mold-ai 0.15.0

Local AI image generation CLI — FLUX, SDXL, SD3.5, Z-Image diffusion models on your GPU
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

//! Shared local-engine preamble for `mold run` (single clip) and
//! `mold run --frames N` chains in `--local` mode: resolve-or-pull the
//! model, apply encoder-variant env overrides, resolve eager/offload,
//! pick a GPU, and construct the engine. One home so the generate and
//! chain paths can't drift.

#[cfg(any(feature = "cuda", feature = "metal"))]
use anyhow::Result;
#[cfg(any(feature = "cuda", feature = "metal"))]
use colored::Colorize;
#[cfg(any(feature = "cuda", feature = "metal"))]
use mold_core::{Config, ModelPaths};

#[cfg(any(feature = "cuda", feature = "metal"))]
use crate::output::status;
#[cfg(any(feature = "cuda", feature = "metal"))]
use crate::theme;

/// Engine-construction knobs collected from CLI flags.
#[cfg(any(feature = "cuda", feature = "metal"))]
#[derive(Debug, Default)]
pub(crate) struct EngineOverrides {
    pub gpus: Option<String>,
    pub t5_variant: Option<String>,
    pub qwen3_variant: Option<String>,
    pub qwen2_variant: Option<String>,
    pub qwen2_text_encoder_mode: Option<String>,
    pub eager: bool,
    pub offload: bool,
}

/// Export encoder-variant overrides as `MOLD_*` env vars so the engine
/// factory's auto-select picks them up during construction.
#[cfg(any(feature = "cuda", feature = "metal", test))]
pub(crate) fn apply_local_engine_env_overrides(
    t5_variant_override: Option<&str>,
    qwen3_variant_override: Option<&str>,
    qwen2_variant_override: Option<&str>,
    qwen2_text_encoder_mode_override: Option<&str>,
) {
    if let Some(variant) = t5_variant_override {
        std::env::set_var("MOLD_T5_VARIANT", variant);
    }
    if let Some(variant) = qwen3_variant_override {
        std::env::set_var("MOLD_QWEN3_VARIANT", variant);
    }
    if let Some(variant) = qwen2_variant_override {
        std::env::set_var("MOLD_QWEN2_VARIANT", variant);
    }
    if let Some(mode) = qwen2_text_encoder_mode_override {
        std::env::set_var("MOLD_QWEN2_TEXT_ENCODER_MODE", mode);
    }
}

/// Resolve model paths, auto-pulling when the model is missing locally or
/// has missing assets (repair pull). Returns the paths, the effective
/// config (post-pull when a pull happened), and whether a pull ran — a
/// `true` tells callers to re-derive request defaults from the refreshed
/// model config.
#[cfg(any(feature = "cuda", feature = "metal"))]
pub(crate) async fn resolve_or_pull_model(
    model: &str,
    config: &Config,
) -> Result<(ModelPaths, Config, bool)> {
    use mold_core::manifest::find_manifest;

    if config.manifest_model_needs_download(model) {
        status!(
            "{} Model '{}' is missing local assets, pulling repair...",
            theme::icon_info(),
            model.bold(),
        );
        let updated =
            super::pull::pull_and_configure(model, &mold_core::download::PullOptions::default())
                .await?;
        let paths = ModelPaths::resolve(model, &updated).ok_or_else(|| {
            anyhow::anyhow!("model '{model}' was pulled but paths could not be resolved")
        })?;
        Ok((paths, updated, true))
    } else if let Some(paths) = ModelPaths::resolve(model, config) {
        Ok((paths, config.clone(), false))
    } else if find_manifest(model).is_some() {
        status!(
            "{} Model '{}' not found locally, pulling...",
            theme::icon_info(),
            model.bold(),
        );
        let updated =
            super::pull::pull_and_configure(model, &mold_core::download::PullOptions::default())
                .await?;
        let paths = ModelPaths::resolve(model, &updated).ok_or_else(|| {
            anyhow::anyhow!("model '{model}' was pulled but paths could not be resolved")
        })?;
        Ok((paths, updated, true))
    } else {
        anyhow::bail!(
            "no model paths configured for '{model}'. Add [models.{model}] to \
             ~/.mold/config.toml, pull via `mold pull {model}`, or set \
             MOLD_TRANSFORMER_PATH / MOLD_VAE_PATH / MOLD_T5_PATH / MOLD_CLIP_PATH \
             / MOLD_T5_TOKENIZER_PATH / MOLD_CLIP_TOKENIZER_PATH env vars.",
        );
    }
}

