diff options
| author | Onur Gündüz <ogunduz@gmail.com> | 2026-06-15 12:16:33 +0200 |
|---|---|---|
| committer | Onur Gündüz <ogunduz@gmail.com> | 2026-06-15 12:16:33 +0200 |
| commit | c6ade0a14b87053d4a19cbac121d03af48cbf515 (patch) | |
| tree | fccea2d9351e310ad308a2336abf766b4884cc11 /tests | |
| parent | b5ada41fedf1ebdaff39ee0fbef76a468b26e5f7 (diff) | |
| download | ffwd-c6ade0a14b87053d4a19cbac121d03af48cbf515.tar.gz ffwd-c6ade0a14b87053d4a19cbac121d03af48cbf515.zip | |
all: support GTE-Qwen2 embeddings
Diffstat (limited to 'tests')
| -rw-r--r-- | tests/test_qwen2.c | 97 | ||||
| -rw-r--r-- | tests/test_server.c | 87 | ||||
| -rw-r--r-- | tests/tiny_model.h | 85 |
3 files changed, 255 insertions, 14 deletions
diff --git a/tests/test_qwen2.c b/tests/test_qwen2.c new file mode 100644 index 0000000..4795f7f --- /dev/null +++ b/tests/test_qwen2.c @@ -0,0 +1,97 @@ +/* tests/test_qwen2.c - hermetic GTE-Qwen2 block and task semantics. */ + +#include "embed.h" +#include "tiny_model.h" + +#include <math.h> +#include <stdio.h> +#include <stdlib.h> +#include <string.h> +#include <unistd.h> + +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; + } + return out; +} + +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); + 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 || + tm_write_qwen2_model_dims(zero_dir, "F32", &dims, 15, 1) != 0) { + fprintf(stderr, "fixture creation failed\n"); + 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); + 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 || + !cfg->append_terminal_token) { + fprintf(stderr, "Qwen2 config or Sentence Transformers metadata was not detected\n"); + return 1; + } + + 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) { + 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 || + max_abs_diff(pooled, expected, 4) > 1e-5f) { + fprintf(stderr, "last-token pooling mismatch\n"); + return 1; + } + if (max_abs_diff(embedding, zero_embedding, 4) < 1e-4f) { + fprintf(stderr, "QKV projection bias did not affect the embedding\n"); + return 1; + } + + int changed_ids[] = {1, 3, 15}; + float changed_states[3 * 4]; + if (embed_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}}; + 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 || + 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); + puts("ok: Qwen2 bias, bidirectional attention, and GTE pooling semantics"); + return 0; +} diff --git a/tests/test_server.c b/tests/test_server.c index ce9668e..9ebb8f2 100644 --- a/tests/test_server.c +++ b/tests/test_server.c @@ -146,41 +146,54 @@ static void test_encoding_from_root_family(void) { static void test_text_type(void) { cJSON *detail = cJSON_CreateArray(); - /* Absent: no-op for both families, no error. */ + /* Absent: no-op for both model types, no error. */ cJSON *empty = cJSON_CreateObject(); - TEST_ASSERT(text_type_is_query(empty, detail, EMBED_API_OPENAI) == 0); - TEST_ASSERT(text_type_is_query(empty, detail, EMBED_API_PERPLEXITY) == 0); + TEST_ASSERT(text_type_is_query(empty, detail, EMBED_QWEN3_QUERY_INSTRUCT) == 0); + TEST_ASSERT(text_type_is_query(empty, detail, NULL) == 0); cJSON_Delete(empty); TEST_ASSERT(cJSON_GetArraySize(detail) == 0); /* Qwen3: query -> 1, document -> 0, both accepted. */ cJSON *q = cJSON_CreateObject(); cJSON_AddStringToObject(q, "text_type", "query"); - TEST_ASSERT(text_type_is_query(q, detail, EMBED_API_OPENAI) == 1); + TEST_ASSERT(text_type_is_query(q, detail, EMBED_QWEN3_QUERY_INSTRUCT) == 1); cJSON_Delete(q); cJSON *d = cJSON_CreateObject(); cJSON_AddStringToObject(d, "text_type", "document"); - TEST_ASSERT(text_type_is_query(d, detail, EMBED_API_OPENAI) == 0); + TEST_ASSERT(text_type_is_query(d, detail, EMBED_QWEN3_QUERY_INSTRUCT) == 0); cJSON_Delete(d); TEST_ASSERT(cJSON_GetArraySize(detail) == 0); /* Invalid enum value on Qwen3 is reported. */ cJSON *bad = cJSON_CreateObject(); cJSON_AddStringToObject(bad, "text_type", "passage"); - TEST_ASSERT(text_type_is_query(bad, detail, EMBED_API_OPENAI) == 0); + TEST_ASSERT(text_type_is_query(bad, detail, EMBED_QWEN3_QUERY_INSTRUCT) == 0); TEST_ASSERT(cJSON_GetArraySize(detail) == 1); cJSON_Delete(bad); /* text_type on a Perplexity-family model is rejected. */ cJSON *p = cJSON_CreateObject(); cJSON_AddStringToObject(p, "text_type", "query"); - TEST_ASSERT(text_type_is_query(p, detail, EMBED_API_PERPLEXITY) == 0); + TEST_ASSERT(text_type_is_query(p, detail, NULL) == 0); TEST_ASSERT(cJSON_GetArraySize(detail) == 2); cJSON_Delete(p); cJSON_Delete(detail); } +static void test_model_registry(void) { + model_slot slot = model_slot_for_id("gte-Qwen2-1.5B-instruct"); + TEST_ASSERT(slot == MODEL_GTE_QWEN2_15); + if (slot == MODEL_UNKNOWN) + return; + TEST_ASSERT(k_models[slot].dim == 1536); + TEST_ASSERT(k_models[slot].min_dim == 1536); + TEST_ASSERT(k_models[slot].attention_mode == EMBED_ATTENTION_BIDIRECTIONAL); + TEST_ASSERT(k_models[slot].pooling_mode == EMBED_POOL_LAST_TOKEN); + TEST_ASSERT(k_models[slot].normalize_embeddings == 1); + TEST_ASSERT(strcmp(k_models[slot].query_instruct, EMBED_GTE_QWEN2_QUERY_INSTRUCT) == 0); +} + static void test_collect_job_batch_deadline(void) { http_server s; memset(&s, 0, sizeof(s)); @@ -410,7 +423,7 @@ static int http_req(int port, } typedef struct { - embed_server_model_spec_t spec[4]; + embed_server_model_spec_t spec[5]; embed_server_config_t cfg; int rc; } srv_ctx; @@ -657,6 +670,45 @@ static void test_http_embeddings(int port) { cJSON_Delete(tt_query); cJSON_Delete(tt_manual); + /* GTE-Qwen2 uses the same retrieval task text but publishes a space after + * "Query:". Verify the registry-specific prompt through the full path. */ + TEST_ASSERT(http_req(port, "POST", "/v1/embeddings", TEST_API_KEY, + "{\"model\":\"gte-Qwen2-1.5B-instruct\",\"input\":\"hello\"," + "\"text_type\":\"query\",\"encoding_format\":\"float\"}", + NULL, &body) == 200); + cJSON *gte_query = body ? parse_embeddings_body(body, 1) : NULL; + free(body); + body = NULL; + TEST_ASSERT(http_req(port, "POST", "/v1/embeddings", TEST_API_KEY, + "{\"model\":\"gte-Qwen2-1.5B-instruct\",\"input\":\"Instruct: Given a " + "web search query, retrieve relevant passages that answer the " + "query\\nQuery: hello\",\"text_type\":\"document\"," + "\"encoding_format\":\"float\"}", + NULL, &body) == 200); + cJSON *gte_manual = body ? parse_embeddings_body(body, 1) : NULL; + free(body); + body = NULL; + if (gte_query && gte_manual) { + cJSON *qe = cJSON_GetObjectItemCaseSensitive( + cJSON_GetArrayItem(cJSON_GetObjectItemCaseSensitive(gte_query, "data"), 0), + "embedding"); + cJSON *me = cJSON_GetObjectItemCaseSensitive( + cJSON_GetArrayItem(cJSON_GetObjectItemCaseSensitive(gte_manual, "data"), 