feat(版本更新): 1. 将 minicpm-convert、minicpm-core 和 minicpm-inference 的版本更新至 0.2.0 2. 更新 README 文件以反映新版本信息

.workbuddy/memory/2026-07-01.md

@@ -19,3 +19,17 @@
 
 - `minicpm-inference/src/lib.rs`: 新增 `generate_stream` 函数和 `MiniCPM::generate_stream` 方法，每 token 通过回调 `impl FnMut(&str)` 输出解码文本
 - `examples/minimal-inference/src/main.rs`: 改为流式调用，`print!` + `flush` 实时输出
+
+## 升级 0.1.2 重新发布（含流式功能）
+
+- minicpm-core, minicpm-convert, minicpm-inference 全部升级到 0.1.2
+- 发布全部成功
+
+## Q8 量化导出和加载
+
+- `minicpm-convert`: 新增 `export_model_q8`，2D 权重 per-tensor INT8 量化（scale = max(|f|)/127），1D norm 权重存 F16；输出 `model.q8.json` + `model.q8.bin`
+- `minicpm-inference`: 新增 `MiniCPM::load_q8`，加载 Q8 模型并在内存中反量化为 f32 推理；新增依赖 `memmap2`、`serde`、`serde_json`
+- `examples/convert`: 支持 Q8 导出
+- `examples/minimal-inference`: 支持 `--full|--half|--q8` 三种模式
+- 量化方式：`q = clamp(round(f / scale), -127, 127)`，`f' = q * scale`
+

Cargo.toml

@@ -3,7 +3,12 @@ members = [
   "crates/minicpm-core",
   "crates/minicpm-convert",
   "crates/minicpm-inference",
-  "examples/minimal-inference",
   "examples/convert",
+  "examples/wgpu-backend",
 ]
 resolver = "2"
+
+[profile.release]
+opt-level = 3
+lto = true
+codegen-units = 1

crates/minicpm-convert/Cargo.toml

@@ -1,12 +1,13 @@
 [package]
 name = "minicpm-convert"
-version = "0.1.1"
+version = "0.2.0"
 edition = "2021"
 publish = ["gitea"]
 
 [dependencies]
-minicpm-core = { path = "../minicpm-core", version = "0.1.1", registry = "gitea" }
+minicpm-core = { path = "../minicpm-core", version = "0.2.0", registry = "gitea" }
 burn = { version = "0.21", default-features = false, features = ["std"] }
 memmap2 = "0.9"
 anyhow = "1.0"
 serde_json = "1.0"
+serde = { version = "1.0", features = ["derive"] }

crates/minicpm-convert/src/format/full.rs

@@ -0,0 +1,31 @@
+use burn::module::Module;
+use burn::record::{FullPrecisionSettings, NamedMpkFileRecorder};
+use burn::tensor::backend::Backend;
+use std::path::Path;
+
+use crate::loader::{load_into_model, TensorStore};
+use minicpm_core::config::LlamaConfig;
+use minicpm_core::model::LlamaForCausalLM;
+
+/// 导出全精度模型（f32）
+pub fn export<B: Backend>(
+    store: &TensorStore,
+    config: &LlamaConfig,
+    output_dir: &Path,
+    device: &B::Device,
+) -> anyhow::Result<()> {
+    println!("创建模型结构...");
+    let model = LlamaForCausalLM::<B>::new(config.clone(), device);
+
+    println!("加载 safetensors 权重...");
+    let model = load_into_model(model, store, device);
+
+    println!("保存为 MPK 格式（全精度）...");
+    std::fs::create_dir_all(output_dir)?;
+    let output_path = output_dir.join("model");
+    let recorder = NamedMpkFileRecorder::<FullPrecisionSettings>::new();
+    model.save_file(&output_path, &recorder)?;
+
+    println!("全精度模型已导出到: {:?}", output_dir);
+    Ok(())
+}

crates/minicpm-convert/src/format/mod.rs

@@ -0,0 +1,44 @@
+pub mod full;
+
+/// 导出格式
+#[derive(Debug, Clone, Copy, PartialEq)]
+pub enum Format {
+    /// 全精度 f32
+    Full,
+}
+
+impl Format {
+    pub fn name(&self) -> &'static str {
+        match self {
+            Format::Full => "全精度",
+        }
+    }
+
+    pub fn dir_name(&self) -> &'static str {
+        match self {
+            Format::Full => "model",
+        }
+    }
+}
+
+/// 导出任务配置
+pub struct ExportTask {
+    pub format: Format,
+    pub output_dir: String,
+}
+
+impl ExportTask {
+    pub fn new(format: Format) -> Self {
+        Self {
+            output_dir: format.dir_name().to_string(),
+            format,
+        }
+    }
+
+    pub fn with_dir(format: Format, output_dir: impl Into<String>) -> Self {
+        Self {
+            output_dir: output_dir.into(),
+            format,
+        }
+    }
+}

crates/minicpm-convert/src/lib.rs

@@ -1,258 +1,46 @@
-use minicpm_core::config::LlamaConfig;
-use minicpm_core::model::LlamaForCausalLM;
-use burn::module::{Module, Param};
-use burn::nn::{EmbeddingRecord, LinearRecord};
-use burn::record::{FullPrecisionSettings, NamedMpkFileRecorder};
+pub mod format;
+pub mod loader;
+pub mod utils;
+
+pub use format::{ExportTask, Format};
+pub use loader::TensorStore;
+
 use burn::tensor::backend::Backend;
-use burn::tensor::{Float, Shape, Tensor, TensorData};
-use memmap2::Mmap;
-use std::collections::HashMap;
 use std::path::Path;
 
-pub fn export_model<B: Backend>(
+use minicpm_core::config::LlamaConfig;
+
+/// 执行导出任务
+pub fn run_export<B: Backend>(
     safetensors_path: &Path,
     config_path: &Path,
     tokenizer_path: &Path,
-    output_dir: &Path,
+    tasks: &[ExportTask],
     device: &B::Device,
 ) -> anyhow::Result<()> {
-    println!("开始转换 MiniCPM 模型为 Burn 格式...");
+    println!("开始转换 MiniCPM 模型...");
+    println!("源文件: {:?}", safetensors_path);
 
-    println!("加载配置文件: {:?}", config_path);
+    println!("加载配置文件...");
     let config = LlamaConfig::from_json(config_path.to_str().unwrap())?;
 
-    println!("创建模型结构...");
-    let model = LlamaForCausalLM::<B>::new(config, device);
+    println!("解析 safetensors 文件...");
+    let store = TensorStore::from_file(safetensors_path)?;
 
-    println!("加载 safetensors 权重...");
-    let model = load_safetensors(model, safetensors_path, device)?;
+    for task in tasks {
+        println!("\n--- 导出 {} ---", task.format.name());
+        let output_dir = Path::new(&task.output_dir);
 
-    println!("保存为 MPK 格式...");
-    std::fs::create_dir_all(output_dir)?;
-    let output_path = output_dir.join("model");
-    let recorder = NamedMpkFileRecorder::<FullPrecisionSettings>::new();
-    model.save_file(&output_path, &recorder)?;
+        match task.format {
+            Format::Full => {
+                format::full::export::<B>(&store, &config, output_dir, device)?;
+            }
+        }
 
-    println!("拷贝配置文件和 tokenizer...");
-    std::fs::copy(config_path, output_dir.join("config.json"))?;
-    std::fs::copy(tokenizer_path, output_dir.join("tokenizer.json"))?;
+        std::fs::copy(config_path, output_dir.join("config.json"))?;
+        std::fs::copy(tokenizer_path, output_dir.join("tokenizer.json"))?;
+    }
 