/// Apply env overrides, resolve eager/offload, select the best GPU from
/// the allowed set (most free VRAM), and construct the engine.
#[cfg(any(feature = "cuda", feature = "metal"))]
pub(crate) fn build_local_engine(
    model: &str,
    paths: ModelPaths,
    config: &Config,
    ov: &EngineOverrides,
) -> Result<Box<dyn mold_inference::InferenceEngine>> {
    use mold_inference::LoadStrategy;

    apply_local_engine_env_overrides(
        ov.t5_variant.as_deref(),
        ov.qwen3_variant.as_deref(),
        ov.qwen2_variant.as_deref(),
        ov.qwen2_text_encoder_mode.as_deref(),
    );

    let is_eager = ov.eager || std::env::var("MOLD_EAGER").is_ok_and(|v| v == "1");
    let load_strategy = if is_eager {
        LoadStrategy::Eager
    } else {
        LoadStrategy::Sequential
    };
    if is_eager {
        std::env::set_var("MOLD_EAGER", "1");
    }
    let is_offload = ov.offload || std::env::var("MOLD_OFFLOAD").is_ok_and(|v| v == "1");

    let gpu_selection = match &ov.gpus {
        Some(s) => mold_core::types::GpuSelection::parse(s)?,
        None => config.gpu_selection(),
    };
    let discovered = mold_inference::device::discover_gpus();
    let available = mold_inference::device::filter_gpus(&discovered, &gpu_selection);
    let gpu_ordinal = mold_inference::device::select_best_gpu(&available)
        .map(|g| g.ordinal)
        .unwrap_or(0);

    Ok(mold_inference::create_engine(
        model.to_string(),
        paths,
        config,
        load_strategy,
        gpu_ordinal,
        is_offload,
    )?)
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::test_support::ENV_LOCK;

    #[test]
    fn apply_local_engine_env_overrides_sets_qwen2_overrides() {
        let _lock = ENV_LOCK.lock().unwrap_or_else(|e| e.into_inner());
        let prior_variant = std::env::var("MOLD_QWEN2_VARIANT").ok();
        let prior_mode = std::env::var("MOLD_QWEN2_TEXT_ENCODER_MODE").ok();

        std::env::remove_var("MOLD_QWEN2_VARIANT");
        std::env::remove_var("MOLD_QWEN2_TEXT_ENCODER_MODE");

        apply_local_engine_env_overrides(None, None, Some("q6"), Some("cpu-stage"));

        assert_eq!(
            std::env::var("MOLD_QWEN2_VARIANT").ok().as_deref(),
            Some("q6")
        );
        assert_eq!(
            std::env::var("MOLD_QWEN2_TEXT_ENCODER_MODE")
                .ok()
                .as_deref(),
            Some("cpu-stage")
        );

        match prior_variant {
            Some(value) => std::env::set_var("MOLD_QWEN2_VARIANT", value),
            None => std::env::remove_var("MOLD_QWEN2_VARIANT"),
        }
        match prior_mode {
            Some(value) => std::env::set_var("MOLD_QWEN2_TEXT_ENCODER_MODE", value),
            None => std::env::remove_var("MOLD_QWEN2_TEXT_ENCODER_MODE"),
        }
    }
}