update init

src/models/minicpm4/generate.rs

@@ -27,8 +27,8 @@ pub struct MiniCPMGenerateModel<'a> {
     im_end_id: u32,
 }
 
-impl<'a> GenerateModel for MiniCPMGenerateModel<'a> {
-    fn init(path: &str, device: Option<&Device>, dtype: Option<DType>) -> Result<Self> {
+impl<'a> MiniCPMGenerateModel<'a> {
+    pub fn init(path: &str, device: Option<&Device>, dtype: Option<DType>) -> Result<Self> {
         let chat_template = ChatTemplate::init(path)?;
         let tokenizer = TokenizerModel::init(path)?;
         let config_path = path.to_string() + "/config.json";
@@ -51,6 +51,9 @@ impl<'a> GenerateModel for MiniCPMGenerateModel<'a> {
             im_end_id,
         })
     }
+}
+
+impl<'a> GenerateModel for MiniCPMGenerateModel<'a> {
 
     fn generate(&mut self, mes: ChatCompletionParameters) -> Result<ChatCompletionResponse> {
         let mut logit_processor = get_logit_processor(mes.temperature, mes.top_p);

src/models/mod.rs

@@ -10,9 +10,9 @@ use openai_dive::v1::resources::chat::{
 use rocket::futures::Stream;
 
 pub trait GenerateModel {
-    fn init(path: &str, device: Option<&Device>, dtype: Option<DType>) -> Result<Self>
-    where
-        Self: Sized;
+    // fn init(path: &str, device: Option<&Device>, dtype: Option<DType>) -> Result<Self>
+    // where
+    //     Self: Sized;
     fn generate(&mut self, mes: ChatCompletionParameters) -> Result<ChatCompletionResponse>;
     fn generate_stream(
         &mut self,

src/models/qwen2_5vl/generate.rs

@@ -31,8 +31,8 @@ pub struct Qwen2_5VLGenerateModel<'a> {
     im_end_id: u32,
 }
 
-impl<'a> GenerateModel for Qwen2_5VLGenerateModel<'a> {
-    fn init(path: &str, device: Option<&Device>, dtype: Option<DType>) -> Result<Self> {
+impl<'a> Qwen2_5VLGenerateModel<'a> {
+    pub fn init(path: &str, device: Option<&Device>, dtype: Option<DType>) -> Result<Self> {
         let chat_template = ChatTemplate::init(path)?;
         let tokenizer = TokenizerModel::init(path)?;
         let config_path = path.to_string() + "/config.json";
@@ -57,6 +57,9 @@ impl<'a> GenerateModel for Qwen2_5VLGenerateModel<'a> {
             im_end_id,
         })
     }
+}
+
+impl<'a> GenerateModel for Qwen2_5VLGenerateModel<'a> {
 
     fn generate(&mut self, mes: ChatCompletionParameters) -> Result<ChatCompletionResponse> {
         let mut logit_processor = get_logit_processor(mes.temperature, mes.top_p);

src/models/voxcpm/config.rs

@@ -0,0 +1,56 @@
+use candle_nn::Activation;
+
+#[derive(Debug, Clone, PartialEq, serde::Deserialize)]
+pub struct RopeScalingConfig {
+    pub rope_type: String,
+    pub long_factor: Vec<f32>,
+    pub short_factor: Vec<f32>,
+    pub original_max_position_embeddings: usize,
+}
+
+#[derive(Debug, Clone, PartialEq, serde::Deserialize)]
+pub struct MiniCPM4Config {
+    pub bos_token_id: u32,
+    pub eos_token_id: u32,
+    pub hidden_size: usize,
+    pub intermediate_size: usize,
+    pub max_position_embeddings: usize,
+    pub num_attention_heads: usize,
+    pub num_hidden_layers: usize,
+    pub num_key_value_heads: usize,
+    pub rms_norm_eps: f64,
+    pub rope_scaling: RopeScalingConfig,
+    pub torch_dtype: String,
+    pub vocab_size: usize,
+    // pub use_mup: bool,
+    pub scale_emb:f32,
+    pub dim_model_base: usize,
+    pub scale_depth: f32,
+    // pub rope_theta: f32,
+    // pub kv_channels: i32,
+}
+
+#[derive(Debug, Clone, PartialEq, serde::Deserialize)]
+pub struct VoxCPMEncoderConfig {
+    hidden_dim: usize,
+    ffn_dim: usize,
+    num_heads: usize,
+    num_layers: usize,
+}
+
+#[derive(Debug, Clone, PartialEq, serde::Deserialize)]
+pub struct CfmConfig {
+    sigma_min: f32,
+    solver: String,
+    t_scheduler: String,
+    inference_cfg_rate: f32,
+}
+
+#[derive(Debug, Clone, PartialEq, serde::Deserialize)]
+pub struct VoxCPMDitConfig {
+    hidden_dim: usize,
+    ffn_dim: usize,
+    num_heads: usize,
+    num_layers: usize,
+    cfm_config: CfmConfig,
+}