0), + "embedding"); + TEST_ASSERT(cJSON_IsArray(qe) && cJSON_IsArray(me)); + TEST_ASSERT(cJSON_GetArraySize(qe) == 1536 && cJSON_GetArraySize(me) == 1536); + double gte_diff = 0.0; + for (int i = 0; i < cJSON_GetArraySize(qe); i++) { + double d = fabs(cJSON_GetArrayItem(qe, i)->valuedouble - + cJSON_GetArrayItem(me, i)->valuedouble); + if (d > gte_diff) + gte_diff = d; + } + TEST_ASSERT(gte_diff < 1e-6); + } + cJSON_Delete(gte_query); + cJSON_Delete(gte_manual); + /* text_type is rejected on a Perplexity-family model, and bad values 422. */ TEST_ASSERT(http_req(port, "POST", "/v1/embeddings", TEST_API_KEY, "{\"model\":\"pplx-embed-v1-0.6b\",\"input\":\"hello\"," @@ -897,7 +949,7 @@ static void test_http_rerank(int port) { } static void test_http_server(void) { - char dir[1024], late_dir[1024], qwen_dir[1024]; + char dir[1024], late_dir[1024], qwen_dir[1024], gte_dir[1024]; snprintf(dir, sizeof(dir), "%s/embed-srv-test-XXXXXX", getenv("TMPDIR") ? getenv("TMPDIR") : "/tmp"); if (!mkdtemp(dir)) { @@ -919,14 +971,24 @@ static void test_http_server(void) { test_failures++; return; } + snprintf(gte_dir, sizeof(gte_dir), "%s/embed-srv-gte-test-XXXXXX", + getenv("TMPDIR") ? getenv("TMPDIR") : "/tmp"); + if (!mkdtemp(gte_dir)) { + fprintf(stderr, "FAIL: GTE mkdtemp\n"); + test_failures++; + return; + } /* 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}; + tm_dims_t gte_dims = {1536, 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 || 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 || + tf_write_vocab(gte_dir) != 0 || + tm_write_qwen2_model_dims(gte_dir, "F32", >e_dims, TF_EOT_ID, 0) != 0) { fprintf(stderr, "FAIL: fixture write\n"); test_failures++; return; @@ -947,8 +1009,10 @@ static void test_http_server(void) { ctx.spec[2].path = late_dir; ctx.spec[3].id = "Qwen3-Embedding-0.6B"; ctx.spec[3].path = qwen_dir; + ctx.spec[4].id = "gte-Qwen2-1.5B-instruct"; + ctx.spec[4].path = gte_dir; ctx.cfg.models = ctx.spec; - ctx.cfg.n_models = 4; + ctx.cfg.n_models = 5; ctx.cfg.port = port; ctx.cfg.batch_size = 2; ctx.cfg.batch_wait_us = 1000; @@ -992,6 +1056,7 @@ int main(void) { test_append_embedding_value_float(); test_encoding_from_root_family(); test_text_type(); + test_model_registry(); test_collect_job_batch_deadline(); test_collect_job_batch_zero_wait(); test_http_server(); diff --git a/tests/tiny_model.h b/tests/tiny_model.h index d2534d2..95496b7 100644 --- a/tests/tiny_model.h +++ b/tests/tiny_model.h @@ -76,8 +76,12 @@ tm_write_config(const char *dir, const tm_dims_t *d, const char *model_type, int /* Write a safetensors file holding `specs` with tm_value() contents. * dtype: "F32" or "BF16". Returns 0 on success. */ -static int tm_write_safetensors_with_prefix( - const char *path, const char *dtype, const tm_spec_t *specs, int n_specs, const char *prefix) { +static int tm_write_safetensors_with_prefix_options(const char *path, + const char *dtype, + const tm_spec_t *specs, + int n_specs, + const char *prefix, + int zero_bias) { int bf16 = strcmp(dtype, "BF16") == 0; size_t esize = bf16 ? 2 : 4; @@ -117,7 +121,7 @@ static int tm_write_safetensors_with_prefix( const tm_spec_t *s = &specs[t]; size_t n = (size_t)s->rows * (s->cols ? (size_t)s->cols : 1); for (size_t i = 0; i < n; i++) { - float v = tm_value(s->name, i); + float v = zero_bias && strstr(s->name, ".bias") ? 