ml: add model_file::float_type() which reads type from GGUF metadata

Author: Acly

include/visp/ml.h

@@ -83,6 +83,7 @@ struct model_file {
 
     VISP_API int64_t n_tensors() const;
     VISP_API std::string_view arch() const;
+    VISP_API ggml_type float_type() const;
     VISP_API tensor_data_layout tensor_layout() const;
 
     VISP_API int64_t key(char const* name) const;

src/cli/cli.cpp

@@ -262,9 +262,13 @@ std::tuple<model_file, model_weights> load_model_weights(
         preferred_layout = file.tensor_layout();
     }
     model_transfer(file, weights, dev, dev.preferred_float_type(), preferred_layout);
-
     printf("done (%s)\n", t.elapsed_str());
-    printf("- float type: %s\n", ggml_type_name(weights.float_type()));
+
+    ggml_type ftype = file.float_type();
+    if (ftype == GGML_TYPE_COUNT) {
+        ftype = weights.float_type();
+    }
+    printf("- float type: %s\n", ggml_type_name(ftype));
     if (preferred_layout != tensor_data_layout::unknown) {
         printf("- tensor layout: %s\n", to_string(preferred_layout));
     }

src/visp/ml.cpp

@@ -227,6 +227,15 @@ std::string_view model_file::arch() const {
     return get_string("general.architecture");
 }
 
+ggml_type model_file::float_type() const {
+    if (int64_t key_id = gguf_find_key(gguf.get(), "general.file_type"); key_id != -1) {
+        if (gguf_get_kv_type(gguf.get(), key_id) == GGUF_TYPE_UINT32) {
+            return (ggml_type)gguf_get_val_u32(gguf.get(), key_id);
+        }
+    }
+    return GGML_TYPE_COUNT;
+}
+
 tensor_data_layout model_file::tensor_layout() const {
     fixed_string<64> str;
     int64_t key = gguf_find_key(gguf.get(), format(str, "{}.tensor_data_layout", arch()));