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

use crate::models::common::MultiModalData;
use crate::models::common::generate::{
    GenerationContext, generate_generic, generate_stream_generic,
};
use crate::params::chat::{ChatCompletionParameters, ChatCompletionResponse};
use crate::{
    chat_template::ChatTemplate,
    models::{
        GenerateModel,
        lfm2::{
            config::{Lfm2Config, Lfm2GenerateConfig},
            model::Lfm2Model,
        },
    },
    tokenizer::TokenizerModel,
    utils::{find_type_files, get_device, get_dtype},
};
use anyhow::Result;
use candle_core::{DType, Device};
use candle_nn::VarBuilder;

pub struct Lfm2GenerateModel<'a> {
    chat_template: ChatTemplate<'a>,
    tokenizer: TokenizerModel,
    device: Device,
    model: Lfm2Model,
    model_name: String,
}
impl<'a> Lfm2GenerateModel<'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 device = get_device(device);
        let gen_cfg_path = path.to_string() + "/generation_config.json";
        let gen_cfg: Lfm2GenerateConfig = serde_json::from_slice(&std::fs::read(gen_cfg_path)?)?;
        let cfg_path = path.to_string() + "/config.json";
        let cfg: Lfm2Config = serde_json::from_slice(&std::fs::read(cfg_path)?)?;
        let model_path = find_type_files(path, "safetensors")?;
        let cfg_dtype = if let Some(dtype) = &cfg.dtype {
            dtype.clone()
        } else if let Some(dtype) = &cfg.torch_dtype {
            dtype.clone()
        } else {
            "bfloat16".to_string()
        };
        let dtype = get_dtype(dtype, &cfg_dtype);
        let vb = unsafe { VarBuilder::from_mmaped_safetensors(&model_path, dtype, &device)? };
        let eos_ids = vec![gen_cfg.eos_token_id];
        let model = Lfm2Model::new(vb, &cfg, eos_ids)?;
        let model_name = std::path::Path::new(path)
            .file_name()
            .and_then(|s| s.to_str())
            .unwrap_or("lfm2")
            .to_string();
        Ok(Self {
            chat_template,
            tokenizer,
            device,
            model,
            model_name,
        })
    }
}

impl<'a> GenerateModel for Lfm2GenerateModel<'a> {
    fn generate(&mut self, mes: ChatCompletionParameters) -> Result<ChatCompletionResponse> {
        let mes_render = self.chat_template.apply_chat_template(&mes)?;
        let input_ids = self.tokenizer.text_encode(mes_render, &self.device)?;
        let sample_len = mes.max_tokens.unwrap_or(1024);
        let seed = mes.seed.unwrap_or(34562) as u64;
        let mut ctx = GenerationContext::new(
            mes.temperature,
            mes.top_p,
            None,
            mes.repeat_penalty,
            mes.repeat_last_n,
            seed,
            input_ids.dim(1)?,
            sample_len,
            self.device.clone(),
        );

        let data = MultiModalData::new(vec![]);
        generate_generic(
            &mut self.model,
            &self.tokenizer,
            input_ids,
            data,
            &mut ctx,
            &self.model_name,
        )
    }

    fn generate_stream(
        &mut self,
        mes: ChatCompletionParameters,
    ) -> Result<
        Box<
            dyn rocket::futures::Stream<
                    Item = Result<crate::params::chat::ChatCompletionChunkResponse, anyhow::Error>,
                > + Send
                + Unpin
                + '_,
        >,
    > {
        let mes_render = self.chat_template.apply_chat_template(&mes)?;
        let input_ids = self.tokenizer.text_encode(mes_render, &self.device)?;
        let sample_len = mes.max_tokens.unwrap_or(1024);
        let data = MultiModalData::new(vec![]);
        let seed = mes.seed.unwrap_or(34562) as u64;
        let stream = generate_stream_generic(
            &mut self.model,
            &self.tokenizer,
            input_ids,
            data,
            mes.temperature,
            mes.top_p,
            None,
            mes.repeat_penalty,
            mes.repeat_last_n,
            seed,
            sample_len,
            false,
            &self.device,
            &self.model_name,
        )?;
        Ok(Box::new(Box::pin(stream)))
    }
}