0.0f : tm_value(s->name, i); if (bf16) { uint16_t h = tm_f32_to_bf16(v); fwrite(&h, sizeof(h), 1, f); @@ -130,6 +134,11 @@ static int tm_write_safetensors_with_prefix( return 0; } +static int tm_write_safetensors_with_prefix( + const char *path, const char *dtype, const tm_spec_t *specs, int n_specs, const char *prefix) { + return tm_write_safetensors_with_prefix_options(path, dtype, specs, n_specs, prefix, 0); +} + static int tm_write_safetensors(const char *path, const char *dtype, const tm_spec_t *specs, int n_specs) { return tm_write_safetensors_with_prefix(path, dtype, specs, n_specs, ""); @@ -184,6 +193,76 @@ static inline int tm_write_qwen3_model_dims(const char *dir, return tm_write_config(dir, d, "qwen3", eos_token_id); } +static inline int tm_write_qwen2_model_dims( + const char *dir, const char *dtype, const tm_dims_t *d, int eos_token_id, int zero_bias) { + int q = d->heads * d->head_dim, kv = d->kv_heads * d->head_dim; + const tm_spec_t specs[] = { + {"embed_tokens.weight", d->vocab, d->hidden}, + {"layers.0.self_attn.q_proj.weight", q, d->hidden}, + {"layers.0.self_attn.q_proj.bias", q, 0}, + {"layers.0.self_attn.k_proj.weight", kv, d->hidden}, + {"layers.0.self_attn.k_proj.bias", kv, 0}, + {"layers.0.self_attn.v_proj.weight", kv, d->hidden}, + {"layers.0.self_attn.v_proj.bias", kv, 0}, + {"layers.0.self_attn.o_proj.weight", d->hidden, q}, + {"layers.0.input_layernorm.weight", d->hidden, 0}, + {"layers.0.post_attention_layernorm.weight", d->hidden, 0}, + {"layers.0.mlp.gate_proj.weight", d->intermediate, d->hidden}, + {"layers.0.mlp.up_proj.weight", d->intermediate, d->hidden}, + {"layers.0.mlp.down_proj.weight", d->hidden, d->intermediate}, + {"norm.weight", d->hidden, 0}, + }; + 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," + "\"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," + "\"model_type\":\"qwen2\",\"eos_token_id\":%d,\"is_causal\":false}", + d->hidden, d->heads, d->kv_heads, d->head_dim, d->intermediate, d->vocab, eos_token_id); + fclose(f); + + snprintf(path, sizeof(path), "%s/model.safetensors", dir); + if (tm_write_safetensors_with_prefix_options( + path, dtype, specs, (int)(sizeof(specs) / sizeof(specs[0])), "model.", zero_bias) != 0) + return -1; + + snprintf(path, sizeof(path), "%s/1_Pooling", dir); + if (mkdir(path, 0755) != 0 && errno != EEXIST) + return -1; + snprintf(path, sizeof(path), "%s/1_Pooling/config.json", dir); + f = fopen(path, "w"); + if (!f) + return -1; + fputs("{\"pooling_mode_cls_token\":false,\"pooling_mode_mean_tokens\":false," + "\"pooling_mode_max_tokens\":false,\"pooling_mode_mean_sqrt_len_tokens\":false," + "\"pooling_mode_weightedmean_tokens\":false,\"pooling_mode_lasttoken\":true}", + f); + fclose(f); + + snprintf(path, sizeof(path), "%s/modules.json", dir); + f = fopen(path, "w"); + if (!f) + return -1; + fputs("[{\"type\":\"sentence_transformers.models.Transformer\"}," + "{\"type\":\"sentence_transformers.models.Pooling\"}," + "{\"type\":\"sentence_transformers.models.Normalize\"}]", + f); + fclose(f); + + snprintf(path, sizeof(path), "%s/tokenizer_config.json", dir); + f = fopen(path, "w"); + if (!f) + return -1; + fputs("{\"add_eos_token\":true}", f); + fclose(f); + return 0; +} + /* 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. */ |
