muna 0.0.14

Run prediction functions in your Rust apps.
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
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231

/*
*   Muna
*   Copyright © 2026 NatML Inc. All Rights Reserved.
*/

use std::collections::HashMap;

use base64::Engine;
use base64::engine::general_purpose::STANDARD as BASE64;

use crate::c;
use crate::client::{MunaError, Result};
use crate::services::{PredictionService, PredictorService};
use crate::types::{Acceleration, Value};

use super::schema::{ImageData, ImageResponse};

/// Image size.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum ImageSize {
    Auto,
    Size256x256,
    Size512x512,
    Size1024x1024,
    Size1536x1024,
    Size1024x1536,
    Size1792x1024,
    Size1024x1792,
}

impl ImageSize {
    /// Requested `(width, height)`; `Auto` requests neither.
    pub fn dimensions(&self) -> (Option<i32>, Option<i32>) {
        match self {
            Self::Auto => (None, None),
            Self::Size256x256 => (Some(256), Some(256)),
            Self::Size512x512 => (Some(512), Some(512)),
            Self::Size1024x1024 => (Some(1024), Some(1024)),
            Self::Size1536x1024 => (Some(1536), Some(1024)),
            Self::Size1024x1536 => (Some(1024), Some(1536)),
            Self::Size1792x1024 => (Some(1792), Some(1024)),
            Self::Size1024x1792 => (Some(1024), Some(1792)),
        }
    }
}

/// Image creation parameters.
pub struct ImageCreateParams {
    /// Text prompt.
    pub prompt: String,
    /// Image predictor tag.
    pub model: String,
    /// Background transparency.
    pub background: Option<String>,
    /// Number of images to generate.
    pub n: Option<i32>,
    /// Output format (`png`, `jpeg`, `webp`, `raw`).
    pub output_format: Option<String>,
    /// Compression level (0-100).
    pub output_compression: Option<i32>,
    /// Image size.
    pub size: Option<ImageSize>,
    /// Prediction acceleration.
    pub acceleration: Option<Acceleration>,
}

struct ImageConfig {
    prompt_param_name: String,
    width_param: Option<String>,
    height_param: Option<String>,
    count_param: Option<String>,
    image_param_idx: usize,
}

/// Image generation service.
#[derive(Clone)]
pub struct ImageService {
    predictors: PredictorService,
    predictions: PredictionService,
    cache: std::sync::Arc<tokio::sync::Mutex<HashMap<String, ImageConfig>>>,
}

impl ImageService {

    pub fn new(
        predictors: PredictorService,
        predictions: PredictionService,
    ) -> Self {
        Self {
            predictors,
            predictions,
            cache: std::sync::Arc::new(tokio::sync::Mutex::new(HashMap::new())),
        }
    }

    /// Generate images from a text prompt.
    pub async fn generate(&self, params: ImageCreateParams) -> Result<ImageResponse> {
        let tag = params.model.clone();
        let output_format = params.output_format.as_deref().unwrap_or("png");
        let acceleration = params.acceleration.unwrap_or(Acceleration::LocalAuto);
        {
            let mut cache = self.cache.lock().await;
            if !cache.contains_key(&tag) {
                let config = self.create_config(&tag).await?;
                cache.insert(tag.clone(), config);
            }
        }
        let cache = self.cache.lock().await;
        let config = cache.get(&tag).unwrap();
        let (req_width, req_height) = params.size.unwrap_or(ImageSize::Auto).dimensions();
        let mut prediction_inputs: HashMap<String, Value> = HashMap::new();
        // The prompt parameter is list-typed so prompts from multiple requests
        // can be batched into one invocation; a single generation is just a
        // batch of one.
        prediction_inputs.insert(
            config.prompt_param_name.clone(),
            Value::List(vec![serde_json::Value::String(params.prompt)])
        );
        if let (Some(n), Some(name)) = (&params.n, &config.count_param) {
            prediction_inputs.insert(name.clone(), Value::Int(*n));
        }
        if let (Some(w), Some(name)) = (req_width, &config.width_param) {
            prediction_inputs.insert(name.clone(), Value::Int(w));
        }
        if let (Some(h), Some(name)) = (req_height, &config.height_param) {
            prediction_inputs.insert(name.clone(), Value::Int(h));
        }
        let image_idx = config.image_param_idx;
        drop(cache);