-    println!("模型已成功导出到: {:?}", output_dir);
+    println!("\n所有导出任务完成！");
     Ok(())
 }
-pub fn load_safetensors<B: Backend>(
-    model: LlamaForCausalLM<B>,
-    path: &Path,
-    device: &B::Device,
-) -> anyhow::Result<LlamaForCausalLM<B>> {
-    let file = std::fs::File::open(path)?;
-    let mmap = unsafe { Mmap::map(&file)? };
-
-    let header_len = u64::from_le_bytes(mmap[..8].try_into().unwrap()) as usize;
-    let header_bytes = &mmap[8..8 + header_len];
-    let header: serde_json::Value = serde_json::from_slice(header_bytes)?;
-
-    let mut tensors = HashMap::new();
-    let mut shapes = HashMap::new();
-    let mut dtypes = HashMap::new();
-
-    let data_offset = 8 + header_len;
-
-    if let Some(obj) = header.as_object() {
-        for (name, info) in obj {
-            if name == "__metadata__" {
-                continue;
-            }
-            if let Some(info_obj) = info.as_object() {
-                if let Some(offsets) = info_obj.get("data_offsets") {
-                    if let Some(arr) = offsets.as_array() {
-                        let start = arr[0].as_u64().unwrap() as usize;
-                        let end = arr[1].as_u64().unwrap() as usize;
-                        let data = &mmap[data_offset + start..data_offset + end];
-                        tensors.insert(name.clone(), data.to_vec());
-                    }
-                }
-                if let Some(shape) = info_obj.get("shape") {
-                    if let Some(arr) = shape.as_array() {
-                        let shape_vec: Vec<usize> = arr
-                            .iter()
-                            .map(|v| v.as_u64().unwrap() as usize)
-                            .collect();
-                        shapes.insert(name.clone(), shape_vec);
-                    }
-                }
-                if let Some(dtype) = info_obj.get("dtype") {
-                    if let Some(s) = dtype.as_str() {
-                        dtypes.insert(name.clone(), s.to_string());
-                    }
-                }
-            }
-        }
-    }
-
-    let mut model = model;
-
-    if let Some(weight) = load_tensor_2d::<B>(&tensors, &shapes, &dtypes, "model.embed_tokens.weight", device) {
-        let record = EmbeddingRecord {
-            weight: Param::from_tensor(weight),
-        };
-        model.model.embed_tokens = model.model.embed_tokens.clone().load_record(record);
-    }
-
-    for i in 0..model.config.num_hidden_layers {
-        let prefix = format!("model.layers.{i}");
-
-        if let Some(w) = load_tensor_2d::<B>(&tensors, &shapes, &dtypes, &format!("{prefix}.self_attn.q_proj.weight"), device) {
-            let record = LinearRecord { weight: Param::from_tensor(w.transpose()), bias: None };
-            model.model.layers[i].self_attn.q_proj = model.model.layers[i].self_attn.q_proj.clone().load_record(record);
-        }
-        if let Some(w) = load_tensor_2d::<B>(&tensors, &shapes, &dtypes, &format!("{prefix}.self_attn.k_proj.weight"), device) {
-            let record = LinearRecord { weight: Param::from_tensor(w.transpose()), bias: None };
-            model.model.layers[i].self_attn.k_proj = model.model.layers[i].self_attn.k_proj.clone().load_record(record);
-        }
-        if let Some(w) = load_tensor_2d::<B>(&tensors, &shapes, &dtypes, &format!("{prefix}.self_attn.v_proj.weight"), device) {
-            let record = LinearRecord { weight: Param::from_tensor(w.transpose()), bias: None };
-            model.model.layers[i].self_attn.v_proj = model.model.layers[i].self_attn.v_proj.clone().load_record(record);
-        }
-        if let Some(w) = load_tensor_2d::<B>(&tensors, &shapes, &dtypes, &format!("{prefix}.self_attn.o_proj.weight"), device) {
-            let record = LinearRecord { weight: Param::from_tensor(w.transpose()), bias: None };
-            model.model.layers[i].self_attn.o_proj = model.model.layers[i].self_attn.o_proj.clone().load_record(record);
-        }
-
-        if let Some(w) = load_tensor_2d::<B>(&tensors, &shapes, &dtypes, &format!("{prefix}.mlp.gate_proj.weight"), device) {
-            let record = LinearRecord { weight: Param::from_tensor(w.transpose()), bias: None };
-            model.model.layers[i].mlp.gate_proj = model.model.layers[i].mlp.gate_proj.clone().load_record(record);
-        }
-        if let Some(w) = load_tensor_2d::<B>(&tensors, &shapes, &dtypes, &format!("{prefix}.mlp.up_proj.weight"), device) {
-            let record = LinearRecord { weight: Param::from_tensor(w.transpose()), bias: None };
-            model.model.layers[i].mlp.up_proj = model.model.layers[i].mlp.up_proj.clone().load_record(record);
-        }
-        if let Some(w) = load_tensor_2d::<B>(&tensors, &shapes, &dtypes, &format!("{prefix}.mlp.down_proj.weight"), device) {
-            let record = LinearRecord { weight: Param::from_tensor(w.transpose()), bias: None };
-            model.model.layers[i].mlp.down_proj = model.model.layers[i].mlp.down_proj.clone().load_record(record);
-        }
-
-        if let Some(w) = load_tensor_1d::<B>(&tensors, &shapes, &dtypes, &format!("{prefix}.input_layernorm.weight"), device) {
-            model.model.layers[i].input_layernorm.weight = Param::from_tensor(w);
-        }
-        if let Some(w) = load_tensor_1d::<B>(&tensors, &shapes, &dtypes, &format!("{prefix}.post_attention_layernorm.weight"), device) {
-            model.model.layers[i].post_attention_layernorm.weight = Param::from_tensor(w);
-        }
-    }
-
-    if let Some(w) = load_tensor_1d::<B>(&tensors, &shapes, &dtypes, "model.norm.weight", device) {
-        model.model.norm.weight = Param::from_tensor(w);
-    }
-
-    if let Some(w) = load_tensor_2d::<B>(&tensors, &shapes, &dtypes, "lm_head.weight", device) {
-        let record = LinearRecord { weight: Param::from_tensor(w.transpose()), bias: None };
-        model.lm_head = model.lm_head.clone().load_record(record);
-    } else if model.config.tie_word_embeddings {
-        let embed_weight = model.model.embed_tokens.weight.val();
-        let record = LinearRecord { weight: Param::from_tensor(embed_weight.transpose()), bias: None };
-        model.lm_head = model.lm_head.clone().load_record(record);
-    }
-
-    Ok(model)
-}
-
-fn load_tensor_2d<B: Backend>(
-    tensors: &HashMap<String, Vec<u8>>,
-    shapes: &HashMap<String, Vec<usize>>,
-    dtypes: &HashMap<String, String>,
-    name: &str,
-    device: &B::Device,
-) -> Option<Tensor<B, 2, Float>> {
-    let data = tensors.get(name)?;
-    let shape = shapes.get(name)?;
-    let dtype = dtypes.get(name)?;
-
-    assert_eq!(shape.len(), 2, "Expected 2D tensor for {name}");
-
-    let f32_data = convert_to_f32(data, dtype);
-    let shape_arr: [usize; 2] = [shape[0], shape[1]];
-    let tensor_data = TensorData::new(f32_data, Shape::new(shape_arr));
-    Some(Tensor::from_data(tensor_data, device))
-}
-
-fn load_tensor_1d<B: Backend>(
-    tensors: &HashMap<String, Vec<u8>>,
-    shapes: &HashMap<String, Vec<usize>>,
-    dtypes: &HashMap<String, String>,
-    name: &str,
-    device: &B::Device,
-) -> Option<Tensor<B, 1, Float>> {
-    let data = tensors.get(name)?;
-    let shape = shapes.get(name)?;
-    let dtype = dtypes.get(name)?;
-
-    assert_eq!(shape.len(), 1, "Expected 1D tensor for {name}");
-
-    let f32_data = convert_to_f32(data, dtype);
-    let shape_arr: [usize; 1] = [shape[0]];
-    let tensor_data = TensorData::new(f32_data, Shape::new(shape_arr));
-    Some(Tensor::from_data(tensor_data, device))
-}
-
-fn convert_to_f32(data: &[u8], dtype: &str) -> Vec<f32> {
-    match dtype {
-        "BF16" => data
-            .chunks_exact(2)
-            .map(|chunk| {
-                let bytes: [u8; 2] = chunk.try_into().unwrap();
-                bf16_to_f32(u16::from_le_bytes(bytes))
-            })
-            .collect(),
-        "F16" => data
-            .chunks_exact(2)
-            .map(|chunk| {
-                let bytes: [u8; 2] = chunk.try_into().unwrap();
-                f16_to_f32(u16::from_le_bytes(bytes))
-            })
-            .collect(),
-        "F32" => data
-            .chunks_exact(4)
-            .map(|chunk| {
-                let bytes: [u8; 4] = chunk.try_into().unwrap();
-                f32::from_le_bytes(bytes)
-            })
-            .collect(),
-        _ => panic!("Unsupported dtype: {dtype}"),
-    }
-}
-
-fn bf16_to_f32(bf16: u16) -> f32 {
-    let f32_bits = (bf16 as u32) << 16;
-    f32::from_bits(f32_bits)
-}
-
-fn f16_to_f32(f16: u16) -> f32 {
-    let sign = (f16 >> 15) & 1;
-    let exp = (f16 >> 10) & 0x1F;
-    let mant = f16 & 0x3FF;
-
-    if exp == 0 {
-        if mant == 0 {
-            return if sign == 0 { 0.0 } else { -0.0 };
-        }
-        let mut m = mant as u32;
-        let mut e = 0u32;
-        while (m & 0x400) == 0 {
-            m <<= 1;
-            e += 1;
-        }
-        m &= 0x3FF;
-        let f32_exp = 127 - 14 - e;
-        let f32_bits = ((sign as u32) << 31) | (f32_exp << 23) | ((m as u32) << 13);
-        return f32::from_bits(f32_bits);
-    }
-
-    if exp == 0x1F {
-        let f32_bits = ((sign as u32) << 31) | (0xFFu32 << 23) | ((mant as u32) << 13);
-        return f32::from_bits(f32_bits);
-    }
-
-    let f32_exp = (exp as u32) + 127 - 15;
-    let f32_bits = ((sign as u32) << 31) | (f32_exp << 23) | ((mant as u32) << 13);
-    f32::from_bits(f32_bits)
-}

crates/minicpm-convert/src/loader/mod.rs

@@ -0,0 +1,3 @@
+pub mod safetensors;
+
+pub use safetensors::{load_into_model, TensorStore};

