aboutsummaryrefslogtreecommitdiff
path: root/liblangdet/config.c
diff options
context:
space:
mode:
authorOnur Gündüz <ogunduz@gmail.com>2026-09-29 13:11:00 +0200
committerOnur Gündüz <ogunduz@gmail.com>2026-09-29 13:11:00 +0200
commit37cbaef9ceb828c667da8f398027ba5415ee7475 (patch)
tree6103e92a9030766e78a35df00f7a1155d0d0ecf5 /liblangdet/config.c
parentcf7d8169ec12cb3adf90d4ff78b3257e74985302 (diff)
downloadlangdet-37cbaef9ceb828c667da8f398027ba5415ee7475.tar.gz
langdet-37cbaef9ceb828c667da8f398027ba5415ee7475.zip
Rename CPU library API and sources to langdet
Diffstat (limited to 'liblangdet/config.c')
-rw-r--r--liblangdet/config.c67
1 files changed, 67 insertions, 0 deletions
diff --git a/liblangdet/config.c b/liblangdet/config.c
new file mode 100644
index 0000000..f22a9d7
--- /dev/null
+++ b/liblangdet/config.c
@@ -0,0 +1,67 @@
+#include "config.h"
+
+#include <yyjson.h>
+
+#include <limits.h>
+#include <math.h>
+#include <stdio.h>
+
+static int json_get_int(yyjson_val *obj, const char *key, int fallback) {
+ yyjson_val *v = yyjson_obj_get(obj, key);
+ if (!v)
+ return fallback;
+ if (!yyjson_is_int(v))
+ return -1;
+ int64_t n = yyjson_get_sint(v);
+ return n >= 0 && n <= INT_MAX ? (int)n : -1;
+}
+
+int langdet_config_parse(langdet_config_t *cfg, const char *model_dir) {
+ char path[1024];
+ int n = snprintf(path, sizeof(path), "%s/config.json", model_dir);
+ if (n < 0 || (size_t)n >= sizeof(path)) {
+ fprintf(stderr, "langdet_config: model path too long\n");
+ return -1;
+ }
+
+ yyjson_read_err error;
+ yyjson_doc *doc = yyjson_read_file(path, 0, NULL, &error);
+ if (!doc) {
+ fprintf(stderr, "langdet_config: cannot read %s: %s\n", path, error.msg);
+ return -1;
+ }
+ yyjson_val *root = yyjson_doc_get_root(doc);
+ yyjson_val *causal = yyjson_obj_get(root, "is_causal");
+ if (!yyjson_equals_str(yyjson_obj_get(root, "model_type"), "xlm-roberta") ||
+ !yyjson_equals_str(yyjson_obj_get(root, "hidden_act"), "gelu") ||
+ (causal && !yyjson_is_false(causal))) {
+ fprintf(stderr, "langdet_config: expected bidirectional XLM-R with GeLU activation\n");
+ yyjson_doc_free(doc);
+ return -1;
+ }
+
+ cfg->hidden_size = json_get_int(root, "hidden_size", 0);
+ cfg->n_layers = json_get_int(root, "num_hidden_layers", 0);
+ cfg->n_heads = json_get_int(root, "num_attention_heads", 0);
+ cfg->head_dim = cfg->n_heads > 0 ? cfg->hidden_size / cfg->n_heads : 0;
+ cfg->intermediate_size = json_get_int(root, "intermediate_size", 0);
+ cfg->vocab_size = json_get_int(root, "vocab_size", 0);
+ cfg->max_position_embeddings = json_get_int(root, "max_position_embeddings", 0);
+ cfg->type_vocab_size = json_get_int(root, "type_vocab_size", 1);
+ yyjson_val *eps = yyjson_obj_get(root, "layer_norm_eps");
+ cfg->layer_norm_eps = eps ? (float)yyjson_get_num(eps) : 1e-5f;
+ int pad_id = json_get_int(root, "pad_token_id", 1);
+ cfg->position_id_offset = pad_id >= 0 && pad_id < INT_MAX ? pad_id + 1 : -1;
+ yyjson_doc_free(doc);
+
+ if (cfg->hidden_size <= 0 || cfg->n_layers <= 0 || cfg->n_layers > LANGDET_MAX_LAYERS ||
+ cfg->n_heads <= 0 || cfg->head_dim <= 0 || cfg->hidden_size % cfg->n_heads != 0 ||
+ cfg->intermediate_size <= 0 || cfg->vocab_size <= 0 || cfg->type_vocab_size <= 0 ||
+ cfg->position_id_offset < 0 ||
+ cfg->max_position_embeddings <= cfg->position_id_offset ||
+ !isfinite(cfg->layer_norm_eps) || cfg->layer_norm_eps <= 0.0f) {
+ fprintf(stderr, "langdet_config: invalid encoder dimensions or normalization in %s\n", path);
+ return -1;
+ }
+ return 0;
+}