        let prediction = self.create_prediction(&tag, prediction_inputs, acceleration).await?;
        if let Some(error) = &prediction.error {
            return Err(MunaError::Prediction(error.clone()));
        }
        let results = prediction.results.as_ref().ok_or_else(|| {
            MunaError::Prediction("No results returned".into())
        })?;
        let images_value = results.get(image_idx).ok_or_else(|| {
            MunaError::Prediction(format!("{tag} did not return images"))
        })?;
        // `create_config` guarantees the output at `image_idx` is an image list.
        let images = match images_value {
            Value::ImageList(list) => {
                let mut image_data = Vec::new();
                for img in list {
                    image_data.push(Self::encode_image(img, output_format)?);
                }
                image_data
            }
            _ => return Err(MunaError::Prediction(format!(
                "{tag} returned unexpected type instead of images"
            ))),
        };
        let created = std::time::SystemTime::now()
            .duration_since(std::time::UNIX_EPOCH)
            .unwrap_or_default()
            .as_secs() as i64;
        Ok(ImageResponse {
            data: Some(images),
            background: Some("opaque".to_string()),
            created,
            usage: None,
        })
    }

    async fn create_config(&self, tag: &str) -> Result<ImageConfig> {
        let predictor = self.predictors.retrieve(tag).await?.ok_or_else(|| {
            MunaError::Prediction(format!("{tag} cannot be used with image API: predictor not found"))
        })?;
        let signature = &predictor.signature;
        let required: Vec<_> = signature.inputs.iter()
            .filter(|p| p.optional != Some(true))
            .collect();
        if required.len() != 1 {
            return Err(MunaError::Prediction(format!(
                "{tag} cannot be used with image API: expected 1 required input"
            )));
        }
        let prompt_param = required.iter()
            .find(|p| p.dtype == Some(crate::types::Dtype::List))
            .ok_or_else(|| MunaError::Prediction(format!("{tag}: no text prompt parameter")))?;
        let is_int = |d: crate::types::Dtype| matches!(d,
            crate::types::Dtype::Int8 | crate::types::Dtype::Int16 | crate::types::Dtype::Int32 | crate::types::Dtype::Int64 |
            crate::types::Dtype::Uint8 | crate::types::Dtype::Uint16 | crate::types::Dtype::Uint32 | crate::types::Dtype::Uint64
        );
        let width_param = signature.inputs.iter()
            .find(|p| p.dtype.map_or(false, is_int) && p.denotation.as_deref() == Some("openai.images.width"))
            .map(|p| p.name.clone());
        let height_param = signature.inputs.iter()
            .find(|p| p.dtype.map_or(false, is_int) && p.denotation.as_deref() == Some("openai.images.height"))
            .map(|p| p.name.clone());
        let count_param = signature.inputs.iter()
            .find(|p| p.dtype.map_or(false, is_int) && p.denotation.as_deref() == Some("openai.images.count"))
            .map(|p| p.name.clone());
        let image_idx = signature.outputs.iter().position(|p| {
            p.dtype == Some(crate::types::Dtype::ImageList)
        }).ok_or_else(|| MunaError::Prediction(format!("{tag}: no image_list output")))?;
        Ok(ImageConfig {
            prompt_param_name: prompt_param.name.clone(),
            width_param,
            height_param,
            count_param,
            image_param_idx: image_idx,
        })
    }

    fn encode_image(
        image: &crate::types::Image,
        output_format: &str,
    ) -> Result<ImageData> {
        if output_format == "raw" {
            return Ok(ImageData { b64_json: None, image: Some(image.clone()) });
        }
        let fxn_value = c::Value::from_object(&Value::Image(image.clone()))?;
        let mime = format!("image/{output_format}");
        let buffer = fxn_value.serialize(Some(&mime))?;
        Ok(ImageData { b64_json: Some(BASE64.encode(&buffer)), image: None })
    }

    async fn create_prediction(
        &self,
        tag: &str,
        inputs: HashMap<String, Value>,
        acceleration: Acceleration,
    ) -> Result<crate::types::Prediction> {
        // `predictions.create` routes remote accelerations internally.
        self.predictions
            .create(tag, Some(inputs), Some(acceleration), None, None)
            .await
    }
}