diff options
| author | Onur Gündüz <ogunduz@gmail.com> | 2026-09-29 14:38:02 +0200 |
|---|---|---|
| committer | Onur Gündüz <ogunduz@gmail.com> | 2026-09-29 14:38:02 +0200 |
| commit | 91c6aa07597462705a7f30ee5b97e2c7943b5616 (patch) | |
| tree | b16f686279a1f1b0691a6c0d65a544084fcc1c1e /liblangdet/model.c | |
| parent | 6f98fc8cf3f7cc4b9d523d16492a07d3e2aea8c9 (diff) | |
| download | langdet-91c6aa07597462705a7f30ee5b97e2c7943b5616.tar.gz langdet-91c6aa07597462705a7f30ee5b97e2c7943b5616.zip | |
Remove unused checkpoint, tokenizer, and attention compatibility paths
Diffstat (limited to 'liblangdet/model.c')
| -rw-r--r-- | liblangdet/model.c | 91 |
1 files changed, 47 insertions, 44 deletions
diff --git a/liblangdet/model.c b/liblangdet/model.c index 146bbe7..53f3692 100644 --- a/liblangdet/model.c +++ b/liblangdet/model.c @@ -46,10 +46,9 @@ int tensor_has_supported_shape(const safetensors_file_t *sf, return 1; } -static const float *load_weight(multi_safetensors_t *ms, const char *name, +static const float *load_weight(const safetensors_file_t *sf, const char *name, const int64_t *shape, int ndim) { - safetensors_file_t *sf = NULL; - const safetensor_t *t = multi_safetensors_find(ms, name, &sf); + const safetensor_t *t = safetensors_find(sf, name); if (!t) { fprintf(stderr, "langdet: tensor not found: %s\n", name); return NULL; @@ -59,20 +58,20 @@ static const float *load_weight(multi_safetensors_t *ms, const char *name, return safetensors_data(sf, t); } -static int load_bert_weights(langdet_model_t *model, multi_safetensors_t *ms) { +static int load_xlm_roberta_weights(langdet_model_t *model, const safetensors_file_t *sf) { const langdet_config_t *cfg = &model->config; langdet_weights_t *w = &model->weights; char name[256]; int hidden = cfg->hidden_size, inter = cfg->intermediate_size; - w->embed_tokens = load_weight(ms, "roberta.embeddings.word_embeddings.weight", + w->embed_tokens = load_weight(sf, "roberta.embeddings.word_embeddings.weight", (int64_t[]){cfg->vocab_size, hidden}, 2); - w->position_embeddings = load_weight(ms, "roberta.embeddings.position_embeddings.weight", + w->position_embeddings = load_weight(sf, "roberta.embeddings.position_embeddings.weight", (int64_t[]){cfg->max_position_embeddings, hidden}, 2); - w->token_type_embeddings = load_weight(ms, "roberta.embeddings.token_type_embeddings.weight", + w->token_type_embeddings = load_weight(sf, "roberta.embeddings.token_type_embeddings.weight", (int64_t[]){cfg->type_vocab_size, hidden}, 2); - w->langdet_ln_w = load_weight(ms, "roberta.embeddings.LayerNorm.weight", (int64_t[]){hidden}, 1); - w->langdet_ln_b = load_weight(ms, "roberta.embeddings.LayerNorm.bias", (int64_t[]){hidden}, 1); + w->langdet_ln_w = load_weight(sf, "roberta.embeddings.LayerNorm.weight", (int64_t[]){hidden}, 1); + w->langdet_ln_b = load_weight(sf, "roberta.embeddings.LayerNorm.bias", (int64_t[]){hidden}, 1); if (!w->embed_tokens || !w->position_embeddings || !w->token_type_embeddings || !w->langdet_ln_w || !w->langdet_ln_b) return -1; @@ -84,42 +83,42 @@ static int load_bert_weights(langdet_model_t *model, multi_safetensors_t *ms) { for (int i = 0; i < cfg->n_layers; i++) { langdet_layer_t *l = &w->layers[i]; -#define BERT_W(field, suffix, d0, d1) \ - do { \ - const int64_t e[2] = {(d0), (d1)}; \ - snprintf(name, sizeof(name), "roberta.encoder.layer.%d." suffix, i); \ - l->field = load_weight(ms, name, e, 2); \ - if (!l->field) \ - return -1; \ +#define LOAD_MATRIX(field, suffix, d0, d1) \ + do { \ + const int64_t e[2] = {(d0), (d1)}; \ + snprintf(name, sizeof(name), "roberta.encoder.layer.