配置

darra_model_config 的每个字段填 0 或 NULL，就表示用默认值。你可以在 create 模型时把配置传进去。

typedef struct darra_model_config {
    const char* provider;
    int32_t gpu_id;
    int32_t timeout;
    const double* default_thresholds;
    size_t default_thresholds_count;
    int threads;
    int workers;
    int queue_capacity;
    int weight;
} darra_model_config;   /* 无 size 字段，全 0/NULL = 默认 */

字段

配置项	类型	默认	说明
provider	const char*	NULL = 自动挑选	后端 ID：onnx-cpu / onnx-cuda / onnx-openvino / onnx-trt-ep / onnx-directml / trt-native / paddle-cpu / paddle-gpu / edge-rknn / torchscript-cpu
gpu_id	int32_t	0 = 第一块卡	多卡机器可填 0 到 3
timeout	int32_t	0 = 不限	推理等待上限，毫秒
default_thresholds / default_thresholds_count	const double* + size_t	NULL = 不按分数过滤	每类最低分，下标 = class_id，取值 0 到 1
threads	int	0 = 自动	推理线程数，同时用于 intra 和 inter
workers	int	0/1 = 单会话	并发会话数
queue_capacity	int	0 = 64	会话池队列上限，满了按 overflow 策略处理
weight	int	0 = 100	同卡多实例份额，只在 open 之前有效

修改配置的函数一览

这些设置要在 open 之前登记，下一次 open 才生效；timeout 和 default_thresholds 运行期间改了也有效。

DARRA_API int32_t darra_model_set_provider(darra_model* h, const char* provider_or_null);
DARRA_API int32_t darra_model_set_gpu_id(darra_model* h, int32_t gpu_id);
DARRA_API int32_t darra_model_set_timeout(darra_model* h, int32_t timeout_ms);
DARRA_API int32_t darra_model_set_default_thresholds(darra_model* h, const double* th, size_t count);
DARRA_API int32_t darra_model_set_threads(darra_model* h, int threads);
DARRA_API int32_t darra_model_set_workers(darra_model* h, int workers);
DARRA_API int32_t darra_model_set_queue(darra_model* h, int queue_capacity);
DARRA_API int32_t darra_model_set_weight(darra_model* h, int weight);

要给类别名赋值，用 darra_model_set_labels / darra_model_clear_labels，随时可调，open 后生效。