Skip to main content

hf_fetch_model/
config.rs

1// SPDX-License-Identifier: MIT OR Apache-2.0
2
3//! Configuration for model downloads.
4//!
5//! [`FetchConfig`] controls revision, authentication, file filtering,
6//! concurrency, timeouts, retry behavior, and progress reporting.
7
8use std::path::PathBuf;
9use std::sync::Arc;
10use std::time::Duration;
11
12use globset::{Glob, GlobSet, GlobSetBuilder};
13
14use crate::error::FetchError;
15use crate::progress::ProgressEvent;
16
17// TRAIT_OBJECT: heterogeneous progress handlers from different callers
18pub(crate) type ProgressCallback = Arc<dyn Fn(&ProgressEvent) + Send + Sync>;
19
20/// Configuration for downloading a model repository.
21///
22/// Use [`FetchConfig::builder()`] to construct.
23///
24/// # Example
25///
26/// ```rust
27/// # fn example() -> Result<(), hf_fetch_model::FetchError> {
28/// use hf_fetch_model::FetchConfig;
29///
30/// let config = FetchConfig::builder()
31///     .revision("main")
32///     .filter("*.safetensors")
33///     .concurrency(4)
34///     .build()?;
35/// # Ok(())
36/// # }
37/// ```
38pub struct FetchConfig {
39    /// Git revision (branch, tag, or commit SHA). `None` means `"main"`.
40    pub(crate) revision: Option<String>,
41    /// Authentication token for gated/private repositories.
42    pub(crate) token: Option<String>,
43    /// Compiled include glob patterns. Only matching files are downloaded.
44    pub(crate) include: Option<GlobSet>,
45    /// Compiled exclude glob patterns. Matching files are skipped.
46    pub(crate) exclude: Option<GlobSet>,
47    /// Number of files to download in parallel.
48    pub(crate) concurrency: usize,
49    /// Custom cache directory (overrides the default HF cache).
50    pub(crate) output_dir: Option<PathBuf>,
51    /// Maximum time allowed for a single file download.
52    pub(crate) timeout_per_file: Option<Duration>,
53    /// Maximum total time for the entire download operation.
54    pub(crate) timeout_total: Option<Duration>,
55    /// Maximum retry attempts per file (exponential backoff with jitter).
56    pub(crate) max_retries: u32,
57    /// Whether to verify SHA256 checksums against HF LFS metadata.
58    pub(crate) verify_checksums: bool,
59    /// Minimum file size (bytes) for multi-connection chunked download.
60    pub(crate) chunk_threshold: u64,
61    /// Number of parallel HTTP Range connections per large file.
62    pub(crate) connections_per_file: usize,
63    // TRAIT_OBJECT: heterogeneous progress handlers from different callers
64    /// Progress callback invoked for each download event.
65    pub(crate) on_progress: Option<ProgressCallback>,
66    /// Tracks which performance fields the user explicitly set via the builder.
67    pub(crate) explicit: ExplicitSettings,
68}
69
70/// Tracks which performance fields the user explicitly set via the builder.
71///
72/// Used by the implicit plan retrofit: when a field was not explicitly set,
73/// the plan-based optimizer may override it with a recommended value.
74#[derive(Debug, Clone, Default)]
75pub(crate) struct ExplicitSettings {
76    /// Whether `concurrency` was explicitly set by the caller.
77    pub(crate) concurrency: bool,
78    /// Whether `connections_per_file` was explicitly set by the caller.
79    pub(crate) connections_per_file: bool,
80    /// Whether `chunk_threshold` was explicitly set by the caller.
81    pub(crate) chunk_threshold: bool,
82}
83
84impl std::fmt::Debug for FetchConfig {
85    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
86        f.debug_struct("FetchConfig")
87            .field("revision", &self.revision)
88            .field("token", &self.token.as_ref().map(|_| "***"))
89            .field("include", &self.include)
90            .field("exclude", &self.exclude)
91            .field("concurrency", &self.concurrency)
92            .field("output_dir", &self.output_dir)
93            .field("timeout_per_file", &self.timeout_per_file)
94            .field("timeout_total", &self.timeout_total)
95            .field("max_retries", &self.max_retries)
96            .field("verify_checksums", &self.verify_checksums)
97            .field("chunk_threshold", &self.chunk_threshold)
98            .field("connections_per_file", &self.connections_per_file)
99            .field(
100                "on_progress",
101                if self.on_progress.is_some() {
102                    &"Some(<fn>)"
103                } else {
104                    &"None"
105                },
106            )
107            .field("explicit", &self.explicit)
108            .finish()
109    }
110}
111
112impl FetchConfig {
113    /// Creates a new [`FetchConfigBuilder`].