crates/minicpm-convert/src/loader/safetensors.rs

@@ -0,0 +1,185 @@
+use burn::module::{Module, Param};
+use burn::nn::{EmbeddingRecord, LinearRecord};
+use burn::tensor::backend::Backend;
+use burn::tensor::{Float, Shape, Tensor, TensorData};
+use memmap2::Mmap;
+use std::collections::HashMap;
+use std::path::Path;
+
+use crate::utils::convert_to_f32;
+use minicpm_core::model::LlamaForCausalLM;
+
+/// 从 safetensors 文件解析出的张量数据
+pub struct TensorStore {
+    tensors: HashMap<String, Vec<u8>>,
+    shapes: HashMap<String, Vec<usize>>,
+    dtypes: HashMap<String, String>,
+}
+
+impl TensorStore {
+    /// 从 safetensors 文件加载所有张量
+    pub fn from_file(path: &Path) -> anyhow::Result<Self> {
+        let file = std::fs::File::open(path)?;
+        let mmap = unsafe { Mmap::map(&file)? };
+
+        let header_len = u64::from_le_bytes(mmap[..8].try_into().unwrap()) as usize;
+        let header_bytes = &mmap[8..8 + header_len];
+        let header: serde_json::Value = serde_json::from_slice(header_bytes)?;
+
+        let mut tensors = HashMap::new();
+        let mut shapes = HashMap::new();
+        let mut dtypes = HashMap::new();
+
+        let data_offset = 8 + header_len;
+
+        if let Some(obj) = header.as_object() {
+            for (name, info) in obj {
+                if name == "__metadata__" {
+                    continue;
+                }
+                if let Some(info_obj) = info.as_object() {
+                    if let Some(offsets) = info_obj.get("data_offsets") {
+                        if let Some(arr) = offsets.as_array() {
+                            let start = arr[0].as_u64().unwrap() as usize;
+                            let end = arr[1].as_u64().unwrap() as usize;
+                            let data = &mmap[data_offset + start..data_offset + end];
+                            tensors.insert(name.clone(), data.to_vec());
+                        }
+                    }
+                    if let Some(shape) = info_obj.get("shape") {
+                        if let Some(arr) = shape.as_array() {
+                            let shape_vec: Vec<usize> = arr
+                                .iter()
+                                .map(|v| v.as_u64().unwrap() as usize)
+                                .collect();
+                            shapes.insert(name.clone(), shape_vec);
+                        }
+                    }
+                    if let Some(dtype) = info_obj.get("dtype") {
+                        if let Some(s) = dtype.as_str() {
+                            dtypes.insert(name.clone(), s.to_string());
+                        }
+                    }
+                }
+            }
+        }
+
+        Ok(Self {
+            tensors,
+            shapes,
+            dtypes,
+        })
+    }
+
+    /// 加载 2D 张量
+    pub fn load_2d<B: Backend>(&self, name: &str, device: &B::Device) -> Option<Tensor<B, 2, Float>> {
+        let data = self.tensors.get(name)?;
+        let shape = self.shapes.get(name)?;
+        let dtype = self.dtypes.get(name)?;
+
+        assert_eq!(shape.len(), 2, "期望 2D 张量: {name}");
+
+        let f32_data = convert_to_f32(data, dtype);
+        let shape_arr: [usize; 2] = [shape[0], shape[1]];
+        let tensor_data = TensorData::new(f32_data, Shape::new(shape_arr));
+        Some(Tensor::from_data(tensor_data, device))
+    }
+
+    /// 加载 1D 张量
+    pub fn load_1d<B: Backend>(&self, name: &str, device: &B::Device) -> Option<Tensor<B, 1, Float>> {
+        let data = self.tensors.get(name)?;
+        let shape = self.shapes.get(name)?;
+        let dtype = self.dtypes.get(name)?;
+
+        assert_eq!(shape.len(), 1, "期望 1D 张量: {name}");
+
+        let f32_data = convert_to_f32(data, dtype);
+        let shape_arr: [usize; 1] = [shape[0]];
+        let tensor_data = TensorData::new(f32_data, Shape::new(shape_arr));
+        Some(Tensor::from_data(tensor_data, device))
+    }
+
+    /// 获取原始字节数据
+    pub fn raw(&self, name: &str) -> Option<&[u8]> {
+        self.tensors.get(name).map(|v| v.as_slice())
+    }
+
+    pub fn shape(&self, name: &str) -> Option<&Vec<usize>> {
+        self.shapes.get(name)
+    }
+
+    pub fn dtype(&self, name: &str) -> Option<&String> {
+        self.dtypes.get(name)
+    }
+
+    pub fn names(&self) -> impl Iterator<Item = &String> {
+        self.tensors.keys()
+    }
+}
+
+/// 将 safetensors 权重加载到模型中
+pub fn load_into_model<B: Backend>(
+    model: LlamaForCausalLM<B>,
+    store: &TensorStore,
+    device: &B::Device,
+) -> LlamaForCausalLM<B> {
+    let mut model = model;
+
+    if let Some(w) = store.load_2d("model.embed_tokens.weight", device) {
+        let record = EmbeddingRecord {
+            weight: Param::from_tensor(w),
+        };
+        model.model.embed_tokens = model.model.embed_tokens.clone().load_record(record);
+    }
+
+    for i in 0..model.config.num_hidden_layers {
+        let prefix = format!("model.layers.{i}");
+
+        load_linear(&mut model, store, device, i, &format!("{prefix}.self_attn.q_proj.weight"), |m| &mut m.model.layers[i].self_attn.q_proj);
+        load_linear(&mut model, store, device, i, &format!("{prefix}.self_attn.k_proj.weight"), |m| &mut m.model.layers[i].self_attn.k_proj);
+        load_linear(&mut model, store, device, i, &format!("{prefix}.self_attn.v_proj.weight"), |m| &mut m.model.layers[i].self_attn.v_proj);
+        load_linear(&mut model, store, device, i, &format!("{prefix}.self_attn.o_proj.weight"), |m| &mut m.model.layers[i].self_attn.o_proj);
+        load_linear(&mut model, store, device, i, &format!("{prefix}.mlp.gate_proj.weight"), |m| &mut m.model.layers[i].mlp.gate_proj);
+        load_linear(&mut model, store, device, i, &format!("{prefix}.mlp.up_proj.weight"), |m| &mut m.model.layers[i].mlp.up_proj);
+        load_linear(&mut model, store, device, i, &format!("{prefix}.mlp.down_proj.weight"), |m| &mut m.model.layers[i].mlp.down_proj);
+
+        if let Some(w) = store.load_1d(&format!("{prefix}.input_layernorm.weight"), device) {
+            model.model.layers[i].input_layernorm.weight = Param::from_tensor(w);
+        }
+        if let Some(w) = store.load_1d(&format!("{prefix}.post_attention_layernorm.weight"), device) {
+            model.model.layers[i].post_attention_layernorm.weight = Param::from_tensor(w);
+        }
+    }
+
+    if let Some(w) = store.load_1d("model.norm.weight", device) {
+        model.model.norm.weight = Param::from_tensor(w);
+    }
+
+    if let Some(w) = store.load_2d("lm_head.weight", device) {
+        let record = LinearRecord { weight: Param::from_tensor(w.transpose()), bias: None };
+        model.lm_head = model.lm_head.clone().load_record(record);
+    } else if model.config.tie_word_embeddings {
+        let embed_weight = model.model.embed_tokens.weight.val();
+        let record = LinearRecord { weight: Param::from_tensor(embed_weight.transpose()), bias: None };
+        model.lm_head = model.lm_head.clone().load_record(record);
+    }
+
+    model
+}
+
+fn load_linear<B: Backend, F>(
+    model: &mut LlamaForCausalLM<B>,
+    store: &TensorStore,
+    device: &B::Device,
+    _layer_idx: usize,
+    name: &str,
+    get_field: F,
+) where
+    F: FnOnce(&mut LlamaForCausalLM<B>) -> &mut burn::nn::Linear<B>,
+{
+    if let Some(w) = store.load_2d(name, device) {
+        let record = LinearRecord { weight: Param::from_tensor(w.transpose()), bias: None };
+        let field = get_field(model);
+        *field = field.clone().load_record(record);
+    }
+}

