diff options
| author | Onur Gündüz <ogunduz@gmail.com> | 2026-06-10 08:48:04 +0200 |
|---|---|---|
| committer | Onur Gündüz <ogunduz@gmail.com> | 2026-06-14 08:27:03 +0200 |
| commit | a960ebb8dec645ab619a68d7412d4392e99f4b7a (patch) | |
| tree | 06ef707b7ab7b7ff4064ec7acf94bd4a4cd92409 /tests | |
| parent | 809c26869b83530d91a5068e5085da7fccbf5399 (diff) | |
| download | ffwd-a960ebb8dec645ab619a68d7412d4392e99f4b7a.tar.gz ffwd-a960ebb8dec645ab619a68d7412d4392e99f4b7a.zip | |
tests,bench: move generated C checkers into standing files
Diffstat (limited to 'tests')
| -rw-r--r-- | tests/test_bf16_model.c | 70 | ||||
| -rw-r--r-- | tests/test_safetensors.c | 36 | ||||
| -rw-r--r-- | tests/test_workspace.c | 196 |
3 files changed, 302 insertions, 0 deletions
diff --git a/tests/test_bf16_model.c b/tests/test_bf16_model.c new file mode 100644 index 0000000..0f3101a --- /dev/null +++ b/tests/test_bf16_model.c @@ -0,0 +1,70 @@ +/* tests/test_bf16_model.c - F32 vs BF16 loader/dispatch parity on a tiny + * synthetic model. Runs via `make test`. */ + +#include "embed.h" + +#include <math.h> +#include <stdio.h> +#include <stdlib.h> + +int main(int argc, char **argv) { + if (argc != 3) { + fprintf(stderr, "usage: %s f32_model bf16_model\n", argv[0]); + return 2; + } + + int ids[] = {1, 2, 3, 4}; + pplx_model_t *mf32 = pplx_model_load(argv[1]); + pplx_model_t *mbf16 = pplx_model_load(argv[2]); + if (!mf32 || !mbf16) { + fprintf(stderr, "model load failed\n"); + return 1; + } + + const pplx_config_t *cfg = pplx_model_config(mf32); + int dim = cfg->hidden_size; + float *a = (float *)malloc((size_t)dim * sizeof(float)); + float *b = (float *)malloc((size_t)dim * sizeof(float)); + pplx_workspace_t *wf32 = pplx_workspace_new(mf32); + pplx_workspace_t *wbf16 = pplx_workspace_new(mbf16); + if (!a || !b || !wf32 || !wbf16) { + fprintf(stderr, "allocation failed\n"); + return 1; + } + + if (pplx_model_embed_into(mf32, wf32, ids, 4, a) != 0 || + pplx_model_embed_into(mbf16, wbf16, ids, 4, b) != 0) { + fprintf(stderr, "embedding failed\n"); + return 1; + } + + float max_diff = 0.0f; + float norm = 0.0f; + for (int i = 0; i < dim; i++) { + if (!isfinite(a[i]) || !isfinite(b[i])) { + fprintf(stderr, "non-finite embedding value\n"); + return 1; + } + float d = fabsf(a[i] - b[i]); + if (d > max_diff) max_diff = d; + norm += b[i] * b[i]; + } + norm = sqrtf(norm); + /* Embeddings are intentionally unnormalized; require a sane nonzero + * magnitude and tight F32-vs-BF16 parity, not unit norm. */ + if (norm <= 1e-6f || max_diff > 2e-5f) { + fprintf(stderr, "bad parity: norm=%.9g max_diff=%.9g\n", norm, max_diff); + return 1; + } + + printf("ok: bf16 model parity dim=%d max_abs_diff=%.9g norm=%.9g\n", + dim, max_diff, norm); + + pplx_workspace_free(wf32); + pplx_workspace_free(wbf16); + pplx_model_free(mf32); + pplx_model_free(mbf16); + free(a); + free(b); + return 0; +} diff --git a/tests/test_safetensors.c b/tests/test_safetensors.c new file mode 100644 index 0000000..42d8f21 --- /dev/null +++ b/tests/test_safetensors.c @@ -0,0 +1,36 @@ +/* tests/test_safetensors.c - safetensors parser limit and rejection tests. + * Runs via `make test`. Fixtures are synthesized in a temp directory. */ + +#include "qwen_safetensors.h" + +#include <stdio.h> +#include <string.h> + +int main(int argc, char **argv) { + if (argc != 4) { + fprintf(stderr, "usage: %s file|multi path ok|fail\n", argv[0]); + return 2; + } + + int want_ok = strcmp(argv[3], "ok") == 0; + int got_ok = 0; + if (strcmp(argv[1], "file") == 0) { + safetensors_file_t *sf = safetensors_open(argv[2]); + got_ok = sf != NULL; + safetensors_close(sf); + } else if (strcmp(argv[1], "multi") == 0) { + multi_safetensors_t *ms = multi_safetensors_open(argv[2]); + got_ok = ms != NULL; + multi_safetensors_close(ms); + } else { + return 2; + } + + if (got_ok != want_ok) { + fprintf(stderr, "%s %s: got %s want %s\n", + argv[1], argv[2], got_ok ? "ok" : "fail", + want_ok ? "ok" : "fail"); + return 1; + } + return 0; +} diff --git a/tests/test_workspace.c b/tests/test_workspace.c new file mode 100644 index 0000000..a87e9f8 --- /dev/null +++ b/tests/test_workspace.c @@ -0,0 +1,196 @@ +/* tests/test_workspace.c - model/workspace API lifecycle tests. + * Runs via `make test`. Uses a tiny synthetic model. */ + +#include "embed.h" +#include "qwen_tokenizer.h" + +#include <math.h> +#include <stdio.h> +#include <stdlib.h> +#include <string.h> + +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; + } + return m; +} + +int main(int argc, char **argv) +{ + if (argc != 2) { + fprintf(stderr, "usage: %s MODEL_DIR\n", argv[0]); + return 2; + } + + const char *model_dir = argv[1]; + char vocab_path[4096]; + snprintf(vocab_path, sizeof(vocab_path), "%s/vocab.json", model_dir); + + pplx_model_t *model = pplx_model_load(model_dir); + if (!model) { + fprintf(stderr, "failed to load model\n"); + return 1; + } + + const pplx_config_t *cfg = pplx_model_config(model); + if (!cfg || cfg->hidden_size <= 0) { + fprintf(stderr, "invalid model config\n"); + pplx_model_free(model); + return 1; + } + + pplx_workspace_t *batch_ws = pplx_workspace_new(model); + pplx_workspace_t *single_ws = pplx_workspace_new(model); + qwen_tokenizer_t *tok = qwen_tokenizer_load(vocab_path); + if (!batch_ws || !single_ws || !tok) { + fprintf(stderr, "failed to allocate workspace or tokenizer\n"); + qwen_tokenizer_free(tok); + pplx_workspace_free(single_ws); + pplx_workspace_free(batch_ws); + pplx_model_free(model); + return 1; + } + + size_t initial_ws_bytes = pplx_workspace_nbytes(batch_ws); + if (initial_ws_bytes == 0) { + fprintf(stderr, "workspace byte accounting returned zero\n"); + goto fail; + } + + 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])); + int *ids[4] = {0}; + int ntok[4] = {0}; + pplx_input_t inputs[4]; + + for (int i = 0; i < batch; i++) { + ids[i] = qwen_tokenizer_encode(tok, texts[i], &ntok[i]); + if (!ids[i] || ntok[i] <= 0) { + fprintf(stderr, "failed to tokenize input %d\n", i); + goto fail; + } + inputs[i].ids = ids[i]; + inputs[i].n_tokens = ntok[i]; + } + + int dim = cfg->hidden_size; + float *batched = (float *)calloc((size_t)batch * dim, sizeof(float)); + float *single = (float *)calloc((size_t)dim, sizeof(float)); + float *normalized = (float *)calloc((size_t)dim, sizeof(float)); + float *states = (float *)calloc((size_t)ntok[0] * dim, sizeof(float)); + if (!batched || !single || !normalized || !states) { + fprintf(stderr, "allocation failure\n"); + free(states); + free(normalized); + free(single); + free(batched); + goto fail; + } + + if (pplx_model_embed_batch(model, batch_ws, inputs, batch, batched) != 0) { + fprintf(stderr, "pplx_model_embed_batch failed\n"); + free(states); + free(normalized); + free(single); + free(batched); + goto fail; + } + + size_t batch_ws_bytes = pplx_workspace_nbytes(batch_ws); + if (batch_ws_bytes <= initial_ws_bytes) { + fprintf(stderr, "workspace byte accounting did not grow after embedding\n"); + free(states); + free(normalized); + free(single); + free(batched); + goto fail; + } + + float worst_diff = 0.0f; + float worst_cos = 1.0f; + for (int i = 0; i < batch; i++) { + if (pplx_model_embed_into(model, single_ws, ids[i], ntok[i], single) != 0) { + fprintf(stderr, "pplx_model_embed_into failed at %d\n", i); + free(states); + free(normalized); + free(single); + free(batched); + goto fail; + } + + const float *row = batched + (size_t)i * dim; + float diff = max_abs_diff(row, single, dim); + float cos = pplx_cosine_similarity(row, single, dim); + if (diff > worst_diff) worst_diff = diff; + if (cos < worst_cos) worst_cos = cos; + } + + memcpy(normalized, batched, (size_t)dim * sizeof(float)); + if (pplx_l2_normalize(normalized, dim) != 0) { + fprintf(stderr, "pplx_l2_normalize failed\n"); + free(states); + free(normalized); + free(single); + free(batched); + goto fail; + } + float normalized_norm_sq = 0.0f; + 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, "pplx_l2_normalize produced norm_sq=%g\n", + normalized_norm_sq); + free(states); + free(normalized); + free(single); + free(batched); + goto fail; + } + + if (pplx_model_forward_into(model, batch_ws, ids[0], ntok[0], states) != 0) { + fprintf(stderr, "pplx_model_forward_into failed\n"); + free(states); + free(normalized); + free(single); + free(batched); + goto fail; + } + + free(states); + free(normalized); + free(single); + free(batched); + + for (int i = 0; i < batch; i++) free(ids[i]); + qwen_tokenizer_free(tok); + pplx_workspace_free(single_ws); + pplx_workspace_free(batch_ws); + pplx_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); + 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); + return 0; + +fail: + for (int i = 0; i < batch; i++) free(ids[i]); + qwen_tokenizer_free(tok); + pplx_workspace_free(single_ws); + pplx_workspace_free(batch_ws); + pplx_model_free(model); + return 1; +} |
