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
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145

#[derive(Debug, Clone, PartialEq, serde::Deserialize)]
pub struct DeepseekV2Config {
    pub bos_token_id: u32,
    pub eos_token_id: u32,
    pub first_k_dense_replace: usize,
    pub hidden_size: usize,
    pub intermediate_size: usize,
    pub kv_lora_rank: Option<usize>,
    pub lm_head: bool,
    pub max_position_embeddings: usize,
    pub moe_intermediate_size: usize,
    #[serde(default = "default_moe_layer_freq")]
    pub moe_layer_freq: usize,
    #[serde(default = "default_routed_scaling_factor")]
    pub routed_scaling_factor: f64,
    #[serde(default = "default_scoring_func")]
    pub scoring_func: String,
    #[serde(default = "default_aux_loss_alpha")]
    pub aux_loss_alpha: f32,
    #[serde(default = "default_true")]
    pub seq_aux: bool,
    #[serde(default = "default_false")]
    pub norm_topk_prob: bool,
    pub n_group: usize,
    pub n_routed_experts: usize,
    pub n_shared_experts: usize,
    pub num_attention_heads: usize,
    pub num_experts_per_tok: usize,
    pub num_hidden_layers: usize,
    pub num_key_value_heads: usize,
    pub q_lora_rank: Option<usize>,
    pub qk_nope_head_dim: usize,
    pub qk_rope_head_dim: usize,
    pub rm_head: bool,
    pub topk_group: usize,
    pub topk_method: String,
    pub torch_dtype: String,
    pub use_mla: bool,
    pub v_head_dim: usize,
    pub vocab_size: usize,
    #[serde(default = "default_rms_norm_eps")]
    pub rms_norm_eps: f64,
}

fn default_moe_layer_freq() -> usize {
    1
}
fn default_routed_scaling_factor() -> f64 {
    1.0
}
fn default_scoring_func() -> String {
    "softmax".to_string()
}
fn default_aux_loss_alpha() -> f32 {
    0.001
}
fn default_true() -> bool {
    true
}
fn default_false() -> bool {
    false
}
fn default_rms_norm_eps() -> f64 {
    1e-6
}

#[derive(Debug, Clone, PartialEq, serde::Deserialize)]
pub struct ProjectorConfig {
    pub input_dim: usize,
    pub model_type: String,
    pub n_embed: usize,
    pub projector_type: String,
}

#[derive(Debug, Clone, PartialEq, serde::Deserialize)]
pub struct ClipL14_224 {
    pub heads: usize,
    pub image_size: usize,
    pub layers: usize,
    pub patch_size: usize,
    pub width: usize,
}

#[derive(Debug, Clone, PartialEq, serde::Deserialize)]
pub struct SamVitB {
    pub downsample_channels: Vec<usize>,
    pub global_attn_indexes: Vec<usize>,
    pub heads: usize,
    pub layers: usize,
    pub width: usize,
}

#[derive(Debug, Clone, PartialEq, serde::Deserialize)]
pub struct Qwen2_0_5B {
    dim: usize,
}

#[derive(Debug, Clone, PartialEq, serde::Deserialize)]
pub struct Width {
    #[serde(rename = "clip-l-14-224")]
    pub clip_l_14_224: Option<ClipL14_224>,
    #[serde(rename = "qwen2-0-5b")]
    pub qwen2_0_5b: Option<Qwen2_0_5B>,
    pub sam_vit_b: SamVitB,
}

#[derive(Debug, Clone, PartialEq, serde::Deserialize)]
pub struct DeepseekOCRVisionConfig {
    pub image_size: usize,
    pub mlp_ratio: f32,
    pub width: Width,
}

#[derive(Debug, Clone, PartialEq, serde::Deserialize)]
pub struct DeepseekOCRConfig {
    pub language_config: DeepseekV2Config,
    pub projector_config: ProjectorConfig,
    pub torch_dtype: String,
    pub vision_config: DeepseekOCRVisionConfig,
    pub bos_token_id: u32,
    pub eos_token_id: u32,
    pub first_k_dense_replace: u32,
    pub hidden_size: usize,
    pub intermediate_size: usize,
    pub kv_lora_rank: Option<usize>,
    pub lm_head: bool,
    pub max_position_embeddings: usize,
    pub moe_intermediate_size: usize,
    pub n_group: usize,
    pub n_routed_experts: usize,
    pub n_shared_experts: usize,
    pub num_attention_heads: usize,
    pub num_experts_per_tok: usize,
    pub num_hidden_layers: usize,
    pub num_key_value_heads: usize,
    pub q_lora_rank: Option<usize>,
    pub qk_nope_head_dim: usize,
    pub qk_rope_head_dim: usize,
    pub rm_head: bool,
    pub topk_group: usize,
    pub topk_method: String,
    pub use_mla: bool,
    pub v_head_dim: usize,
    pub vocab_size: usize,
}