i < n; i++) strcat(got, qwen_tokenizer_decode(tok, ids[i])); if (strcmp(got, want) != 0) { - fprintf(stderr, "FAIL %s roundtrip \"%s\": got \"%s\" want \"%s\"\n", - label, text, got, want); + fprintf(stderr, "FAIL %s roundtrip \"%s\": got \"%s\" want \"%s\"\n", label, text, got, + want); failures++; } free(got); @@ -79,15 +86,21 @@ static void check_roundtrip(const qwen_tokenizer_t *tok, /* Two spellings of the same text must encode identically. */ static void check_same_ids(const qwen_tokenizer_t *tok, qwen_tokenizer_workspace_t *ws, - const char *a, const char *b, const char *label) -{ + const char *a, + const char *b, + const char *label) { int na = 0, nb = 0; - int *ia = qwen_tokenizer_encode_with_workspace( - (qwen_tokenizer_t *)tok, ws, a, &na); - if (!ia) { TEST_ASSERT(0); return; } - int *ib = qwen_tokenizer_encode_with_workspace( - (qwen_tokenizer_t *)tok, ws, b, &nb); - if (!ib) { free(ia); TEST_ASSERT(0); return; } + int *ia = qwen_tokenizer_encode_with_workspace((qwen_tokenizer_t *)tok, ws, a, &na); + if (!ia) { + TEST_ASSERT(0); + return; + } + int *ib = qwen_tokenizer_encode_with_workspace((qwen_tokenizer_t *)tok, ws, b, &nb); + if (!ib) { + free(ia); + TEST_ASSERT(0); + return; + } if (na != nb || memcmp(ia, ib, (size_t)na * sizeof(int)) != 0) { fprintf(stderr, "FAIL %s same-ids \"%s\" vs \"%s\"\n", label, a, b); failures++; @@ -96,9 +109,8 @@ static void check_same_ids(const qwen_tokenizer_t *tok, free(ia); } -static void run_cases(const qwen_tokenizer_t *tok, - qwen_tokenizer_workspace_t *ws, const char *label) -{ +static void +run_cases(const qwen_tokenizer_t *tok, qwen_tokenizer_workspace_t *ws, const char *label) { /* h e l l o -> he -> he,ll -> hell -> hello */ const int hello[] = {259}; check_ids(tok, ws, "hello", hello, 1, label); @@ -147,14 +159,14 @@ static void run_cases(const qwen_tokenizer_t *tok, /* NFC: decomposed base+combining-mark spellings must match the * precomposed codepoint (one pair per mark in the composition table). */ - check_same_ids(tok, ws, "e\xCC\x81", "\xC3\xA9", label); /* acute */ - check_same_ids(tok, ws, "a\xCC\x80", "\xC3\xA0", label); /* grave */ - check_same_ids(tok, ws, "e\xCC\x82", "\xC3\xAA", label); /* circumflex */ - check_same_ids(tok, ws, "n\xCC\x83", "\xC3\xB1", label); /* tilde */ - check_same_ids(tok, ws, "u\xCC\x88", "\xC3\xBC", label); /* diaeresis */ - check_same_ids(tok, ws, "a\xCC\x8A", "\xC3\xA5", label); /* ring */ - check_same_ids(tok, ws, "c\xCC\xA7", "\xC3\xA7", label); /* cedilla */ - check_same_ids(tok, ws, "A\xCC\x80", "\xC3\x80", label); /* upper */ + check_same_ids(tok, ws, "e\xCC\x81", "\xC3\xA9", label); /* acute */ + check_same_ids(tok, ws, "a\xCC\x80", "\xC3\xA0", label); /* grave */ + check_same_ids(tok, ws, "e\xCC\x82", "\xC3\xAA", label); /* circumflex */ + check_same_ids(tok, ws, "n\xCC\x83", "\xC3\xB1", label); /* tilde */ + check_same_ids(tok, ws, "u\xCC\x88", "\xC3\xBC", label); /* diaeresis */ + check_same_ids(tok, ws, "a\xCC\x8A", "\xC3\xA5", label); /* ring */ + check_same_ids(tok, ws, "c\xCC\xA7", "\xC3\xA7", label); /* cedilla */ + check_same_ids(tok, ws, "A\xCC\x80", "\xC3\x80", label); /* upper */ check_roundtrip(tok, ws, "cafe\xCC\x81", "caf\xC3\xA9", label); /* 'x' + acute has no composition: bytes pass through untouched. */ check_roundtrip(tok, ws, "x\xCC\x81", "x\xCC\x81", label); @@ -177,30 +189,37 @@ static void run_cases(const qwen_tokenizer_t *tok, check_roundtrip(tok, ws, "\xF0\x9F\x98\x80", "\xF0\x9F\x98\x80", label); check_roundtrip(tok, ws, "tab\tand\nnl 42!", "tab\tand\nnl 42!", label); const int ellipsis[] = {'a', 0xE2, 0x80, 0xA6, 'b'}; - check_ids(tok, ws, "a\xE2\x80\xA6" "b", ellipsis, 5, label); + check_ids(tok, ws, + "a\xE2\x80\xA6" + "b", + ellipsis, 5, label); } -static void write_file(const char *dir, const char *name, const char *body) -{ +static void write_file(const char *dir, const char *name, const char *body) { char path[2048]; snprintf(path, sizeof(path), "%s/%s", dir, name); FILE *f = fopen(path, "w"); - if (!f) { perror(path); exit(2); } + if (!f) { + perror(path); + exit(2); + } fputs(body, f); fclose(f); } /* Loader rejection paths and vocab.json escape handling, in throwaway * subdirectories of the main fixture dir. */ -static void test_loader_edges(const char *root) -{ +static void test_loader_edges(const char *root) { char dir[1152], path[2048]; TEST_ASSERT(qwen_tokenizer_load("/nonexistent/vocab.json") == NULL); /* Not a JSON object. */ snprintf(dir, sizeof(dir), "%s/badjson", root); - if (mkdir(dir, 0755) != 0) { perror(dir); exit(2); } + if (mkdir(dir, 0755) != 0) { + perror(dir); + exit(2); + } write_file(dir, "vocab.json", "[1, 2]\n"); write_file(dir, "merges.txt", "a b\n"); snprintf(path, sizeof(path), "%s/vocab.json", dir); @@ -209,7 +228,10 @@ static void test_loader_edges(const char *root) /* Missing or pair-less merges.txt is tolerated: the tokenizer loads * and encoding falls back to byte/vocab-level (no merges applied). */ snprintf(dir, sizeof(dir), "%s/nomerges", root); - if (mkdir(dir, 0755) != 0) { perror(dir); exit(2); } + if (mkdir(dir, 0755) != 0) { + perror(dir); + exit(2); + } write_file(dir, "vocab.json", "{\"a\":0,\"b\":1}\n"); snprintf(path, sizeof(path), "%s/vocab.json", dir); qwen_tokenizer_t *nom = qwen_tokenizer_load(path); @@ -225,7 +247,10 @@ static void test_loader_edges(const char *root) } snprintf(dir, sizeof(dir), "%s/emptymerges", root); - if (mkdir(dir, 0755) != 0) { perror(dir); exit(2); } + if (mkdir(dir, 0755) != 0) { + perror(dir); + exit(2); + } write_file(dir, "vocab.json", "{\"a\":0,\"b\":1}\n"); write_file(dir, "merges.txt", "#version: 0.2\n\n \n"); snprintf(path, sizeof(path), "%s/vocab.json", dir); @@ -235,7 +260,10 @@ static void test_loader_edges(const char *root) /* Escape forms in vocab keys; a malformed merge line is skipped. */ snprintf(dir, sizeof(dir), "%s/escapes", root); - if (mkdir(dir, 0755) != 0) { perror(dir); exit(2); } + if (mkdir(dir, 0755) != 0) { + perror(dir); + exit(2); + } write_file(dir, "vocab.json", "{\"a\":0, \"b\":1, \"ab\":2,\n" " \"\\u20AC\":3, \"x\\ny\":4, \"q\\\"r\":5,\n" @@ -264,12 +292,14 @@ static void test_loader_edges(const char *root) qwen_tokenizer_t *rel = qwen_tokenizer_load("vocab.json"); TEST_ASSERT(rel != NULL); qwen_tokenizer_free(rel); - if (chdir(saved_cwd) != 0) { perror("chdir