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
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213

use anyhow::Result;
use candle_core::{DType, Device, Tensor};

use super::config::GlmOcrPreprocessorConfig;
use crate::tokenizer::TokenizerModel;
use crate::utils::img_utils::get_image;
use crate::utils::video_utils::video_smart_resize;

/// GLM-OCR Processor for image and text preprocessing.
///
/// Matches Python's Glm46VImageProcessor behavior:
/// - Uses smart_resize to compute target dimensions
/// - Outputs flattened patches format [num_patches, patch_dim]
pub struct GlmOcrProcessor {
    image_mean: Vec<f32>,
    image_std: Vec<f32>,
    shortest_edge: usize, // min_pixels in Python
    longest_edge: usize,  // max_pixels in Python
    patch_size: usize,
    merge_size: usize,
    temporal_patch_size: usize,
    device: Device,
    dtype: DType,
}

pub struct ProcessedImage {
    pub pixel_values: Tensor, // Shape: [num_patches, patch_dim]
    pub grid_h: usize,
    pub grid_w: usize,
}

pub struct ProcessedInput {
    pub input_ids: Tensor,
    pub pixel_values: Tensor, // Shape: [num_patches, patch_dim]
    pub image_mask: Tensor,
    pub grid_thw: Tensor,
}

impl GlmOcrProcessor {
    pub fn new(path: &str, device: &Device, dtype: DType) -> Result<Self> {
        assert!(
            std::path::Path::new(path).exists(),
            "model path file not exists"
        );
        let config_path = format!("{}/preprocessor_config.json", path);
        assert!(
            std::path::Path::new(&config_path).exists(),
            "preprocessor_config.json not exists in model path"
        );
        let process_cfg: GlmOcrPreprocessorConfig =
            serde_json::from_slice(&std::fs::read(config_path)?)?;

        Ok(Self {
            image_mean: process_cfg.image_mean.clone(),
            image_std: process_cfg.image_std.clone(),
            shortest_edge: process_cfg.size.shortest_edge,
            longest_edge: process_cfg.size.longest_edge,
            patch_size: process_cfg.patch_size,
            merge_size: process_cfg.merge_size,
            temporal_patch_size: process_cfg.temporal_patch_size, // Fixed for images
            device: device.clone(),
            dtype,
        })
    }

    /// Process image for vision encoder.
    ///
    /// Matches Python's Glm46VImageProcessor._preprocess():
    /// 1. Resize using smart_resize
    /// 2. Normalize
    /// 3. Reshape into flattened patches [num_patches, patch_dim]
    ///
    /// Output format: [grid_t * grid_h * grid_w, channels * temporal_patch_size * patch_size * patch_size]
    /// For images: grid_t = 1, so [grid_h * grid_w, 3 * 2 * 14 * 14] = [num_patches, 1176]
    pub fn process_image(&self, image_path: &str) -> Result<ProcessedImage> {
        let img = get_image(image_path)?;
        // Use smart_resize to compute target dimensions
        let (target_h, target_w) = video_smart_resize(
            self.temporal_patch_size as u32,
            img.height(),
            img.width(),
            self.temporal_patch_size as u32,
            (self.patch_size * self.merge_size) as u32,
            self.shortest_edge as u32,
            self.longest_edge as u32,
            None,
        )?;

        // Resize image
        let img = img.resize_exact(target_w, target_h, image::imageops::FilterType::Lanczos3);

        let target_h = target_h as usize;
        let target_w = target_w as usize;
        // Convert to RGB and normalize
        let img = img.to_rgb8();
        let pixels: Vec<f32> = img
            .pixels()
            .flat_map(|p| {
                vec![
                    p[0] as f32 / 255.0,
                    p[1] as f32 / 255.0,
                    p[2] as f32 / 255.0,
                ]
            })
            .collect();

        let tensor = Tensor::from_vec(pixels, (target_h, target_w, 3), &self.device)?;

        let mean = Tensor::new(self.image_mean.clone(), &self.device)?.reshape((1, 1, 3))?;
        let std = Tensor::new(self.image_std.clone(), &self.device)?.reshape((1, 1, 3))?;
        let tensor = tensor.broadcast_sub(&mean)?.broadcast_div(&std)?;

