feat(model): 新增 ModelLoader 链式构建 API 及 ORT GPU/Tract 多线程配置

- 在 ddddocr-core 中定义 ModelBuilder Trait 及其错误类型
- ddddocr-ort 支持 use_gpu、device_id 及 num_threads 链式配置与 CUDA 硬件加速
- ddddocr-tract 基于 multithread-mm 特性支持 CPU 线程数控制
- 支持基于 tract-linalg 配置推理线程数，显式引入 tract-linalg 的 multithread-mm 特性，控制 GEMM 算子并发
- 优化线程池加载策略，适配 Tokio 异步及 CLI 等多场景

ddddocr-tract/src/det/session.rs

@@ -2,7 +2,8 @@ use crate::types::Session;
 use ddddocr_core::error::{Result, TensorError};
 use ddddocr_core::{DetEngine, DetOutput, InferenceEngine};
 use ndarray::Ix3;
-use tract_onnx::prelude::{tvec, IntoTensor, Tensor};
+// use tract_onnx::prelude::{tvec, IntoTensor, Tensor};
+use tract_onnx::prelude::*;
 #[derive(Debug)]
 pub struct DetSession {
     pub session: Session,
@@ -29,7 +30,8 @@ impl InferenceEngine for DetSession {
         println!("模型输出原始数据: {:?}", result);
         // Ok(result.swap_remove(0).into_tensor())
         let raw_tensor = result.swap_remove(0).into_tensor();
-        let array_d = raw_tensor.into_array::<f32>().map_err(|_| {
+        // raw_tensor.into_plain_array()?
+        let array_d = raw_tensor.into_plain_array::<f32>().map_err(|_| {
             TensorError::Engine("Tract 实体张量无法转换为 ndarray::ArrayD".to_string())
         })?;
         // 提前利用克隆(Clone)备份好当前未转维度前的真实 shape (Vec<usize>)

ddddocr-tract/src/lib.rs

@@ -4,7 +4,6 @@ mod ocr;
 mod types;
 mod error;
 
-pub use ddddocr_core::ocr::OcrBuilder;
-pub use ddddocr_core::{SlideResult, Slider};
+pub use ddddocr_core::{SlideResult, Slider,OcrBuilder};
 pub use det::session::DetSession;
 pub use ocr::session::OcrSession;

ddddocr-tract/src/loader.rs

@@ -3,5 +3,5 @@ mod metadata;
 mod model;
 
 pub use error::{Error, ParseError, Result};
-pub use metadata::ModelMetadata;
-pub use model::ModelLoader;
+pub use metadata::{ModelMetadataDto, NormalizationDto, TractModelMetadata};
+pub use model::TractModelLoader;

ddddocr-tract/src/loader/metadata.rs

@@ -1,14 +1,14 @@
 use crate::loader::error::{Error, Result};
 
-pub use ddddocr_core::ModelMetadata;
-use ddddocr_core::ocr::Resize;
+use ddddocr_core::ModelMetadata;
+use ddddocr_core::Resize;
 use ddddocr_core::{Charset, Normalization};
 use serde::Deserialize;
 use std::borrow::Cow;
 
 #[derive(Deserialize)]
 #[serde(rename_all = "snake_case")] // 支持 json 中写 "zero_to_one" 或 "minus_one_to_one"
-enum NormalizationDto {
+pub enum NormalizationDto {
     /// 映射到 [0.0, 1.0] -> pixel / 255.0
     ZeroToOne,
     /// 映射到 [-1.0, 1.0] -> (pixel / 255.0 - 0.5) / 0.5
@@ -26,7 +26,7 @@ impl From<NormalizationDto> for Normalization {
 
 /// 仅用于反序列化 JSON 的中间临时结构体（DTO）
 #[derive(Deserialize)]
-struct ModelMetadataDto {
+pub struct ModelMetadataDto {
     charset: Vec<String>,
     word: bool,
     #[serde(alias = "image")]

ddddocr-tract/src/loader/model.rs

@@ -1,17 +1,54 @@
-use crate::loader::error;
 use crate::loader::error::{Error, ParseError, Result};
 use crate::types::Session;
+use ddddocr_core::ModelBuilder;
 use std::io::Cursor;
+use tract_linalg::multithread::{Executor, set_default_executor};
 use tract_onnx::onnx;
 use tract_onnx::prelude::*;
 
-pub struct ModelLoader;
+pub struct TractModelLoader;
+impl TractModelLoader {
+    /// 获取针对 Tract 后端的链式构建器
+    pub fn builder() -> TractModelBuilder {
+        TractModelBuilder::default()
+    }
+}
 