back"); exit(2); } + if (chdir(saved_cwd) != 0) { + perror("chdir back"); + exit(2); + } } } -int main(void) -{ +int main(void) { char dir[1024]; snprintf(dir, sizeof(dir), "%s/embed-tok-test-XXXXXX", getenv("TMPDIR") ? getenv("TMPDIR") : "/tmp"); @@ -282,7 +312,8 @@ int main(void) snprintf(vocab_path, sizeof(vocab_path), "%s/vocab.json", dir); qwen_tokenizer_t *tok = qwen_tokenizer_load(vocab_path); TEST_ASSERT(tok != NULL); - if (!tok) return 1; + if (!tok) + return 1; /* The complete byte alphabet means the integer-pair tables must exist. */ TEST_ASSERT(tok->int_merges != NULL); @@ -312,21 +343,17 @@ int main(void) { int out[8]; int n = -1; - TEST_ASSERT(qwen_tokenizer_encode_into(tok, ws, "hello world", out, 8, - &n) == 0 && n == 3); + TEST_ASSERT(qwen_tokenizer_encode_into(tok, ws, "hello world", out, 8, &n) == 0 && n == 3); TEST_ASSERT(out[0] == 259 && out[1] == 263 && out[2] == 262); /* On overflow the prefix is written and n reports the total * required count, so a caller can re-size and retry. */ n = -1; - TEST_ASSERT(qwen_tokenizer_encode_into(tok, ws, "hello world", out, 2, - &n) == -2 && n == 3); + TEST_ASSERT(qwen_tokenizer_encode_into(tok, ws, "hello world", out, 2, &n) == -2 && n == 3); TEST_ASSERT(out[0] == 259 && out[1] == 263); n = -1; - TEST_ASSERT(qwen_tokenizer_encode_into(tok, ws, "hello world", out, 0, - &n) == -2 && n == 3); + TEST_ASSERT(qwen_tokenizer_encode_into(tok, ws, "hello world", out, 0, &n) == -2 && n == 3); n = -1; - TEST_ASSERT(qwen_tokenizer_encode_into(tok, ws, "", out, 8, &n) == 0 && - n == 0); + TEST_ASSERT(qwen_tokenizer_encode_into(tok, ws, "", out, 8, &n) == 0 && n == 0); TEST_ASSERT(qwen_tokenizer_encode_into(tok, ws, NULL, out, 8, &n) == -1); TEST_ASSERT(qwen_tokenizer_encode_into(NULL, ws, "x", out, 8, &n) == -1); } diff --git a/tests/test_workspace.c b/tests/test_workspace.c index 4f83936..a906689 100644 --- a/tests/test_workspace.c +++ b/tests/test_workspace.c @@ -14,12 +14,12 @@ #include <string.h> #include <unistd.h> -static float max_abs_diff(const float *a, const float *b, int n) -{ +static float max_abs_diff(const float *a, const float *b, int n) { float m = 0.0f; for (int i = 0; i < n; i++) { float d = fabsf(a[i] - b[i]); - if (d > m) m = d; + if (d > m) + m = d; } return m; } @@ -28,11 +28,11 @@ static float max_abs_diff(const float *a, const float *b, int n) * results validated in main. reference is the batched row for ids0. */ static int check_alloc_and_pooling_variants(const embed_model_t *model, embed_workspace_t *ws, - const int *ids0, int n0, + const int *ids0, + int n0, const float *states, const float *reference, - const embed_config_t *cfg) -{ + const embed_config_t *cfg) { int rc = 1; int dim = cfg->hidden_size; float *emb = NULL, *fwd = NULL, *pooled = NULL, *span_out = NULL; @@ -73,22 +73,19 @@ static int check_alloc_and_pooling_variants(const embed_model_t *model, /* One span covering the whole sequence pools to the embedding, and * two halves recombine into it by token-count weighting. */ embed_span_t whole = {0, n0}; - if (embed_model_encode_spans(model, ws, ids0, n0, &whole, 1, - span_out) != 0 || + if (embed_model_encode_spans(model, ws, ids0, n0, &whole, 