crates/minicpm-convert/src/utils.rs

@@ -0,0 +1,84 @@
+/// BF16 → F32
+pub fn bf16_to_f32(bf16: u16) -> f32 {
+    let f32_bits = (bf16 as u32) << 16;
+    f32::from_bits(f32_bits)
+}
+
+/// F16 → F32
+pub fn f16_to_f32(f16: u16) -> f32 {
+    let sign = (f16 >> 15) & 1;
+    let exp = (f16 >> 10) & 0x1F;
+    let mant = f16 & 0x3FF;
+
+    if exp == 0 {
+        if mant == 0 {
+            return if sign == 0 { 0.0 } else { -0.0 };
+        }
+        let mut m = mant as u32;
+        let mut e = 0u32;
+        while (m & 0x400) == 0 {
+            m <<= 1;
+            e += 1;
+        }
+        m &= 0x3FF;
+        let f32_exp = 127 - 14 - e;
+        let f32_bits = ((sign as u32) << 31) | (f32_exp << 23) | ((m as u32) << 13);
+        return f32::from_bits(f32_bits);
+    }
+
+    if exp == 0x1F {
+        let f32_bits = ((sign as u32) << 31) | (0xFFu32 << 23) | ((mant as u32) << 13);
+        return f32::from_bits(f32_bits);
+    }
+
+    let f32_exp = (exp as u32) + 127 - 15;
+    let f32_bits = ((sign as u32) << 31) | (f32_exp << 23) | ((mant as u32) << 13);
+    f32::from_bits(f32_bits)
+}
+
+/// F32 → F16 位表示
+pub fn f32_to_f16_bits(v: f32) -> u16 {
+    let bits = v.to_bits();
+    let sign = bits >> 31;
+    let exp = (bits >> 23) & 0xFF;
+    let mant = bits & 0x7FFFFF;
+
+    let res: u32 = if exp == 0xFF {
+        let f16_exp = 0x1F;
+        let f16_mant = if mant == 0 { 0 } else { 0x3FF };
+        (sign << 15) | (f16_exp << 10) | f16_mant
+    } else if exp < 103 {
+        0
+    } else {
+        let new_exp = (exp as i32) - 127 + 15;
+        if new_exp >= 31 {
+            (sign << 15) | (0x1F << 10)
+        } else if new_exp <= 0 {
+            let mant_new = (mant | 0x800000) >> (113 - new_exp);
+            (sign << 15) | (mant_new & 0x3FF)
+        } else {
+            let f16_mant = mant >> 13;
+            (sign << 15) | ((new_exp as u32) << 10) | f16_mant
+        }
+    };
+    res as u16
+}
+
+/// 将原始字节按 dtype 转换为 f32 向量
+pub fn convert_to_f32(data: &[u8], dtype: &str) -> Vec<f32> {
+    match dtype {
+        "BF16" => data
+            .chunks_exact(2)
+            .map(|chunk| bf16_to_f32(u16::from_le_bytes(chunk.try_into().unwrap())))
+            .collect(),
+        "F16" => data
+            .chunks_exact(2)
+            .map(|chunk| f16_to_f32(u16::from_le_bytes(chunk.try_into().unwrap())))
+            .collect(),
+        "F32" => data
+            .chunks_exact(4)
+            .map(|chunk| f32::from_le_bytes(chunk.try_into().unwrap()))
+            .collect(),
+        _ => panic!("不支持的 dtype: {dtype}"),
+    }
+}

crates/minicpm-core/Cargo.toml

@@ -1,9 +1,13 @@
 [package]
 name = "minicpm-core"
-version = "0.1.1"
+version = "0.2.0"
 edition = "2021"
 publish = ["gitea"]
 
+[features]
+default = ["avx2"]
+avx2 = []
+
 [dependencies]
 burn = { version = "0.21", default-features = false, features = ["std"] }
 serde = { version = "1.0", features = ["derive"] }

crates/minicpm-core/src/model/attention.rs

@@ -129,7 +129,6 @@ impl<B: Backend> Attention<B> {
         let q = rope.apply(q, offset);
         let k = rope.apply(k, offset);
 
-        // 合并 KV Cache
         let (k_full, v_full) = match cache {
             Some(cache) => {
                 let k_full = Tensor::cat(vec![cache.k.clone(), k], 2);

crates/minicpm-core/src/model/decoder.rs

@@ -4,15 +4,15 @@ use burn::tensor::{Float, Tensor};
 
 use super::attention::{Attention, KVCache};
 use super::ffn::FeedForward;
-use super::norm::RmsNorm;
+use super::norm::RMSNorm;
 use super::rope::RoPE;
 
 #[derive(Module, Debug)]
 pub struct DecoderLayer<B: Backend> {
     pub self_attn: Attention<B>,
     pub mlp: FeedForward<B>,
-    pub input_layernorm: RmsNorm<B>,
-    pub post_attention_layernorm: RmsNorm<B>,
+    pub input_layernorm: RMSNorm<B>,
+    pub post_attention_layernorm: RMSNorm<B>,
 }
 
 impl<B: Backend> DecoderLayer<B> {
@@ -27,8 +27,8 @@ impl<B: Backend> DecoderLayer<B> {
     ) -> Self {
         let self_attn = Attention::new(hidden_size, n_heads, n_kv_heads, head_dim, device);
         let mlp = FeedForward::new(hidden_size, intermediate_size, device);
-        let input_layernorm = RmsNorm::new(hidden_size, rms_norm_eps, device);
-        let post_attention_layernorm = RmsNorm::new(hidden_size, rms_norm_eps, device);
+        let input_layernorm = RMSNorm::new(hidden_size, rms_norm_eps, device);
+        let post_attention_layernorm = RMSNorm::new(hidden_size, rms_norm_eps, device);
 
         Self {
             self_attn,

crates/minicpm-core/src/model/model.rs

@@ -6,7 +6,7 @@ use burn::tensor::{Float, Int, Tensor};
 use super::attention::KVCache;
 use crate::config::LlamaConfig;
 use super::decoder::DecoderLayer;
-use super::norm::RmsNorm;
+use super::norm::RMSNorm;
 use super::rope::RoPE;
 
 #[derive(Debug, Clone)]
@@ -26,7 +26,7 @@ impl<B: Backend> LlamaKVCache<B> {
 pub struct LlamaModel<B: Backend> {
     pub embed_tokens: Embedding<B>,
     pub layers: Vec<DecoderLayer<B>>,
-    pub norm: RmsNorm<B>,
+    pub norm: RMSNorm<B>,
     pub rope: RoPE,
 }
 
@@ -58,7 +58,7 @@ impl<B: Backend> LlamaForCausalLM<B> {
             ));
         }
 
-        let norm = RmsNorm::new(config.hidden_size, config.rms_norm_eps, device);
+        let norm = RMSNorm::new(config.hidden_size, config.rms_norm_eps, device);
 
         let lm_head = LinearConfig::new(config.hidden_size, config.vocab_size).init(device);
 

crates/minicpm-core/src/model/norm.rs

@@ -4,12 +4,12 @@ use burn::tensor::backend::Backend;
 use burn::tensor::{Float, Tensor};
 
 #[derive(Module, Debug)]
-pub struct RmsNorm<B: Backend> {
+pub struct RMSNorm<B: Backend> {
     pub weight: Param<Tensor<B, 1, Float>>,
     pub eps: f64,
 }
 
-impl<B: Backend> RmsNorm<B> {
+impl<B: Backend> RMSNorm<B> {
     pub fn new(dim: usize, eps: f64, device: &B::Device) -> Self {
         let weight = Initializer::Ones.init([dim], device);
         Self {

crates/minicpm-inference/Cargo.toml

@@ -1,12 +1,14 @@
 [package]
 name = "minicpm-inference"
-version = "0.1.1"
+version = "0.2.0"
 edition = "2021"
 publish = ["gitea"]
 
 [dependencies]
-minicpm-core = { path = "../minicpm-core", version = "0.1.1", registry = "gitea" }
-burn = { version = "0.21", default-features = false, features = ["std"] }
+minicpm-core = { path = "../minicpm-core", version = "0.2.0", registry = "gitea" }
+burn = { version = "0.21", default-features = false, features = ["std", "autotune"] }
 tokenizers = "0.20"
 rand = "0.8"
 anyhow = "1.0"
+serde = { version = "1.0", features = ["derive"] }
+serde_json = "1.0"

crates/minicpm-inference/src/config.rs

@@ -0,0 +1,16 @@
+#[derive(Debug, Clone)]
+pub struct GenerationConfig {
+    pub max_new_tokens: Option<usize>,
+    pub temperature: f32,
+    pub top_p: f32,
+}
+
+impl Default for GenerationConfig {
+    fn default() -> Self {
+        Self {
+            max_new_tokens: None,
+            temperature: 1.0,
+            top_p: 1.0,
+        }
+    }
+}