        // Now reshape into flattened patches like Python
        let grid_h = target_h / self.patch_size;
        let grid_w = target_w / self.patch_size;
        let channels = 3;
        let tensor =
            tensor.reshape((grid_h, self.patch_size, grid_w, self.patch_size, channels))?;
        let tensor = tensor.permute((0, 2, 4, 1, 3))?;
        let num_patches = grid_h * grid_w;
        let tensor = tensor.reshape((num_patches, channels, self.patch_size, self.patch_size))?;
        let tensor = tensor.unsqueeze(2)?;
        let tensor = tensor.repeat((1, 1, self.temporal_patch_size, 1, 1))?;
        let patch_dim = channels * self.temporal_patch_size * self.patch_size * self.patch_size;
        let tensor = tensor.reshape((num_patches, patch_dim))?;

        let tensor = tensor.to_dtype(self.dtype)?;

        Ok(ProcessedImage {
            pixel_values: tensor,
            grid_h,
            grid_w,
        })
    }

    /// Process image and text for multimodal input
    pub fn process_info(
        &self,
        image_path: &str,
        prompt: &str,
        tokenizer: &TokenizerModel,
        image_token_id: u32,
        image_start_token_id: u32,
        image_end_token_id: u32,
        _patch_size: usize,
        _temporal_patch_size: usize,
        spatial_merge_size: usize,
    ) -> Result<ProcessedInput> {
        let processed_image = self.process_image(image_path)?;
        let pixel_values = processed_image.pixel_values;
        let grid_h = processed_image.grid_h;
        let grid_w = processed_image.grid_w;

        // After spatial merge, each spatial_merge_size x spatial_merge_size block becomes 1 token
        let merged_h = grid_h / spatial_merge_size;
        let merged_w = grid_w / spatial_merge_size;
        let num_image_tokens = merged_h * merged_w;

        // GLM-OCR format: [gMASK] <sop> <|user|> \n <|begin_of_image|> <|image|>*N <|end_of_image|> text <|assistant|> \n
        // Special token IDs:
        //   59248 = [gMASK]
        //   59250 = <sop>
        //   59253 = <|user|>
        //   59256 = <|begin_of_image|>
        //   59280 = <|image|>
        //   59257 = <|end_of_image|>
        //   59254 = <|assistant|>

        // Build input_ids following Python format
        // Header: [gMASK] <sop> <|user|> \n Image tokens: <|begin_of_image|> <|image|>*N <|end_of_image|>
        //          59248  59250  59253    10
        let mut input_ids_vec = vec![59248, 59250, 59253, 10, image_start_token_id];

        for _ in 0..num_image_tokens {
            input_ids_vec.push(image_token_id); // <|image|>
        }
        input_ids_vec.push(image_end_token_id); // <|end_of_image|>

        // Text prompt (without special tokens - they're already added)
        let text_ids = tokenizer.text_encode_vec(prompt.to_string(), false)?;
        input_ids_vec.extend(text_ids);

        // Generation prompt: <|assistant|> \n
        input_ids_vec.push(59254); // <|assistant|>
        input_ids_vec.push(10); // newline

        let input_ids = Tensor::from_vec(
            input_ids_vec.clone(),
            (1, input_ids_vec.len()),
            &self.device,
        )?;

        // Create image mask (1s at image token positions, 0 elsewhere)
        // Image tokens start after header (4 tokens) + start token (1 token) = index 5
        let mut image_mask_vec = vec![0u32; input_ids_vec.len()];
        let image_start_idx = 5; // After [gMASK, sop, user, newline, begin_image]
        for i in 0..num_image_tokens {
            image_mask_vec[image_start_idx + i] = 1;
        }
        let image_mask = Tensor::from_vec(image_mask_vec, (1, input_ids_vec.len()), &self.device)?;

        // Compute grid_thw for RoPE
        let grid_thw =
            Tensor::from_vec(vec![1u32, grid_h as u32, grid_w as u32], (3,), &self.device)?;

        Ok(ProcessedInput {
            input_ids,
            pixel_values,
            image_mask,
            grid_thw,
        })
    }
}