1, span_out) != 0 || max_abs_diff(span_out, reference, dim) > 0.00005f) { fprintf(stderr, "whole-sequence span disagrees with embedding\n"); goto done; } int h = n0 / 2; embed_span_t halves[2] = {{0, h}, {h, n0 - h}}; - if (embed_model_encode_spans(model, ws, ids0, n0, halves, 2, - span_out) != 0) { + if (embed_model_encode_spans(model, ws, ids0, n0, halves, 2, span_out) != 0) { fprintf(stderr, "embed_model_encode_spans failed for two spans\n"); goto done; } for (int d = 0; d < dim; d++) { - float combined = ((float)h * span_out[d] + - (float)(n0 - h) * span_out[dim + d]) / (float)n0; + float combined = ((float)h * span_out[d] + (float)(n0 - h) * span_out[dim + d]) / (float)n0; if (fabsf(combined - reference[d]) > 0.00005f) { fprintf(stderr, "half spans do not recombine into embedding\n"); goto done; @@ -108,10 +105,10 @@ done: * doc0 is chunks 0 and 1 joined by a separator token; doc1 is chunk 2. */ static int check_spans_batch_parity(const embed_model_t *model, embed_workspace_t *ws, - int *const ids[], const int *ntok, + int *const ids[], + const int *ntok, int separator_id, - const embed_config_t *cfg) -{ + const embed_config_t *cfg) { int rc = 1; int dim = cfg->hidden_size; int n0 = ntok[0] + 1 + ntok[1]; @@ -128,10 +125,9 @@ static int check_spans_batch_parity(const embed_model_t *model, embed_span_t spans0[2] = {{0, ntok[0]}, {ntok[0] + 1, ntok[1]}}; embed_span_t span1 = {0, ntok[2]}; - if (embed_model_encode_spans(model, ws, doc0, n0, spans0, 2, - expected) != 0 || + if (embed_model_encode_spans(model, ws, doc0, n0, spans0, 2, expected) != 0 || embed_model_encode_spans(model, ws, ids[2], ntok[2], &span1, 1, - expected + (size_t)2 * dim) != 0) { + expected + (size_t)2 * dim) != 0) { fprintf(stderr, "singleton contextual embedding failed\n"); goto done; } @@ -164,8 +160,7 @@ done: return rc; } -int main(int argc, char **argv) -{ +int main(int argc, char **argv) { char fixture_dir[1024]; const char *model_dir; if (argc == 2) { @@ -214,12 +209,10 @@ int main(int argc, char **argv) return 1; } - const char *texts[] = { - "query: what is the capital of France?", - "document: Paris is the capital of France.", - "document: Berlin is the capital of Germany.", - "document: Istanbul is a major city in Turkey." - }; + const char *texts[] = {"query: what is the capital of France?", + "document: Paris is the capital of France.", + "document: Berlin is the capital of Germany.", + "document: Istanbul is a major city in Turkey."}; const int batch = (int)(sizeof(texts) / sizeof(texts[0])); /* Initialized before the first `goto fail` so the cleanup loop never * sees indeterminate pointers. */ @@ -291,8 +284,10 @@ int main(int argc, char **argv) const float *row = batched + (size_t)i * dim; float diff = max_abs_diff(row, single, dim); float cos = embed_cosine_similarity(row, single, dim); - if (diff > worst_diff) worst_diff = diff; - if (cos < worst_cos) worst_cos = cos; + if (diff > worst_diff) + worst_diff = diff; + if (cos < worst_cos) + worst_cos = cos; } memcpy(normalized, batched, (size_t)dim * sizeof(float)); @@ -308,8 +303,7 @@ int main(int argc, char **argv) for (int i = 0; i < dim; i++) normalized_norm_sq += normalized[i] * normalized[i]; if (fabsf(normalized_norm_sq - 1.0f) > 0.00005f) { - fprintf(stderr, "embed_l2_normalize produced norm_sq=%g\n", - normalized_norm_sq); + fprintf(stderr, "embed_l2_normalize produced norm_sq=%g\n", normalized_norm_sq); free(states); free(normalized); free(single); @@ -326,8 +320,8 @@ int main(int argc, char **argv) goto fail; } - if (check_alloc_and_pooling_variants(model, batch_ws, ids[0], ntok[0], - states, batched, cfg) != 0) { + if (check_alloc_and_pooling_variants(model, batch_ws, ids[0], ntok[0], states, batched, cfg) != + 0) { free(states); free(normalized); free(single); @@ -338,9 +332,9 @@ int main(int argc, char **argv) /* Same separator resolution as the server: the fixture vocab defines * <|endoftext|>, real snapshots use the reserved id. */ int sep_id = qwen_tokenizer_token_id(tok, "<|endoftext|>"); - if (sep_id < 0) sep_id = EMBED_CONTEXT_SEPARATOR_TOKEN_ID; - if (check_spans_batch_parity(model, batch_ws, ids, ntok, sep_id, - cfg) != 0) { + if (sep_id < 0) + sep_id = EMBED_CONTEXT_SEPARATOR_TOKEN_ID; + if (check_spans_batch_parity(model, batch_ws, ids, ntok, sep_id, cfg) != 0) { free(states); free(normalized); free(single); @@ -353,24 +347,26 @@ int main(int argc, char **argv) free(single); free(batched); - for (int i = 0; i < batch; i++) free(ids[i]); + for (int i = 0; i < batch; i++) + free(ids[i]); qwen_tokenizer_free(tok); embed_workspace_free(single_ws); embed_workspace_free(batch_ws); embed_model_free(model); if (worst_diff > 0.00005f || worst_cos < 0.99999f) { - fprintf(stderr, "workspace API parity failed: max_abs_diff=%g cosine=%g\n", - worst_diff, worst_cos); + fprintf(stderr, "workspace API parity failed: max_abs_diff=%g cosine=%g\n", worst_diff, + worst_cos); return 1; } - printf("ok: workspace API dim=%d batch=%d workspace=%zu bytes max_abs_diff=%g cosine=%g\n", - dim, batch, batch_ws_bytes, worst_diff, worst_cos); + printf("ok: workspace API dim=%d batch=%d workspace=%zu bytes max_abs_diff=%g cosine=%g\n", dim, + batch, batch_ws_bytes, worst_diff, worst_cos); return 0; fail: - for (int i = 0; i < batch; i++) free(ids[i]); + for (int i = 0; i < batch; i++) + free(ids[i]); qwen_tokenizer_free(tok); embed_workspace_free(single_ws); embed_workspace_free(batch_ws); diff --git a/tests/tiny_model.h b/tests/tiny_model.h index d4ae431..42542c1 100644 --- a/tests/tiny_model.h +++ b/tests/tiny_model.h @@ -14,15 +14,13 @@ #include <sys/stat.h> #include <sys/types.h> -static uint16_t tm_f32_to_bf16(float x) -{ +static uint16_t tm_f32_to_bf16(float x) { uint32_t u; memcpy(&u, &x, sizeof(u)); return (uint16_t)(u >> 16); } -static float tm_bf16_to_f32(uint16_t b) -{ +static float tm_bf16_to_f32(uint16_t b) { uint32_t u = ((uint32_t)b) << 16; float f; memcpy(&f, &u, sizeof(f)); @@ -30,8 +28,7 @@ static float tm_bf16_to_f32(uint16_t b) } /* Deterministic, BF16-exact value for element i of tensor `name`. */ -static float tm_value(const char *name, size_t i) -{ +static float tm_value(const char *name, size_t i) { float v; size_t len = strlen(name); if ((len >= 11 && strcmp(name + len - 11, "norm.weight") == 0) || @@ -46,7 +43,7 @@ static float tm_value(const