%d." suffix, i); \ + l->field = load_weight(sf, name, e, 2); \ + if (!l->field) \ + return -1; \ } while (0) -#define BERT_V(field, suffix, d0) \ - do { \ - const int64_t e[1] = {(d0)}; \ - snprintf(name, sizeof(name), "roberta.encoder.layer.%d." suffix, i); \ - l->field = load_weight(ms, name, e, 1); \ - if (!l->field) \ - return -1; \ +#define LOAD_VECTOR(field, suffix, d0) \ + do { \ + const int64_t e[1] = {(d0)}; \ + snprintf(name, sizeof(name), "roberta.encoder.layer.%d." suffix, i); \ + l->field = load_weight(sf, name, e, 1); \ + if (!l->field) \ + return -1; \ } while (0) - BERT_W(wq, "attention.self.query.weight", hidden, hidden); - BERT_V(q_bias, "attention.self.query.bias", hidden); - BERT_W(wk, "attention.self.key.weight", hidden, hidden); - BERT_V(k_bias, "attention.self.key.bias", hidden); - BERT_W(wv, "attention.self.value.weight", hidden, hidden); - BERT_V(v_bias, "attention.self.value.bias", hidden); - BERT_W(wo, "attention.output.dense.weight", hidden, hidden); - BERT_V(o_bias, "attention.output.dense.bias", hidden); - BERT_V(input_norm, "attention.output.LayerNorm.weight", hidden); - BERT_V(attn_ln_bias, "attention.output.LayerNorm.bias", hidden); - BERT_W(up_proj, "intermediate.dense.weight", inter, hidden); - BERT_V(ffn_inter_bias, "intermediate.dense.bias", inter); - BERT_W(down_proj, "output.dense.weight", hidden, inter); - BERT_V(ffn_out_bias, "output.dense.bias", hidden); - BERT_V(post_attn_norm, "output.LayerNorm.weight", hidden); - BERT_V(ffn_ln_bias, "output.LayerNorm.bias", hidden); + LOAD_MATRIX(wq, "attention.self.query.weight", hidden, hidden); + LOAD_VECTOR(q_bias, "attention.self.query.bias", hidden); + LOAD_MATRIX(wk, "attention.self.key.weight", hidden, hidden); + LOAD_VECTOR(k_bias, "attention.self.key.bias", hidden); + LOAD_MATRIX(wv, "attention.self.value.weight", hidden, hidden); + LOAD_VECTOR(v_bias, "attention.self.value.bias", hidden); + LOAD_MATRIX(wo, "attention.output.dense.weight", hidden, hidden); + LOAD_VECTOR(o_bias, "attention.output.dense.bias", hidden); + LOAD_VECTOR(input_norm, "attention.output.LayerNorm.weight", hidden); + LOAD_VECTOR(attn_ln_bias, "attention.output.LayerNorm.bias", hidden); + LOAD_MATRIX(up_proj, "intermediate.dense.weight", inter, hidden); + LOAD_VECTOR(ffn_inter_bias, "intermediate.dense.bias", inter); + LOAD_MATRIX(down_proj, "output.dense.weight", hidden, inter); + LOAD_VECTOR(ffn_out_bias, "output.dense.bias", hidden); + LOAD_VECTOR(post_attn_norm, "output.LayerNorm.weight", hidden); + LOAD_VECTOR(ffn_ln_bias, "output.LayerNorm.bias", hidden); -#undef BERT_W -#undef BERT_V +#undef LOAD_MATRIX +#undef LOAD_VECTOR } return 0; } @@ -132,12 +131,16 @@ langdet_model_t *langdet_model_load(const char *model_dir) { return NULL; if (langdet_config_parse(&model->config, model_dir) != 0) goto fail; - model->safetensors = multi_safetensors_open(model_dir); + char path[1024]; + int length = snprintf(path, sizeof(path), "%s/model.safetensors", model_dir); + if (length < 0 || (size_t)length >= sizeof(path)) + goto fail; + model->safetensors = safetensors_open(path); if (!model->safetensors) { fprintf(stderr, "langdet_model_load: failed to open safetensors in %s\n", model_dir); goto fail; } - if (load_bert_weights(model, model->safetensors) != 0) + if (load_xlm_roberta_weights(model, model->safetensors) != 0) goto fail; return model; @@ -149,7 +152,7 @@ fail: void langdet_model_free(langdet_model_t *model) { if (!model) return; - multi_safetensors_close(model->safetensors); + safetensors_close(model->safetensors); free(model->weights.layers); free(model); } |