-impl ModelLoader {
-    pub fn model_for_path<P>(model_path: P) -> Result<Session>
+/// Tract 专用的链式构建器
+#[derive(Debug, Clone, Default)]
+pub struct TractModelBuilder {
+    num_threads: Option<usize>,
+}
+
+impl TractModelBuilder {
+    /// 可选扩展：设置 CPU 线程数（不提供任何 GPU 相关的 API）
+    pub fn num_threads(mut self, threads: usize) -> Self {
+        self.num_threads = Some(threads);
+        self
+    }
+    fn setup_tract_threads(&self) {
+        // 💡 1. 如果设置了线程数，可以通过 tract 的 multithread 配置应用给 model
+        if let Some(threads) = self.num_threads {
+            // 在 Tract 中可以通过 set_num_threads 或设置底层环境控制并发
+            // （注：Tract 0.20+ 版本支持全局/局部线程控制）
+            let executor = if threads <= 1 {
+                Executor::SingleThread
+            } else {
+                Executor::multithread(threads)
+            };
+            set_default_executor(executor);
+        }
+    }
+}
+
+impl ModelBuilder for TractModelBuilder {
+    type Session = Session;
+    type Error = Error;
+    fn model_for_path<P>(&self, model_path: P) -> Result<Session>
     where
         P: AsRef<std::path::Path>,
     {
+        self.setup_tract_threads();
         let path_ref = model_path.as_ref();
 
         let session = onnx()
@@ -27,7 +64,8 @@ impl ModelLoader {
         Ok(session)
     }
     /// 策略 B：从内存字节流加载模型（配合 include_bytes! 使用）
-    pub fn model_from_bytes(model_bytes: &[u8]) -> Result<Session> {
+    fn model_from_bytes(&self, model_bytes: &[u8]) -> Result<Session> {
+        self.setup_tract_threads();
         // 使用 std::io::Cursor 将 &[u8] 包装为可读的流（实现 std::io::Read）
         let mut cursor = Cursor::new(model_bytes);
 
@@ -52,7 +90,7 @@ mod tests {
 
     /// 辅助函数：动态构建一个简单的 ONNX/Tract 内存模型图用于测试
     fn create_test_model() -> std::result::Result<Session, anyhow::Error> {
-        let mut rect = tract_onnx::prelude::Graph::default();
+        let mut rect = TypedModel::default();
 
         // 0.21.10 最稳妥的静态 Fact 构建
         let input_fact = TypedFact::dt_shape(DatumType::F32, &[1, 3, 224, 224]);
@@ -63,8 +101,6 @@ mod tests {
 
         rect.set_input_outlets(&[input_node.into()])
             .map_err(|e| anyhow::anyhow!("{:?}", e))?;
-        rect.set_output_outlets(&[input_node.into()])
-            .map_err(|e| anyhow::anyhow!("{:?}", e))?;
 
         let typed = rect
             .into_optimized()

ddddocr-tract/src/ocr/session.rs

@@ -1,5 +1,5 @@
-use crate::loader::ModelMetadata;
 use crate::types::Session;
+use ddddocr_core::ModelMetadata;
 use ddddocr_core::error::{DdddError, Result, TensorError};
 use ddddocr_core::utils::normalize_ocr_logits;
 use ddddocr_core::{InferenceEngine, OcrEngine, OcrOutput};
@@ -83,7 +83,7 @@ impl InferenceEngine for OcrSession {
         match raw_tensor.datum_type() {
             DatumType::I64 => {
                 let array_d = raw_tensor
-                    .into_array::<i64>()
+                    .into_plain_array::<i64>()
                     .map_err(|_| TensorError::Engine("Tract 无法获取 i64 内存视图".to_string()))?;
                 // .context("Tract 无法获取 i64 内存视图")?;
                 // 🌟 提前提取真实维度
@@ -101,8 +101,9 @@ impl InferenceEngine for OcrSession {
             DatumType::F32 => {
                 let shape = raw_tensor.shape();
                 println!("模型输出shape数据: {:?}", shape);
+                // raw_tensor.to_plain_array_view()
                 let view = raw_tensor
-                    .to_array_view::<f32>()
+                    .to_plain_array_view::<f32>()
                     .map_err(|_| TensorError::Engine("Tract 无法获取 f32 内存视图".to_string()))?;
                 // 1. 极其纯粹的、无拷贝的多维 Shape 压扁清洗
                 normalize_ocr_logits(view, shape)

ddddocr-tract/src/types.rs

@@ -1,3 +1,4 @@
-use tract_onnx::prelude::{Graph, RunnableModel, TypedFact, TypedOp};
+use std::sync::Arc;
+use tract_onnx::prelude::TypedRunnableModel;
 
-pub type Session = RunnableModel<TypedFact, Box<dyn TypedOp>, Graph<TypedFact, Box<dyn TypedOp>>>;
+pub type Session = Arc<TypedRunnableModel>;