chore(代码变更): 1. 清理无用代码 2. 更新依赖库 3. 优化项目结构 4. 修复小型bug 5. 增加注释以提高可读性

2026-07-08 10:44:32 +08:00
parent 4b342ef62f
commit d88afb5fe7
22 changed files with 1068 additions and 654372 deletions
@@ -1,24 +1,14 @@
-use burn::backend::Wgpu;
-use minicpm_convert::export_model;
+use minicpm_convert::export_model_q8;
 use std::path::Path;

 fn main() -> anyhow::Result<()> {
-    let device = Default::default();
-
    let safetensors_path = Path::new("MiniCPM5-1B/model-00000-of-00001.safetensors");
    let config_path = Path::new("MiniCPM5-1B/config.json");
    let tokenizer_path = Path::new("MiniCPM5-1B/tokenizer.json");
-    let output_dir = Path::new("model");

-    export_model::<Wgpu>(
-        safetensors_path,
-        config_path,
-        tokenizer_path,
-        output_dir,
-        &device,
-    )?;
+    // Q8 量化导出（INT8，文件约 1/4 大小）
+    export_model_q8(safetensors_path, config_path, tokenizer_path, Path::new("model_q8"))?;

    println!("模型转换完成！");
-
    Ok(())
 }
@@ -1,9 +1,9 @@
 [package]
-name = "minimal-inference"
+name = "flex-backend"
 version = "0.1.0"
 edition = "2021"