crates/minicpm-inference/src/generator.rs

@@ -0,0 +1,132 @@
+use burn::tensor::backend::Backend;
+use burn::tensor::{Int, Tensor};
+
+use minicpm_core::config::EosTokenId;
+use minicpm_core::model::{LlamaForCausalLM, LlamaKVCache};
+
+use crate::config::GenerationConfig;
+use crate::sampling::{CachedSampler, sample_last};
+
+pub struct TokenStream<'a, B: Backend> {
+    model: &'a LlamaForCausalLM<B>,
+    config: &'a GenerationConfig,
+    eos_token_id: &'a EosTokenId,
+    device: &'a B::Device,
+    cache: Option<LlamaKVCache<B>>,
+    last_token: Option<u32>,
+    count: usize,
+    finished: bool,
+    first_step: bool,
+    input_ids: Vec<u32>,
+    sampler: CachedSampler,
+}
+
+impl<'a, B: Backend> TokenStream<'a, B> {
+    pub fn new(
+        model: &'a LlamaForCausalLM<B>,
+        input_ids: &[u32],
+        config: &'a GenerationConfig,
+        eos_token_id: &'a EosTokenId,
+        device: &'a B::Device,
+    ) -> Self {
+        Self {
+            model,
+            config,
+            eos_token_id,
+            device,
+            cache: None,
+            last_token: None,
+            count: 0,
+            finished: false,
+            first_step: true,
+            input_ids: input_ids.to_vec(),
+            sampler: CachedSampler::new(),
+        }
+    }
+}
+
+impl<'a, B: Backend> Iterator for TokenStream<'a, B> {
+    type Item = u32;
+
+    fn next(&mut self) -> Option<Self::Item> {
+        if self.finished {
+            return None;
+        }
+
+        if let Some(max) = self.config.max_new_tokens {
+            if self.count >= max {
+                self.finished = true;
+                return None;
+            }
+        }
+
+        let next_token = if self.first_step {
+            self.first_step = false;
+            self.step_first()
+        } else {
+            self.step_next()
+        };
+
+        match next_token {
+            Some(token) => {
+                self.count += 1;
+                self.last_token = Some(token);
+
+                if self.eos_token_id.contains(token) {
+                    self.finished = true;
+                }
+
+                Some(token)
+            }
+            None => {
+                self.finished = true;
+                None
+            }
+        }
+    }
+}
+
+impl<'a, B: Backend> TokenStream<'a, B> {
+    fn step_first(&mut self) -> Option<u32> {
+        let input_ints: Vec<i64> = self.input_ids.iter().map(|&x| x as i64).collect();
+        let input_tensor =
+            Tensor::<B, 1, Int>::from_ints(input_ints.as_slice(), self.device)
+                .unsqueeze::<2>();
+
+        let (logits, new_cache) = self.model.forward_with_cache(input_tensor, self.cache.as_ref());
+        self.cache = Some(new_cache);
+
+        Some(sample_last(&logits, self.config.temperature, self.config.top_p))
+    }
+
+    fn step_next(&mut self) -> Option<u32> {
+        let last_token = self.last_token?;
+        let input_ints: Vec<i64> = vec![last_token as i64];
+        let input_tensor =
+            Tensor::<B, 1, Int>::from_ints(input_ints.as_slice(), self.device)
+                .unsqueeze::<2>();
+
+        let (logits, new_cache) = self.model.forward_with_cache(input_tensor, self.cache.as_ref());
+        self.cache = Some(new_cache);
+
+        let vocab_size = self.model.config.vocab_size;
+        let next_token_logits = logits
+            .slice([0..1, 0..1, 0..vocab_size])
+            .reshape([vocab_size]);
+
+        let data = next_token_logits.to_data();
+        let values: Vec<f32> = data.as_slice().unwrap().to_vec();
+        Some(self.sampler.sample(&values, self.config.temperature, self.config.top_p))
+    }
+}
+
+pub fn generate_tokens<B: Backend>(
+    model: &LlamaForCausalLM<B>,
+    input_ids: &[u32],
+    config: &GenerationConfig,
+    eos_token_id: &EosTokenId,
+    device: &B::Device,
+) -> Vec<u32> {
+    let stream = TokenStream::new(model, input_ids, config, eos_token_id, device);
+    stream.collect()
+}

crates/minicpm-inference/src/lib.rs

@@ -1,305 +1,18 @@
+pub mod config;
+pub mod generator;
+pub mod loader;
+pub mod sampling;
+pub mod tokenizer;
+
+pub use config::GenerationConfig;
+pub use generator::{TokenStream};
 pub use minicpm_core::config::{EosTokenId, LlamaConfig};
 pub use minicpm_core::model::{LlamaForCausalLM, LlamaKVCache};
+pub use tokenizer::TokenizerWrapper;
 
-use burn::module::Module;
-use burn::record::{FullPrecisionSettings, NamedMpkFileRecorder};
 use burn::tensor::backend::Backend;
-use burn::tensor::{Float, Int, Tensor};
-use rand::Rng;
-use tokenizers::Tokenizer;
-
-
-pub struct TokenizerWrapper {
-    tokenizer: Tokenizer,
-}
-
-impl TokenizerWrapper {
-    pub fn from_file(path: &str) -> anyhow::Result<Self> {
-        let tokenizer = Tokenizer::from_file(path)
-            .map_err(|e| anyhow::anyhow!("Failed to load tokenizer: {}", e))?;
-        Ok(Self { tokenizer })
-    }
-
-    pub fn encode(&self, text: &str, add_special_tokens: bool) -> anyhow::Result<Vec<u32>> {
-        let encoding = self
-            .tokenizer
-            .encode(text, add_special_tokens)
-            .map_err(|e| anyhow::anyhow!("Failed to encode: {}", e))?;
-        Ok(encoding.get_ids().to_vec())
-    }
-
-    pub fn decode(&self, ids: &[u32], skip_special_tokens: bool) -> anyhow::Result<String> {
-        let text = self
-            .tokenizer
-            .decode(ids, skip_special_tokens)
-            .map_err(|e| anyhow::anyhow!("Failed to decode: {}", e))?;
-        Ok(text)
-    }
-
-    pub fn vocab_size(&self) -> usize {
-        self.tokenizer.get_vocab_size(true)
-    }
-
-    /// 应用 MiniCPM5 chat template
-    pub fn apply_chat_template(&self, user_msg: &str, enable_thinking: bool) -> anyhow::Result<Vec<u32>> {
-        let prompt = if enable_thinking {
-            format!(
-                "<s><|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n<think>\n",
-                user_msg
-            )
-        } else {
-            format!(
-                "<s><|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n",
-                user_msg
-            )
-        };
-        self.encode(&prompt, false)
-    }
-}
-
-
-pub struct GenerationConfig {
-    pub max_new_tokens: Option<usize>,
-    pub temperature: f32,
-    pub top_p: f32,
-}
-
-impl Default for GenerationConfig {
-    fn default() -> Self {
-        Self {
-            max_new_tokens: None,
-            temperature: 1.0,
-            top_p: 1.0,
-        }
-    }
-}
-
-pub fn generate_with_cache<B: Backend>(
-    model: &LlamaForCausalLM<B>,
-    input_ids: &[u32],
-    config: &GenerationConfig,
-    eos_token_id: &EosTokenId,
-    device: &B::Device,
-) -> Vec<u32> {
-    let mut output_ids = input_ids.to_vec();
-    let mut cache: Option<LlamaKVCache<B>> = None;
-
-    // 第一步：输入完整 prompt，建立初始 cache
-    {
-        let input_ints: Vec<i64> = input_ids.iter().map(|&x| x as i64).collect();
-        let input_tensor =
-            Tensor::<B, 1, Int>::from_ints(input_ints.as_slice(), device)
-                .unsqueeze::<2>();
-
-        let (logits, new_cache) = model.forward_with_cache(input_tensor, cache.as_ref());
-        cache = Some(new_cache);
-
-        let next_token = sample_last(&logits, config.temperature, config.top_p);
-        output_ids.push(next_token);
-
-        if eos_token_id.contains(next_token) {
-            return output_ids;
-        }
-    }
-
-    // 后续步骤：每次只输入 1 个 token，使用 cache
-    let mut count = 1;
-    loop {
-        if let Some(max) = config.max_new_tokens {
-            if count >= max {
-                break;
-            }
-        }
-
-        let last_token = output_ids[output_ids.len() - 1];
-        let input_ints: Vec<i64> = vec![last_token as i64];
-        let input_tensor =
-            Tensor::<B, 1, Int>::from_ints(input_ints.as_slice(), device)
-                .unsqueeze::<2>();
-
-        let (logits, new_cache) = model.forward_with_cache(input_tensor, cache.as_ref());
-        cache = Some(new_cache);
-
-        let vocab_size = model.config.vocab_size;
-        let next_token_logits = logits
-            .clone()
-            .slice([0..1, 0..1, 0..vocab_size])
-            .reshape([vocab_size]);
-
-        let next_token = sample(&next_token_logits, config.temperature, config.top_p);
-        output_ids.push(next_token);
-
-        if eos_token_id.contains(next_token) {
-            break;
-        }
-
-        count += 1;
-    }
-
-    output_ids
-}
-
-/// 流式生成：每生成一个新 token，立即调用 `on_token` 回调输出该 token 的解码文本。
-/// 返回完整的 output token IDs。
-pub fn generate_stream<B: Backend>(
-    model: &LlamaForCausalLM<B>,
-    tokenizer: &TokenizerWrapper,
-    input_ids: &[u32],
-    config: &GenerationConfig,
-    eos_token_id: &EosTokenId,
-    device: &B::Device,
-    mut on_token: impl FnMut(&str),
-) -> Vec<u32> {
-    let mut output_ids = input_ids.to_vec();
-    let mut cache: Option<LlamaKVCache<B>> = None;
-
-    // 第一步：完整 prompt 输入
-    {
-        let input_ints: Vec<i64> = input_ids.iter().map(|&x| x as i64).collect();
-        let input_tensor =
-            Tensor::<B, 1, Int>::from_ints(input_ints.as_slice(), device)
-                .unsqueeze::<2>();
-
-        let (logits, new_cache) = model.forward_with_cache(input_tensor, cache.as_ref());
-        cache = Some(new_cache);
-
-        let next_token = sample_last(&logits, config.temperature, config.top_p);
-        output_ids.push(next_token);
-
-        // 流式输出第一个 token
-        if let Ok(text) = tokenizer.decode(&[next_token], true) {
-            on_token(&text);
-        }
-
-        if eos_token_id.contains(next_token) {
-            return output_ids;
-        }
-    }
-
-    // 后续步骤：每步生成一个 token
-    let mut count = 1;
-    loop {
-        if let Some(max) = config.max_new_tokens {
-            if count >= max {
-                break;
-            }
-        }
-
-        let last_token = output_ids[output_ids.len() - 1];
-        let input_ints: Vec<i64> = vec![last_token as i64];
-        let input_tensor =
-            Tensor::<B, 1, Int>::from_ints(input_ints.as_slice(), device)
-                .unsqueeze::<2>();
-
-        let (logits, new_cache) = model.forward_with_cache(input_tensor, cache.as_ref());
-        cache = Some(new_cache);
-
-        let vocab_size = model.config.vocab_size;
-        let next_token_logits = logits
-            .clone()
-            .slice([0..1, 0..1, 0..vocab_size])
-            .reshape([vocab_size]);
-
-        let next_token = sample(&next_token_logits, config.temperature, config.top_p);
-        output_ids.push(next_token);
-
-        // 流式输出新 token
-        if let Ok(text) = tokenizer.decode(&[next_token], true) {
-            on_token(&text);
-        }
-
-        if eos_token_id.contains(next_token) {
-            break;
-        }
-
-        count += 1;
-    }
-
-    output_ids
-}
-
-fn sample_last<B: Backend>(logits: &Tensor<B, 3, Float>, temperature: f32, top_p: f32) -> u32 {
-    let shape = logits.shape();
-    let dims: [usize; 3] = shape.dims();
-    let seq_len = dims[1];
-    let vocab_size = dims[2];
-
-    let last_logits = logits
-        .clone()
-        .slice([0..1, seq_len - 1..seq_len, 0..vocab_size])
-        .reshape([vocab_size]);
-
-    sample(&last_logits, temperature, top_p)
-}
-
-fn sample<B: Backend>(logits: &Tensor<B, 1, Float>, temperature: f32, top_p: f32) -> u32 {
-    let data = logits.clone().to_data();
-    let values: Vec<f32> = data.as_slice().unwrap().to_vec();
-
-    // temperature = 0 时退化为 greedy
-    if temperature <= 0.001 {
-        let mut max_idx = 0;
-        let mut max_val = f32::NEG_INFINITY;
-        for (i, &val) in values.iter().enumerate() {
-            if val > max_val {
-                max_val = val;
-                max_idx = i;
-            }
-        }
-        return max_idx as u32;
-    }
-
-    // 应用 temperature
-    let scaled: Vec<f32> = values.iter().map(|&v| v / temperature).collect();
-
-    // softmax
-    let max_val = scaled.iter().cloned().fold(f32::NEG_INFINITY, f32::max);
-    let exp_sum: f32 = scaled.iter().map(|&v| (v - max_val).exp()).sum();
-    let mut probs: Vec<f32> = scaled.iter().map(|&v| (v - max_val).exp() / exp_sum).collect();
-
-    // top_p 过滤
-    if top_p < 1.0 {
-        let mut indexed: Vec<(usize, f32)> = probs.iter().enumerate().map(|(i, &p)| (i, p)).collect();
-        indexed.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap());
-
-        let mut cumsum = 0.0;
-        let mut cutoff = indexed.len();
-        for (i, &(_, p)) in indexed.iter().enumerate() {
-            cumsum += p;
-            if cumsum > top_p {
-                cutoff = i + 1;
-                break;
-            }
-        }
-
-        let keep: std::collections::HashSet<usize> = indexed[..cutoff].iter().map(|(i, _)| *i).collect();
-        for (i, p) in probs.iter_mut().enumerate() {
-            if !keep.contains(&i) {
-                *p = 0.0;
-            }
-        }
-
-        let sum: f32 = probs.iter().sum();
-        if sum > 0.0 {
-            for p in probs.iter_mut() {
-                *p /= sum;
-            }
-        }
-    }
-
-    // 多项式采样
-    let mut rng = rand::thread_rng();
-    let r: f32 = rng.gen();
-    let mut cumsum = 0.0;
-    for (i, &p) in probs.iter().enumerate() {
-        cumsum += p;
-        if r < cumsum {
-            return i as u32;
-        }
-    }
-    probs.len() as u32 - 1
-}
 
