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

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 {