char *name, size_t i) typedef struct { const char *name; - int rows, cols; /* cols == 0 means a 1-D tensor of `rows` values */ + int rows, cols; /* cols == 0 means a 1-D tensor of `rows` values */ } tm_spec_t; /* Always one layer; everything else is free so a test can match a real @@ -57,22 +54,21 @@ typedef struct { enum { TM_N_SPECS = 13 }; -static int tm_write_config(const char *dir, const tm_dims_t *d, - const char *model_type, int eos_token_id) -{ +static int +tm_write_config(const char *dir, const tm_dims_t *d, const char *model_type, int eos_token_id) { char path[2048]; snprintf(path, sizeof(path), "%s/config.json", dir); FILE *f = fopen(path, "w"); - if (!f) return -1; - fprintf(f, "{\"hidden_size\":%d,\"num_hidden_layers\":1," + if (!f) + return -1; + fprintf(f, + "{\"hidden_size\":%d,\"num_hidden_layers\":1," "\"num_attention_heads\":%d,\"num_key_value_heads\":%d," "\"head_dim\":%d,\"intermediate_size\":%d,\"vocab_size\":%d," "\"rms_norm_eps\":1e-6,\"rope_theta\":10000.0", - d->hidden, d->heads, d->kv_heads, d->head_dim, - d->intermediate, d->vocab); + d->hidden, d->heads, d->kv_heads, d->head_dim, d->intermediate, d->vocab); if (model_type) - fprintf(f, ",\"model_type\":\"%s\",\"eos_token_id\":%d", - model_type, eos_token_id); + fprintf(f, ",\"model_type\":\"%s\",\"eos_token_id\":%d", model_type, eos_token_id); fputs("}", f); fclose(f); return 0; @@ -80,9 +76,8 @@ static int tm_write_config(const char *dir, const tm_dims_t *d, /* Write a safetensors file holding `specs` with tm_value() contents. * dtype: "F32" or "BF16". Returns 0 on success. */ -static int tm_write_safetensors(const char *path, const char *dtype, - const tm_spec_t *specs, int n_specs) -{ +static int +tm_write_safetensors(const char *path, const char *dtype, const tm_spec_t *specs, int n_specs) { int bf16 = strcmp(dtype, "BF16") == 0; size_t esize = bf16 ? 2 : 4; @@ -95,23 +90,26 @@ static int tm_write_safetensors(const char *path, const char *dtype, size_t end = doff + n * esize; if (s->cols) hoff += (size_t)snprintf(header + hoff, sizeof(header) - hoff, - "%s\"%s\":{\"dtype\":\"%s\",\"shape\":[%d,%d]," - "\"data_offsets\":[%zu,%zu]}", - t ? "," : "", s->name, dtype, s->rows, s->cols, doff, end); + "%s\"%s\":{\"dtype\":\"%s\",\"shape\":[%d,%d]," + "\"data_offsets\":[%zu,%zu]}", + t ? "," : "", s->name, dtype, s->rows, s->cols, doff, end); else hoff += (size_t)snprintf(header + hoff, sizeof(header) - hoff, - "%s\"%s\":{\"dtype\":\"%s\",\"shape\":[%d]," - "\"data_offsets\":[%zu,%zu]}", - t ? "," : "", s->name, dtype, s->rows, doff, end); + "%s\"%s\":{\"dtype\":\"%s\",\"shape\":[%d]," + "\"data_offsets\":[%zu,%zu]}", + t ? "," : "", s->name, dtype, s->rows, doff, end); doff = end; } hoff += (size_t)snprintf(header + hoff, sizeof(header) - hoff, "}"); - if (hoff >= sizeof(header) - 1) return -1; + if (hoff >= sizeof(header) - 1) + return -1; FILE *f = fopen(path, "wb"); - if (!f) return -1; + if (!f) + return -1; unsigned char len8[8]; - for (int i = 0; i < 8; i++) len8[i] = (unsigned char)(hoff >> (8 * i)); + for (int i = 0; i < 8; i++) + len8[i] = (unsigned char)(hoff >> (8 * i)); fwrite(len8, 1, 8, f); fwrite(header, 1, hoff, f); for (int t = 0; t < n_specs; t++) { @@ -133,9 +131,7 @@ static int tm_write_safetensors(const char *path, const char *dtype, /* Write config.json + model.safetensors with the given dimensions into dir. * dtype: "F32" or "BF16". Returns 0 on success. */ -static int tm_write_model_dims(const char *dir, const char *dtype, - const tm_dims_t *d) -{ +static int tm_write_model_dims(const char *dir, const char *dtype, const tm_dims_t *d) { int q = d->heads * d->head_dim, kv = d->kv_heads * d->head_dim; const tm_spec_t specs[TM_N_SPECS] = { {"embed_tokens.weight", d->vocab, d->hidden}, @@ -154,37 +150,37 @@ static int tm_write_model_dims(const char *dir, const char *dtype, }; char path[2048]; - if (tm_write_config(dir, d, NULL, -1) != 0) return -1; + if (tm_write_config(dir, d, NULL, -1) != 0) + return -1; snprintf(path, sizeof(path), "%s/model.safetensors", dir); return tm_write_safetensors(path, dtype, specs, TM_N_SPECS); } -static inline int tm_write_qwen3_model_dims(const char *dir, const char *dtype, +static inline int tm_write_qwen3_model_dims(const char *dir, + const char *dtype, const tm_dims_t *d, - int eos_token_id) -{ - if (tm_write_model_dims(dir, dtype, d) != 0) return -1; + int eos_token_id) { + if (tm_write_model_dims(dir, dtype, d) != 0) + return -1; return tm_write_config(dir, d, "qwen3", eos_token_id); } /* Add the 1_Dense per-token projection head that turns a base model dir into * a late-interaction snapshot. weight_name is "linear.weight" for a valid * head; pass another name to synthesize a malformed head. */ -static inline int tm_write_late_projection_named(const char *dir, - const char *dtype, - int token_dim, int hidden, - const char *weight_name) -{ +static inline int tm_write_late_projection_named( + const char *dir, const char *dtype, int token_dim, int hidden, const char *weight_name) { char path[2048]; snprintf(path, sizeof(path), "%s/1_Dense", dir); - if (mkdir(path, 0755) != 0 && errno != EEXIST) return -1; + if (mkdir(path, 0755) != 0 && errno != EEXIST) + return -1; snprintf(path, sizeof(path), "%s/1_Dense/config.json", dir); FILE *f = fopen(path, "w"); - if (!f) return -1; - fprintf(f, "{\"in_features\":%d,\"out_features\":%d,\"bias\":false}", - hidden, token_dim); + if (!f) + return -1; + fprintf(f, "{\"in_features\":%d,\"out_features\":%d,\"bias\":false}", hidden, token_dim); fclose(f); const tm_spec_t spec = {weight_name, token_dim, hidden}; @@ -192,16 +188,13 @@ static inline int tm_write_late_projection_named(const char *dir, return tm_write_safetensors(path, dtype, &spec, 1); } -static inline int tm_write_late_projection(const char *dir, const char *dtype, - int token_dim, int hidden) -{ - return tm_write_late_projection_named(dir, dtype, token_dim, hidden, - "linear.weight"); +static inline int +tm_write_late_projection(const char *dir, const char *dtype, int token_dim, int hidden) { + return tm_write_late_projection_named(dir, dtype, token_dim, hidden, "linear.weight"); } /* The historical default fixture: 4-dim hidden, 16-token vocab. */ -static inline int tm_write_model(const char *dir, const char *dtype) -{ +static inline int tm_write_model(const char *dir, const