+use loader::load_model;
 
 pub struct MiniCPM<B: Backend> {
     model: LlamaForCausalLM<B>,
@@ -317,10 +30,7 @@ impl<B: Backend> MiniCPM<B> {
     ) -> anyhow::Result<Self> {
         let config = LlamaConfig::from_json(config_path)?;
         let tokenizer = TokenizerWrapper::from_file(tokenizer_path)?;
-
-        let model = LlamaForCausalLM::<B>::new(config.clone(), device);
-        let recorder = NamedMpkFileRecorder::<FullPrecisionSettings>::new();
-        let model = model.load_file(model_path, &recorder, device)?;
+        let model = load_model::<B>(model_path, &config, device)?;
 
         Ok(Self {
             model,
@@ -337,7 +47,7 @@ impl<B: Backend> MiniCPM<B> {
         config: &GenerationConfig,
     ) -> anyhow::Result<String> {
         let input_ids = self.tokenizer.apply_chat_template(prompt, think)?;
-        let output_ids = generate_with_cache(
+        let output_ids = generator::generate_tokens(
             &self.model,
             &input_ids,
             config,
@@ -348,27 +58,26 @@ impl<B: Backend> MiniCPM<B> {
         self.tokenizer.decode(new_ids, true)
     }
 
-    /// 流式生成：每生成一个 token 立即调用 `on_token` 回调，
-    /// 参数为该 token 解码后的文本。
-    pub fn generate_stream(
-        &self,
+    pub fn generate_stream<'a>(
+        &'a self,
         prompt: &str,
         think: bool,
-        config: &GenerationConfig,
-        on_token: impl FnMut(&str),
-    ) -> anyhow::Result<String> {
+        config: &'a GenerationConfig,
+    ) -> anyhow::Result<TextStream<'a, B>> {
         let input_ids = self.tokenizer.apply_chat_template(prompt, think)?;
-        let output_ids = generate_stream(
+        let token_stream = TokenStream::new(
             &self.model,
-            &self.tokenizer,
             &input_ids,
             config,
             &self.config.eos_token_id,
             &self.device,
-            on_token,
         );
-        let new_ids = &output_ids[input_ids.len()..];
-        self.tokenizer.decode(new_ids, true)
+
+        Ok(TextStream {
+            token_stream,
+            tokenizer: &self.tokenizer,
+            buffer: Vec::new(),
+        })
     }
 
     pub fn config(&self) -> &LlamaConfig {
@@ -383,3 +92,32 @@ impl<B: Backend> MiniCPM<B> {
         &self.model
     }
 }
+
+pub struct TextStream<'a, B: Backend> {
+    token_stream: TokenStream<'a, B>,
+    tokenizer: &'a TokenizerWrapper,
+    buffer: Vec<u32>,
+}
+
+impl<'a, B: Backend> Iterator for TextStream<'a, B> {
+    type Item = String;
+
+    fn next(&mut self) -> Option<Self::Item> {
+        match self.token_stream.next() {
+            Some(token) => {
+                self.buffer.push(token);
+                match self.tokenizer.decode(&[token], true) {
+                    Ok(text) => Some(text),
+                    Err(_) => Some(String::new()),
+                }
+            }
+            None => None,
+        }
+    }
+}
+
+impl<'a, B: Backend> TextStream<'a, B> {
+    pub fn all_tokens(&self) -> &[u32] {
+        &self.buffer
+    }
+}

crates/minicpm-inference/src/loader/mod.rs

@@ -0,0 +1,17 @@
+use burn::module::Module;
+use burn::record::{FullPrecisionSettings, NamedMpkFileRecorder};
+use burn::tensor::backend::Backend;
+
+use minicpm_core::config::LlamaConfig;
+use minicpm_core::model::LlamaForCausalLM;
+
+pub fn load_model<B: Backend>(
+    model_path: &str,
+    config: &LlamaConfig,
+    device: &B::Device,
+) -> anyhow::Result<LlamaForCausalLM<B>> {
+    let model = LlamaForCausalLM::<B>::new(config.clone(), device);
+    let recorder = NamedMpkFileRecorder::<FullPrecisionSettings>::new();
+    let model = model.load_file(model_path, &recorder, device)?;
+    Ok(model)
+}

