aha 0.2.5

aha model inference library, now supports Qwen(2.5VL/3/3VL/3.5/ASR/3Embedding/3Reranker), MiniCPM4, VoxCPM/1.5, DeepSeek-OCR/2, Hunyuan-OCR, PaddleOCR-VL/1.5, RMBG2.0, GLM(ASR-Nano-2512/OCR), Fun-ASR-Nano-2512, LFM(2/2.5/2VL/2.5VL)
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29

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: Vec<u32>,
    pub hidden_act: Activation,
    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 scale_emb: f64,
    pub dim_model_base: usize,
    pub scale_depth: f32,
}