char *dtype) { static const tm_dims_t d = {4, 2, 1, 2, 8, 16}; return tm_write_model_dims(dir, dtype, &d); } diff --git a/tests/tok_fixture.h b/tests/tok_fixture.h index 7890927..24e1563 100644 --- a/tests/tok_fixture.h +++ b/tests/tok_fixture.h @@ -10,15 +10,13 @@ #include <string.h> /* GPT-2 byte -> unicode codepoint mapping (mirrors the tokenizer's rules). */ -static int tf_byte_to_cp(int b) -{ +static int tf_byte_to_cp(int b) { static int table[256]; static int init = 0; if (!init) { int n = 0; for (int i = 0; i < 256; i++) { - int normal = (i >= 33 && i <= 126) || (i >= 161 && i <= 172) || - (i >= 174 && i <= 255); + int normal = (i >= 33 && i <= 126) || (i >= 161 && i <= 172) || (i >= 174 && i <= 255); table[i] = normal ? i : 256 + n++; } init = 1; @@ -26,9 +24,12 @@ static int tf_byte_to_cp(int b) return table[b]; } -static int tf_cp_to_utf8(int cp, char *out) -{ - if (cp < 0x80) { out[0] = (char)cp; out[1] = '\0'; return 1; } +static int tf_cp_to_utf8(int cp, char *out) { + if (cp < 0x80) { + out[0] = (char)cp; + out[1] = '\0'; + return 1; + } out[0] = (char)(0xC0 | (cp >> 6)); out[1] = (char)(0x80 | (cp & 0x3F)); out[2] = '\0'; @@ -36,11 +37,11 @@ static int tf_cp_to_utf8(int cp, char *out) } /* JSON-escape a byte-unicode token (the byte alphabet needs only \" and \\). */ -static const char *tf_json_escape(const char *s, char *buf, size_t cap) -{ +static const char *tf_json_escape(const char *s, char *buf, size_t cap) { size_t o = 0; for (; *s && o + 2 < cap; s++) { - if (*s == '"' || *s == '\\') buf[o++] = '\\'; + if (*s == '"' || *s == '\\') + buf[o++] = '\\'; buf[o++] = *s; } buf[o] = '\0'; @@ -52,8 +53,7 @@ static const char *TF_MERGED[] = { "he", "ll", "hell", "hello", "\xC4\xA0w", "or", "ld", "\xC4\xA0wor", }; static const char *TF_MERGES[] = { - "h e", "l l", "he ll", "hell o", "\xC4\xA0 w", "o r", "l d", - "\xC4\xA0w or", + "h e", "l l", "he ll", "hell o", "\xC4\xA0 w", "o r", "l d", "\xC4\xA0w or", }; enum { TF_N_MERGED = sizeof(TF_MERGED) / sizeof(TF_MERGED[0]) }; /* Special tokens go last so contextual and late-interaction ids resolve @@ -65,17 +65,16 @@ enum { TF_DOCUMENT_ID = TF_QUERY_ID + 1 }; enum { TF_VOCAB_SIZE = TF_DOCUMENT_ID + 1 }; /* Write vocab.json + merges.txt into dir. Returns 0 on success. */ -static int tf_write_vocab(const char *dir) -{ +static int tf_write_vocab(const char *dir) { char path[2048], tok[4], esc[16]; snprintf(path, sizeof(path), "%s/vocab.json", dir); FILE *f = fopen(path, "w"); - if (!f) return -1; + if (!f) + return -1; fputs("{", f); for (int b = 0; b < 256; b++) { tf_cp_to_utf8(tf_byte_to_cp(b), tok); - fprintf(f, "%s\"%s\": %d", b ? ", " : "", - tf_json_escape(tok, esc, sizeof(esc)), b); + fprintf(f, "%s\"%s\": %d", b ? ", " : "", tf_json_escape(tok, esc, sizeof(esc)), b); } for (int i = 0; i < TF_N_MERGED; i++) fprintf(f, ", \"%s\": %d", TF_MERGED[i], 256 + i); @@ -88,7 +87,8 @@ static int tf_write_vocab(const char *dir) snprintf(path, sizeof(path), "%s/merges.txt", dir); f = fopen(path, "w"); - if (!f) return -1; + if (!f) + return -1; for (size_t i = 0; i < sizeof(TF_MERGES) / sizeof(TF_MERGES[0]); i++) fprintf(f, "%s\n", TF_MERGES[i]); fclose(f); |
