refactor(项目结构) 重构为 workspace 多成员项目

- 将项目拆分为三个 crate：minicpm-core（核心模型）、minicpm-convert（转换功能）、minicpm-inference（推理功能） - 添加两个示例：minimal-inference（最小推理）和 convert（模型转换） - 转换后自动拷贝 config.json 和 tokenizer.json 到 model 目录 - 更新 README 说明 workspace 结构和使用方式
2026-07-01 14:33:28 +08:00
parent 7e0585421f
commit 69b37ced07
29 changed files with 654446 additions and 421 deletions
@@ -0,0 +1,83 @@
+use burn::tensor::backend::Backend;
+use burn::tensor::{Float, Tensor};
+
+#[derive(Debug, Clone)]
+pub struct RoPE {
+    pub dim: usize,
+    pub theta: f64,
+    pub max_seq_len: usize,
+    cos_cache: Vec<f32>,
+    sin_cache: Vec<f32>,
+    half_dim: usize,
+}
+
+impl RoPE {
+    pub fn new(dim: usize, theta: f64, max_seq_len: usize) -> Self {
+        let half_dim = dim / 2;
+
+        let mut inv_freqs = Vec::with_capacity(half_dim);
+        for i in 0..half_dim {
+            let freq = 1.0 / (theta.powf((2.0 * i as f64) / dim as f64));
+            inv_freqs.push(freq as f32);
+        }
+
+        let mut cos_cache = Vec::with_capacity(max_seq_len * half_dim);
+        let mut sin_cache = Vec::with_capacity(max_seq_len * half_dim);
+
+        for pos in 0..max_seq_len {
+            for &freq in &inv_freqs {
+                let angle = pos as f32 * freq;
+                cos_cache.push(angle.cos());
+                sin_cache.push(angle.sin());
+            }
+        }
+
+        Self {
+            dim,
+            theta,
+            max_seq_len,
+            cos_cache,
+            sin_cache,
+            half_dim,
+        }
+    }
+
+    pub fn get_cache(&self, offset: usize, len: usize) -> (Vec<f32>, Vec<f32>) {
+        let cos = self.cos_cache[offset * self.half_dim..(offset + len) * self.half_dim].to_vec();
+        let sin = self.sin_cache[offset * self.half_dim..(offset + len) * self.half_dim].to_vec();
+        (cos, sin)
+    }
+
+    pub fn apply<B: Backend>(
+        &self,
+        x: Tensor<B, 4, Float>,
+        offset: usize,
+    ) -> Tensor<B, 4, Float> {
+        let device = x.device();
+        let shape = x.shape();
+        let dims: [usize; 4] = shape.dims();
+        let batch = dims[0];
+        let heads = dims[1];
+        let seq_len = dims[2];
+        let dim = dims[3];
+        let half_dim = self.half_dim;
+
+        let cos_vals = &self.cos_cache[offset * half_dim..(offset + seq_len) * half_dim];
+        let sin_vals = &self.sin_cache[offset * half_dim..(offset + seq_len) * half_dim];
+
+        let cos = Tensor::<B, 1, Float>::from_floats(cos_vals, &device)
+            .reshape([1, 1, seq_len, half_dim])
+            .expand([batch, heads, seq_len, half_dim]);
+        let sin = Tensor::<B, 1, Float>::from_floats(sin_vals, &device)
+            .reshape([1, 1, seq_len, half_dim])
+            .expand([batch, heads, seq_len, half_dim]);
+
+        let x1 = x.clone().slice([0..batch, 0..heads, 0..seq_len, 0..half_dim]);
+        let x2 = x.slice([0..batch, 0..heads, 0..seq_len, half_dim..dim]);
+
+        let x1_rot = x1.clone() * cos.clone() - x2.clone() * sin.clone();
+        let x2_rot = x2 * cos + x1 * sin;
+
+        Tensor::cat(vec![x1_rot, x2_rot], 3)
+    }
+}