diff options
| author | Onur Gündüz <ogunduz@gmail.com> | 2026-06-20 20:36:55 +0200 |
|---|---|---|
| committer | Onur Gündüz <ogunduz@gmail.com> | 2026-06-21 06:00:26 +0200 |
| commit | 6652fcbc653d84e75cb5c08c5f5edf52263ed50f (patch) | |
| tree | 1ccffa48e1dc86c59c61dcd4672ffedcf1e0c873 /tests/test_qwen2.c | |
| parent | dbdf714bb57578431924e5d069089bdfd9f4c604 (diff) | |
| download | ffwd-6652fcbc653d84e75cb5c08c5f5edf52263ed50f.tar.gz ffwd-6652fcbc653d84e75cb5c08c5f5edf52263ed50f.zip | |
tests: upgrade unit tests and add integration/parity tests
Diffstat (limited to 'tests/test_qwen2.c')
| -rw-r--r-- | tests/test_qwen2.c | 46 |
1 files changed, 23 insertions, 23 deletions
diff --git a/tests/test_qwen2.c b/tests/test_qwen2.c index 4795f7f..9567aad 100644 --- a/tests/test_qwen2.c +++ b/tests/test_qwen2.c @@ -1,6 +1,6 @@ /* tests/test_qwen2.c - hermetic GTE-Qwen2 block and task semantics. */ -#include "embed.h" +#include "internal.h" #include "tiny_model.h" #include <math.h> @@ -22,8 +22,8 @@ static float max_abs_diff(const float *a, const float *b, int n) { int main(void) { const char *tmp = getenv("TMPDIR") ? getenv("TMPDIR") : "/tmp"; char dir[1024], zero_dir[1024]; - snprintf(dir, sizeof(dir), "%s/embed-qwen2-test-XXXXXX", tmp); - snprintf(zero_dir, sizeof(zero_dir), "%s/embed-qwen2-zero-bias-test-XXXXXX", tmp); + snprintf(dir, sizeof(dir), "%s/ffwd-qwen2-test-XXXXXX", tmp); + snprintf(zero_dir, sizeof(zero_dir), "%s/ffwd-qwen2-zero-bias-test-XXXXXX", tmp); tm_dims_t dims = {4, 2, 1, 2, 8, 16}; if (!mkdtemp(dir) || !mkdtemp(zero_dir) || tm_write_qwen2_model_dims(dir, "F32", &dims, 15, 0) != 0 || @@ -32,17 +32,17 @@ int main(void) { return 2; } - embed_model_t *model = embed_model_load(dir); - embed_model_t *zero_model = embed_model_load(zero_dir); - embed_workspace_t *ws = embed_workspace_new(model); - embed_workspace_t *zero_ws = embed_workspace_new(zero_model); - const embed_config_t *cfg = embed_model_config(model); + ffwd_model_t *model = ffwd_model_load(dir); + ffwd_model_t *zero_model = ffwd_model_load(zero_dir); + ffwd_workspace_t *ws = ffwd_workspace_new(model); + ffwd_workspace_t *zero_ws = ffwd_workspace_new(zero_model); + const ffwd_config_t *cfg = ffwd_model_config(model); if (!model || !zero_model || !ws || !zero_ws || !cfg) { fprintf(stderr, "model setup failed\n"); return 1; } - if (cfg->qk_norm || !cfg->qkv_bias || cfg->attention_mode != EMBED_ATTENTION_BIDIRECTIONAL || - cfg->pooling_mode != EMBED_POOL_LAST_TOKEN || !cfg->normalize_embeddings || + if (cfg->qk_norm || !cfg->qkv_bias || cfg->attention_mode != FFWD_ATTENTION_BIDIRECTIONAL || + cfg->pooling_mode != FFWD_POOL_LAST_TOKEN || !cfg->normalize_embeddings || !cfg->append_terminal_token) { fprintf(stderr, "Qwen2 config or Sentence Transformers metadata was not detected\n"); return 1; @@ -50,17 +50,17 @@ int main(void) { int ids[] = {1, 2, 15}; float states[3 * 4], embedding[4], pooled[4], zero_embedding[4]; - if (embed_model_forward_into(model, ws, ids, 3, states) != 0 || - embed_model_encode_into(model, ws, ids, 3, embedding) != 0 || - embed_model_encode_into(zero_model, zero_ws, ids, 3, zero_embedding) != 0 || - embed_pool_batch(cfg, states, (int[]){3}, 1, pooled) != 0) { + if (ffwd_model_forward_into(model, ws, ids, 3, states) != 0 || + ffwd_model_encode_into(model, ws, ids, 3, embedding) != 0 || + ffwd_model_encode_into(zero_model, zero_ws, ids, 3, zero_embedding) != 0 || + ffwd_pool_batch(cfg, states, (int[]){3}, 1, pooled) != 0) { fprintf(stderr, "Qwen2 inference failed\n"); return 1; } float expected[4]; memcpy(expected, states + 2 * 4, sizeof(expected)); - if (embed_l2_normalize(expected, 4) != 0 || max_abs_diff(embedding, expected, 4) > 1e-5f || + if (ffwd_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; @@ -72,26 +72,26 @@ int main(void) { int changed_ids[] = {1, 3, 15}; float changed_states[3 * 4]; - if (embed_model_forward_into(model, ws, changed_ids, 3, changed_states) != 0 || + if (ffwd_model_forward_into(model, ws, changed_ids, 3, changed_states) != 0 || max_abs_diff(states, changed_states, 4) < 1e-6f) { fprintf(stderr, "bidirectional attention did not affect the first token\n"); return 1; } int short_ids[] = {4, 15}; - embed_input_t inputs[] = {{ids, 3}, {short_ids, 2}}; + ffwd_input_t inputs[] = {{ids, 3}, {short_ids, 2}}; float batch[8], single[4]; - if (embed_model_encode_batch(model, ws, inputs, 2, batch) != 0 || - embed_model_encode_into(model, ws, short_ids, 2, single) != 0 || + if (ffwd_model_encode_batch(model, ws, inputs, 2, batch) != 0 || + ffwd_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) { fprintf(stderr, "Qwen2 packed batch parity failed\n"); return 1; } - embed_workspace_free(zero_ws); - embed_workspace_free(ws); - embed_model_free(zero_model); - embed_model_free(model); + ffwd_workspace_free(zero_ws); + ffwd_workspace_free(ws); + ffwd_model_free(zero_model); + ffwd_model_free(model); puts("ok: Qwen2 bias, bidirectional attention, and GTE pooling semantics"); return 0; } |