 [dependencies]
 minicpm-inference = { path = "../../crates/minicpm-inference" }
-burn = { version = "0.21", features = ["std", "wgpu"] }
+burn = { version = "0.21", default-features = false, features = ["std", "flex"] }
 anyhow = "1.0"
@@ -0,0 +1,82 @@
+// Flex 后端推理示例 —— 适用于 Tauri / WebAssembly 环境
+//
+// burn-flex 是纯 Rust CPU 后端，无 GPU 依赖，完美适配 Tauri 应用。
+// 支持 SIMD 加速（AVX2/NEON）和多线程（rayon）。
+//
+// 运行方式：cargo run --release -p flex-backend [--full|--half|--q8]
+
+use burn::backend::{flex::FlexDevice, Flex};
+use minicpm_inference::{GenerationConfig, MiniCPM};
+use std::io::Write;
+use std::time::Instant;
+
+fn main() -> anyhow::Result<()> {
+    let args: Vec<String> = std::env::args().collect();
+    let mode = args.get(1).map(|s| s.as_str()).unwrap_or("--full");
+
+    let (model_dir, config_path, tokenizer_path, is_q8) = match mode {
+        "--half" => (
+            "model_half",
+            "MiniCPM5-1B/config.json",
+            "MiniCPM5-1B/tokenizer.json",
+            false,
+        ),
+        "--q8" => (
+            "model_q8",
+            "MiniCPM5-1B/config.json",
+            "MiniCPM5-1B/tokenizer.json",
+            true,
+        ),
+        _ => (
+            "model",
+            "MiniCPM5-1B/config.json",
+            "MiniCPM5-1B/tokenizer.json",
+            false,
+        ),
+    };
+
+    println!("正在加载模型（后端：Flex，模式：{mode}）...");
+    println!("Flex 后端特性：纯 Rust、SIMD 加速、多线程");
+    let start = Instant::now();
+
+    let device = FlexDevice::default();
+    let model: MiniCPM<Flex> = if is_q8 {
+        MiniCPM::<Flex>::load_q8(model_dir, config_path, tokenizer_path, &device)?
+    } else {
+        let model_path = format!("{model_dir}/model");
+        MiniCPM::<Flex>::load(&model_path, config_path, tokenizer_path, &device)?
+    };
+
+    println!("模型加载完成，耗时: {:.2?}", start.elapsed());
+
+    let prompt = "需要处理生成一个河南的旅游计划，要求如下：
+1. 旅游计划为三天两夜的行程安排。
+2. 每天的行程安排包括景点、餐饮和住宿建议。
+3. 景点安排要考虑交通便利性和时间合理性。
+请生成详细的旅游计划，包含每天的行程安排、景点介绍、餐饮推荐和住宿建议。
+";
+    println!("\n用户: {prompt}");
+    print!("Assistant: ");
+    std::io::stdout().flush()?;
+
+    let config = GenerationConfig {
+        max_new_tokens: Some(1200),
+        temperature: 0.7,
+        top_p: 0.95,
+    };
+
+    let start = Instant::now();
+    let stream = model.generate_stream(prompt, false, &config)?;
+
+    for text in stream {
+        print!("{text}");
+        std::io::stdout().flush().ok();
+    }
+
+    let elapsed = start.elapsed();
+
+    println!();
+    println!("\n生成耗时: {:.2?}", elapsed);
+
+    Ok(())
+}
@@ -1,54 +0,0 @@
-// 运行方式：
-// 1. 确保已在项目根目录
-// 2. 运行：cargo run --release -p minimal-inference
-// 3. 确保模型文件存在：
-//    - model/model.mpk (模型权重)
-//    - MiniCPM5-1B/config.json (模型配置)
-//    - MiniCPM5-1B/tokenizer.json (分词器)
-
-use burn::backend::Wgpu;
-use minicpm_inference::{GenerationConfig, MiniCPM};
-use std::time::Instant;
-
-fn main() -> anyhow::Result<()> {
-    let device = Default::default();
-
-    println!("正在加载模型...");
-    let start = Instant::now();
-
-    let model = MiniCPM::<Wgpu>::load(
-        "model/model",
-        "MiniCPM5-1B/config.json",
-        "MiniCPM5-1B/tokenizer.json",
-        &device,
-    )?;
-
-    println!("模型加载完成，耗时: {:.2?}", start.elapsed());
-
-    let prompt = "我怕人家一说要改前端文件，明天咔嚓甩给我了";
-    println!("\n用户: {}", prompt);
-    println!("Assistant: ");
-
-    let config = GenerationConfig {
-        max_new_tokens: Some(200),
-        temperature: 0.7,
-        top_p: 0.95,
-    };
-
-    let start = Instant::now();
-    let _response = model.generate_stream(
-        prompt,
-        false,
-        &config,
-        |token| {
-            print!("{}", token);
-            std::io::Write::flush(&mut std::io::stdout()).ok();
-        },
-    )?;
-    let elapsed = start.elapsed();
-
-    println!();
-    println!("\n生成耗时: {:.2?}", elapsed);
-
-    Ok(())
-}
@@ -0,0 +1,9 @@
+[package]
+name = "wgpu-backend"
+version = "0.1.0"
+edition = "2021"
+
+[dependencies]
+minicpm-inference = { path = "../../crates/minicpm-inference" }
+burn = { version = "0.21", default-features = false, features = ["std", "wgpu"] }
+anyhow = "1.0"
@@ -0,0 +1,82 @@
+// Wgpu 后端推理示例 —— GPU 加速推理
+//
+// 依赖 WebGPU API，支持 CUDA/Metal/Vulkan 等平台。
+// 注意：与 Tauri 应用可能存在兼容性冲突，建议使用 Flex 后端。
+//
+// 运行方式：cargo run --release -p wgpu-backend [--full|--half|--q8]
+
+use burn::backend::{wgpu::WgpuDevice, Wgpu};
+use minicpm_inference::{GenerationConfig, MiniCPM};
+use std::io::Write;
+use std::time::Instant;
+
+fn main() -> anyhow::Result<()> {
+    let args: Vec<String> = std::env::args().collect();
+    let mode = args.get(1).map(|s| s.as_str()).unwrap_or("--full");
+
+    let (model_dir, config_path, tokenizer_path, is_q8) = match mode {
+        "--half" => (
+            "model_half",
+            "MiniCPM5-1B/config.json",
+            "MiniCPM5-1B/tokenizer.json",
+            false,
+        ),
+        "--q8" => (
+            "model_q8",
+            "MiniCPM5-1B/config.json",
+            "MiniCPM5-1B/tokenizer.json",
+            true,
+        ),
+        _ => (
+            "model",
+            "MiniCPM5-1B/config.json",
+            "MiniCPM5-1B/tokenizer.json",
+            false,
+        ),
+    };
+
+    println!("正在加载模型（后端：Wgpu，模式：{mode}）...");
+    println!("Wgpu 后端特性：GPU 加速（CUDA/Metal/Vulkan）");
+    let start = Instant::now();
+
+    let device = WgpuDevice::default();
+    let model: MiniCPM<Wgpu> = if is_q8 {
+        MiniCPM::<Wgpu>::load_q8(model_dir, config_path, tokenizer_path, &device)?
+    } else {
+        let model_path = format!("{model_dir}/model");
+        MiniCPM::<Wgpu>::load(&model_path, config_path, tokenizer_path, &device)?
+    };
+
+    println!("模型加载完成，耗时: {:.2?}", start.elapsed());
+
+    let prompt = "需要处理生成一个河南的旅游计划，要求如下：
+1. 旅游计划为三天两夜的行程安排。
+2. 每天的行程安排包括景点、餐饮和住宿建议。
+3. 景点安排要考虑交通便利性和时间合理性。
+请生成详细的旅游计划，包含每天的行程安排、景点介绍、餐饮推荐和住宿建议。
+";
+    println!("\n用户: {prompt}");
+    print!("Assistant: ");
+    std::io::stdout().flush()?;
+
+    let config = GenerationConfig {
+        max_new_tokens: Some(1200),
+        temperature: 0.7,
+        top_p: 0.95,
+    };
+
+    let start = Instant::now();
+    let stream = model.generate_stream(prompt, false, &config)?;
+
+    for text in stream {
+        print!("{text}");
+        std::io::stdout().flush().ok();
+    }
+
+    let elapsed = start.elapsed();
+
+    println!();
+    println!("\n生成耗时: {:.2?}", elapsed);
+
+    Ok(())
+}