crates/minicpm-inference/src/sampling.rs

@@ -0,0 +1,163 @@
+use burn::tensor::backend::Backend;
+use burn::tensor::{Float, Tensor};
+use rand::{rngs::ThreadRng, Rng};
+
+struct Sampler {
+    rng: ThreadRng,
+}
+
+impl Sampler {
+    fn new() -> Self {
+        Self {
+            rng: rand::thread_rng(),
+        }
+    }
+
+    fn sample(&mut self, logits: &[f32], temperature: f32, top_p: f32) -> u32 {
+        if temperature <= 0.001 {
+            let mut max_idx = 0;
+            let mut max_val = f32::NEG_INFINITY;
+            for (i, &val) in logits.iter().enumerate() {
+                if val > max_val {
+                    max_val = val;
+                    max_idx = i;
+                }
+            }
+            return max_idx as u32;
+        }
+
+        let max_val = logits.iter().cloned().fold(f32::NEG_INFINITY, f32::max);
+        let exp_sum: f32 = logits.iter().map(|&v| (v - max_val).exp()).sum();
+        let mut probs: Vec<f32> = logits.iter().map(|&v| (v - max_val).exp() / exp_sum).collect();
+
+        if top_p < 1.0 {
+            let mut indexed: Vec<(usize, f32)> = probs.iter().enumerate().map(|(i, &p)| (i, p)).collect();
+            indexed.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap());
+
+            let mut cumsum = 0.0;
+            let mut cutoff = indexed.len();
+            for (i, &(_, p)) in indexed.iter().enumerate() {
+                cumsum += p;
+                if cumsum > top_p {
+                    cutoff = i + 1;
+                    break;
+                }
+            }
+
+            let keep: std::collections::HashSet<usize> = indexed[..cutoff].iter().map(|(i, _)| *i).collect();
+            for (i, p) in probs.iter_mut().enumerate() {
+                if !keep.contains(&i) {
+                    *p = 0.0;
+                }
+            }
+
+            let sum: f32 = probs.iter().sum();
+            if sum > 0.0 {
+                for p in probs.iter_mut() {
+                    *p /= sum;
+                }
+            }
+        }
+
+        let r: f32 = self.rng.gen();
+        let mut cumsum = 0.0;
+        for (i, &p) in probs.iter().enumerate() {
+            cumsum += p;
+            if r < cumsum {
+                return i as u32;
+            }
+        }
+        probs.len() as u32 - 1
+    }
+}
+
+pub struct CachedSampler {
+    sampler: std::cell::RefCell<Sampler>,
+}
+
+impl CachedSampler {
+    pub fn new() -> Self {
+        Self {
+            sampler: std::cell::RefCell::new(Sampler::new()),
+        }
+    }
+
+    pub fn sample(&self, logits: &[f32], temperature: f32, top_p: f32) -> u32 {
+        self.sampler.borrow_mut().sample(logits, temperature, top_p)
+    }
+}
+
+pub fn sample_last<B: Backend>(logits: &Tensor<B, 3, Float>, temperature: f32, top_p: f32) -> u32 {
+    let shape = logits.shape();
+    let dims: [usize; 3] = shape.dims();
+    let seq_len = dims[1];
+    let vocab_size = dims[2];
+
+    let last_logits = logits
+        .clone()
+        .slice([0..1, seq_len - 1..seq_len, 0..vocab_size])
+        .reshape([vocab_size]);
+
+    sample(&last_logits, temperature, top_p)
+}
+
+pub fn sample<B: Backend>(logits: &Tensor<B, 1, Float>, temperature: f32, top_p: f32) -> u32 {
+    let data = logits.clone().to_data();
+    let values: Vec<f32> = data.as_slice().unwrap().to_vec();
+
+    if temperature <= 0.001 {
+        let mut max_idx = 0;
+        let mut max_val = f32::NEG_INFINITY;
+        for (i, &val) in values.iter().enumerate() {
+            if val > max_val {
+                max_val = val;
+                max_idx = i;
+            }
+        }
+        return max_idx as u32;
+    }
+
+    let max_val = values.iter().cloned().fold(f32::NEG_INFINITY, f32::max);
+    let exp_sum: f32 = values.iter().map(|&v| (v - max_val).exp()).sum();
+    let mut probs: Vec<f32> = values.iter().map(|&v| (v - max_val).exp() / exp_sum).collect();
+
+    if top_p < 1.0 {
+        let mut indexed: Vec<(usize, f32)> = probs.iter().enumerate().map(|(i, &p)| (i, p)).collect();
+        indexed.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap());
+
+        let mut cumsum = 0.0;
+        let mut cutoff = indexed.len();
+        for (i, &(_, p)) in indexed.iter().enumerate() {
+            cumsum += p;
+            if cumsum > top_p {
+                cutoff = i + 1;
+                break;
+            }
+        }
+
+        let keep: std::collections::HashSet<usize> = indexed[..cutoff].iter().map(|(i, _)| *i).collect();
+        for (i, p) in probs.iter_mut().enumerate() {
+            if !keep.contains(&i) {
+                *p = 0.0;
+            }
+        }
+
+        let sum: f32 = probs.iter().sum();
+        if sum > 0.0 {
+            for p in probs.iter_mut() {
+                *p /= sum;
+            }
+        }
+    }
+
+    let mut rng = rand::thread_rng();
+    let r: f32 = rng.gen();
+    let mut cumsum = 0.0;
+    for (i, &p) in probs.iter().enumerate() {
+        cumsum += p;
+        if r < cumsum {
+            return i as u32;
+        }
+    }
+    probs.len() as u32 - 1
+}

crates/minicpm-inference/src/tokenizer.rs

@@ -0,0 +1,48 @@
+use tokenizers::Tokenizer;
+
+pub struct TokenizerWrapper {
+    tokenizer: Tokenizer,
+}
+
+impl TokenizerWrapper {
+    pub fn from_file(path: &str) -> anyhow::Result<Self> {
+        let tokenizer = Tokenizer::from_file(path)
+            .map_err(|e| anyhow::anyhow!("加载 tokenizer 失败: {}", e))?;
+        Ok(Self { tokenizer })
+    }
+
+    pub fn encode(&self, text: &str, add_special_tokens: bool) -> anyhow::Result<Vec<u32>> {
+        let encoding = self
+            .tokenizer
+            .encode(text, add_special_tokens)
+            .map_err(|e| anyhow::anyhow!("编码失败: {}", e))?;
+        Ok(encoding.get_ids().to_vec())
+    }
+
+    pub fn decode(&self, ids: &[u32], skip_special_tokens: bool) -> anyhow::Result<String> {
+        let text = self
+            .tokenizer
+            .decode(ids, skip_special_tokens)
+            .map_err(|e| anyhow::anyhow!("解码失败: {}", e))?;
+        Ok(text)
+    }
+
+    pub fn vocab_size(&self) -> usize {
+        self.tokenizer.get_vocab_size(true)
+    }
+
+    pub fn apply_chat_template(&self, user_msg: &str, enable_thinking: bool) -> anyhow::Result<Vec<u32>> {
+        let prompt = if enable_thinking {
+            format!(
+                "<s><|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n<think>\n",
+                user_msg
+            )
+        } else {
+            format!(
+                "<s><|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n",
+                user_msg
+            )
+        };
+        self.encode(&prompt, false)
+    }
+}

examples/convert/Cargo.toml

@@ -3,7 +3,11 @@ name = "convert"
 version = "0.1.0"
 edition = "2021"
 
+[[bin]]
+name = "convert"
+path = "src/main.rs"
+
 [dependencies]
 minicpm-convert = { path = "../../crates/minicpm-convert" }
-burn = { version = "0.21", features = ["std", "wgpu"] }
+burn = { version = "0.21", default-features = false, features = ["std", "wgpu", "fusion", "autotune"] }
 anyhow = "1.0"

examples/convert/src/main.rs

@@ -1,24 +1,20 @@
-use burn::backend::Wgpu;
-use minicpm_convert::export_model;
+use burn::backend::{wgpu::WgpuDevice, Wgpu};
+use minicpm_convert::{run_export, ExportTask, Format};
 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,
-    )?;
+    println!("模型转换工具 (Wgpu GPU 加速)");
+    println!("源文件: {:?}", safetensors_path);
+    println!();
 
-    println!("模型转换完成！");
+    let device = WgpuDevice::default();
+    let tasks = vec![ExportTask::new(Format::Full)];
+
+    run_export::<Wgpu>(safetensors_path, config_path, tokenizer_path, &tasks, &device)?;
 
     Ok(())
 }

examples/minimal-inference/src/main.rs

@@ -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(())
-}

examples/wgpu-backend/Cargo.toml

@@ -1,9 +1,9 @@
 [package]
-name = "minimal-inference"
+name = "wgpu-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", "wgpu", "fusion", "autotune"] }
 anyhow = "1.0"

examples/wgpu-backend/src/main.rs

@@ -0,0 +1,50 @@
+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 think = args.iter().any(|s| s == "--think");
+
+    println!("========================================");
+    println!("后端: Wgpu (GPU 加速)");
+    println!("模型目录: model");
+    if think {
+        println!("思考模式: 开启");
+    } else {
+        println!("思考模式: 关闭");
+    }
+    println!("========================================");
+
+    let start = Instant::now();
+    let device = WgpuDevice::default();
+
+    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 = "需要处理生成一个河南的旅游计划，要求如下：\n1. 旅游计划为三天两夜的行程安排。\n2. 每天的行程安排包括景点、餐饮和住宿建议。\n3. 景点安排要考虑交通便利性和时间合理性。\n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。\n";
+    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, think, &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(())
+}

readme.md

@@ -1,186 +1,126 @@
-# MiniCPM5-1B-rust
+# MiniCPM5-1B-rust (v0.2.0)
 
