aha 0.2.6

aha model inference library, now supports Qwen(2.5VL/3/3VL/3.5/ASR/3Embedding/3Reranker), MiniCPM(4/5), VoxCPM(0.5B/1.5/2), 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

use crate::{
    models::common::{
        MultiModalData,
        generate::{GenerationDataProvider, PrepareData},
    },
    params::chat::ChatCompletionParameters,
};
use anyhow::Result;
use candle_core::{DType, Device};
use candle_nn::VarBuilder;

use crate::{
    chat_template::ChatTemplate,
    models::{
        lfm2::config::Lfm2GenerateConfig,
        lfm2vl::{config::Lfm2VLConfig, model::Lfm2VLModel, processor::Lfm2VLProcessor},
    },
    tokenizer::TokenizerModel,
    utils::{find_type_files, get_device, get_dtype},
};

pub struct Lfm2VLGenerateModel<'a> {
    chat_template: ChatTemplate<'a>,
    tokenizer: TokenizerModel,
    device: Device,
    model: Lfm2VLModel,
    processor: Lfm2VLProcessor,
    model_name: String,
}
impl<'a> Lfm2VLGenerateModel<'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: Lfm2VLConfig = serde_json::from_slice(&std::fs::read(cfg_path)?)?;

        let model_path = find_type_files(path, "safetensors")?;
        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 = Lfm2VLModel::new(vb, &cfg, eos_ids)?;
        let processor = Lfm2VLProcessor::new(path, dtype, &device)?;
        let model_name = std::path::Path::new(path)
            .file_name()
            .and_then(|s| s.to_str())
            .unwrap_or("lfm2.5-vl")
            .to_string();
        Ok(Self {
            chat_template,
            tokenizer,
            device,
            model,
            processor,
            model_name,
        })
    }
}

impl<'a> GenerationDataProvider for Lfm2VLGenerateModel<'a> {
    fn get_data(&self, mes: &ChatCompletionParameters) -> Result<PrepareData> {
        let mes_render = self.chat_template.apply_chat_template(mes)?;
        let (pixel_values, pixel_attention_mask, spatial_shapes, text) =
            self.processor.process_info(mes, &mes_render)?;
        let input_ids = self.tokenizer.text_encode(text, &self.device)?;
        let data_vec = vec![
            pixel_values.into(),
            pixel_attention_mask.into(),
            spatial_shapes.into(),
        ];
        let multi_model_data = MultiModalData::new(data_vec);
        Ok(PrepareData {
            in_reasoning: false,
            input_ids,
            multi_model_data,
        })
    }
}

crate::impl_generate_model!(Lfm2VLGenerateModel<'a>);