-基于 [Burn](https://burn.dev/) 深度学习框架实现的 MiniCPM5-1B 大语言模型推理库，纯 Rust 编写，支持 GPU 加速。
+MiniCPM5-1B 模型的 Rust 推理实现，基于 [Burn](https://github.com/tracel-ai/burn) 框架。
 
-## 模型架构参数
+## 特性
 
-| 参数 | 值 |
-|------|-----|
-| 层数 | 24 |
-| 注意力机制 | GQA (16 Query / 2 KV) |
-| head_dim | 128 |
-| hidden_size | 1536 |
-| intermediate_size | 8960 |
-| vocab_size | 151936 |
-| RoPE theta | 5000000 |
-| 归一化 | RMSNorm |
-| 激活函数 | SiLU (SwiGLU) |
-| KV Cache | 支持 |
+- 🔥 **Wgpu GPU 加速**：支持 CUDA / Metal / Vulkan / DX12 等多种 GPU 后端
+- 🚀 **算子融合**：启用 fusion + autotune 优化，自动选择最优 kernel
+- 📝 **流式输出**：实时生成文本，支持逐字输出
+- 🧠 **思考模式**：支持开启模型的推理思考过程
+- 📦 **纯 Rust**：无 C/C++ 依赖，跨平台编译简单
 
-## Workspace 结构
+## 项目结构
 
-项目采用 Cargo workspace 多成员结构，包含三个 crate：
-
-### minicpm-core
-
-核心模型定义，包含：
-
-- `LlamaConfig` — 模型配置（从 `config.json` 加载）
-- `LlamaForCausalLM` — 因果语言模型主体
-- `LlamaKVCache` — KV 缓存结构
-- `EosTokenId` — EOS token 标识（支持单个或多个）
-- 各模块实现：Attention、Decoder、FFN、RMSNorm、RoPE
-
-### minicpm-convert
-
-模型格式转换工具，负责将 HuggingFace safetensors 格式的权重转换为 Burn MPK 格式：
-
-- `export_model()` — 完整转换流程（加载 safetensors → 构建模型 → 导出 MPK）
-- 支持 BF16 / F16 / F32 精度输入
-- 支持 `tie_word_embeddings` 权重共享
-
-### minicpm-inference
-
-推理功能封装，提供高层 API：
-
-- `MiniCPM` — 高层封装，整合模型 + tokenizer + 推理逻辑
-- `TokenizerWrapper` — tokenizer 封装，支持 MiniCPM5 chat template
-- `GenerationConfig` — 生成配置（max_new_tokens、temperature、top_p）
-- `generate_with_cache()` — 带 KV Cache 的自回归生成
-- 支持 temperature 采样、top-p 采样、greedy 解码
+```
+MiniCPM5-1B-rust/
+├── crates/
+│   ├── minicpm-core/        # 核心模型实现（Transformer、Attention、FFN）
+│   ├── minicpm-inference/   # 推理引擎（生成、采样、tokenizer）
+│   └── minicpm-convert/     # 模型转换库
+└── examples/
+    ├── convert/             # 模型转换示例（safetensors → Burn 格式）
+    └── wgpu-backend/        # Wgpu GPU 推理示例
+```
 
 ## 快速开始
 
-### 1. 转换模型
+### 1. 准备模型
 
-首先将 HuggingFace 格式的 MiniCPM5-1B 模型转换为 Burn MPK 格式：
-
-```rust
-use minicpm_convert::export_model;
-use burn::backend::Wgpu;
-use std::path::Path;
-
-fn main() -> anyhow::Result<()> {
-    let device = Default::default();
-
-    export_model::<Wgpu>(
-        Path::new("MiniCPM5-1B/model.safetensors"),
-        Path::new("MiniCPM5-1B/config.json"),
-        Path::new("MiniCPM5-1B-burn"),
-        &device,
-    )?;
-
-    Ok(())
-}
-```
-
-转换完成后，`MiniCPM5-1B-burn/` 目录下会生成 `model.mp` 文件。
-
-### 2. 推理使用
-
-使用 `MiniCPM` 高层 API 进行文本生成：
-
-```rust
-use minicpm_inference::{MiniCPM, GenerationConfig};
-use burn::backend::Wgpu;
-
-fn main() -> anyhow::Result<()> {
-    let device = Default::default();
-
-    // 加载模型
-    let model = MiniCPM::<Wgpu>::load(
-        "MiniCPM5-1B-burn/model",
-        "MiniCPM5-1B/config.json",
-        "MiniCPM5-1B/tokenizer.json",
-        &device,
-    )?;
-
-    // 生成配置
-    let gen_config = GenerationConfig {
-        max_new_tokens: Some(512),
-        temperature: 0.7,
-        top_p: 0.8,
-    };
-
-    // 生成回答（think = true 启用思考模式）
-    let response = model.generate("用 Rust 写一个 Hello World", true, &gen_config)?;
-
-    println!("{}", response);
-
-    Ok(())
-}
-```
-
-## 移植到其他项目
-
-### 依赖配置
-
-在你的 `Cargo.toml` 中添加：
-
-```toml
-[dependencies]
-minicpm-inference = { path = "../MiniCPM5-1B-rust/crates/minicpm-inference" }
-burn = { version = "0.21", features = ["std", "wgpu"] }
-anyhow = "1.0"
-```
-
-> 也可以根据需要选择其他 backend（如 `tch-gpu`、`cuda` 等）。
-
-### 最小示例
-
-```rust
-use minicpm_inference::{MiniCPM, GenerationConfig};
-use burn::backend::Wgpu;
-
-fn main() -> anyhow::Result<()> {
-    let device = Default::default();
-
-    let model = MiniCPM::<Wgpu>::load(
-        "model",
-        "config.json",
-        "tokenizer.json",
-        &device,
-    )?;
-
-    let config = GenerationConfig {
-        max_new_tokens: Some(256),
-        temperature: 1.0,
-        top_p: 1.0,
-    };
-
-    let output = model.generate("你好", false, &config)?;
-    println!("{}", output);
-
-    Ok(())
-}
-```
-
-## 模型文件准备
-
-`MiniCPM5-1B/` 目录需要包含以下文件：
+将 MiniCPM5-1B safetensors 模型放到 `MiniCPM5-1B/` 目录：
 
 ```
 MiniCPM5-1B/
-├── config.json        # 模型配置文件
-├── model.safetensors  # safetensors 格式权重
-└── tokenizer.json     # tokenizer 配置
+├── model-00000-of-00001.safetensors
+├── config.json
+└── tokenizer.json
 ```
 
-转换后 Burn 模型目录结构：
+### 2. 转换模型
 
-```
-MiniCPM5-1B-burn/
-└── model.mp           # Burn MPK 格式权重
+```bash
+cargo run --release -p convert
 ```
 
-## 性能说明
+转换完成后会生成 `model/` 目录（全精度 f32，Burn mpk 格式）。
 
-| Backend | 设备 | 速度 | 显存占用 |
-|---------|------|------|----------|
-| WGPU (Vulkan) | RTX 4060 | ~10 tokens/s | ~4 GB (F32) |
+### 3. 运行推理
 
-> 以上数据仅供参考，实际性能因硬件配置、生成长度等因素而异。
+```bash
+# 普通模式
+cargo run --release -p wgpu-backend
 
-## 许可证
+# 开启思考模式
+cargo run --release -p wgpu-backend -- --think
+```
+
+## 使用说明
+
+### wgpu-backend 命令行参数
+
+| 参数 | 说明 |
+|------|------|
+| `--think` | 开启思考模式，模型会在回答前进行推理思考 |
+
+### GenerationConfig
+
+```rust
+GenerationConfig {
+    max_new_tokens: Some(1200), // 最大生成 token 数
+    temperature: 0.7,            // 温度，越高越随机
+    top_p: 0.95,                 // nucleus sampling 参数
+}
+```
+
+### 作为库使用
+
+```rust
+use burn::backend::{wgpu::WgpuDevice, Wgpu};
+use minicpm_inference::{GenerationConfig, MiniCPM};
+
+let device = WgpuDevice::default();
+let model = MiniCPM::<Wgpu>::load(
+    "model/model",
+    "config.json",
+    "tokenizer.json",
+    &device,
+)?;
+
+let config = GenerationConfig {
+    max_new_tokens: Some(500),
+    temperature: 0.7,
+    top_p: 0.95,
+};
+
+// 流式生成
+let stream = model.generate_stream("你好", false, &config)?;
+for text in stream {
+    print!("{}", text);
+}
+```
+
+## 发布构建
+
+```bash
+# Release 构建（启用 LTO + 最高优化级别）
+cargo build --release
+
+# 发布到 gitea registry
+cargo publish -p minicpm-core --registry gitea
+cargo publish -p minicpm-inference --registry gitea
+cargo publish -p minicpm-convert --registry gitea
+```
+
+Release profile 配置：
+- `opt-level = 3`：最高优化级别
+- `lto = true`：链接时优化
+- `codegen-units = 1`：单代码生成单元，最大化优化
+
+## 依赖
+
+- Rust 1.75+
+- 支持 Wgpu 的 GPU（CUDA / Metal / Vulkan / DX12）
+
+## License
 
 MIT