114    #[must_use]
115    pub fn builder() -> FetchConfigBuilder {
116        FetchConfigBuilder::default()
117    }
118
119    /// Returns the configured concurrency level (parallel file downloads).
120    #[must_use]
121    pub const fn concurrency(&self) -> usize {
122        self.concurrency
123    }
124
125    /// Returns the configured number of parallel HTTP connections per file.
126    #[must_use]
127    pub const fn connections_per_file(&self) -> usize {
128        self.connections_per_file
129    }
130
131    /// Returns the chunk threshold in bytes (minimum file size for
132    /// multi-connection chunked downloads).
133    #[must_use]
134    pub const fn chunk_threshold(&self) -> u64 {
135        self.chunk_threshold
136    }
137}
138
139/// Builder for [`FetchConfig`].
140#[derive(Default)]
141pub struct FetchConfigBuilder {
142    revision: Option<String>,
143    token: Option<String>,
144    include_patterns: Vec<String>,
145    exclude_patterns: Vec<String>,
146    concurrency: Option<usize>,
147    output_dir: Option<PathBuf>,
148    timeout_per_file: Option<Duration>,
149    timeout_total: Option<Duration>,
150    max_retries: Option<u32>,
151    verify_checksums: Option<bool>,
152    chunk_threshold: Option<u64>,
153    connections_per_file: Option<usize>,
154    on_progress: Option<ProgressCallback>,
155}
156
157impl FetchConfigBuilder {
158    /// Sets the git revision (branch, tag, or commit SHA) to download.
159    ///
160    /// Defaults to `"main"` if not set.
161    #[must_use]
162    pub fn revision(mut self, revision: &str) -> Self {
163        // BORROW: explicit .to_owned() for &str → owned String
164        self.revision = Some(revision.to_owned());
165        self
166    }
167
168    /// Sets the authentication token.
169    #[must_use]
170    pub fn token(mut self, token: &str) -> Self {
171        // BORROW: explicit .to_owned() for &str → owned String
172        self.token = Some(token.to_owned());
173        self
174    }
175
176    /// Reads the authentication token from the `HF_TOKEN` environment variable.
177    #[must_use]
178    pub fn token_from_env(mut self) -> Self {
179        self.token = std::env::var("HF_TOKEN").ok();
180        self
181    }
182
183    /// Adds an include glob pattern. Only files matching at least one
184    /// include pattern will be downloaded.
185    ///
186    /// Can be called multiple times to add multiple patterns.
187    #[must_use]
188    pub fn filter(mut self, pattern: &str) -> Self {
189        // BORROW: explicit .to_owned() for &str → owned String
190        self.include_patterns.push(pattern.to_owned());
191        self
192    }
193
194    /// Adds an exclude glob pattern. Files matching any exclude pattern
195    /// will be skipped, even if they match an include pattern.
196    ///
197    /// Can be called multiple times to add multiple patterns.
198    #[must_use]
199    pub fn exclude(mut self, pattern: &str) -> Self {
200        // BORROW: explicit .to_owned() for &str → owned String
201        self.exclude_patterns.push(pattern.to_owned());
202        self
203    }
204
205    /// Sets the number of files to download concurrently.
206    ///
207    /// When omitted, the download plan optimizer auto-tunes this value
208    /// based on file count and size distribution. Falls back to 4 if
209    /// no plan recommendation is available.
210    #[must_use]
211    pub fn concurrency(mut self, concurrency: usize) -> Self {
212        self.concurrency = Some(concurrency);
213        self
214    }
215
216    /// Sets a custom output directory for downloaded files.
217    ///
218    /// By default, files are stored in the standard `HuggingFace` cache directory
219    /// (`~/.cache/huggingface/hub/`). When set, the `HuggingFace` cache hierarchy
220    /// is created inside this directory instead.
221    #[must_use]
222    pub fn output_dir(mut self, dir: PathBuf) -> Self {
223        self.output_dir = Some(dir);
224        self
225    }
226
227    /// Sets the maximum time allowed per file download.
228    ///
229    /// Defaults to 300 seconds when not set. If a single file download
230    /// exceeds this duration, it is aborted and may be retried according
231    /// to the retry policy.
232    ///
233    /// For chunked (multi-connection) downloads, an abort here does **not**
234    /// discard work already on disk: the partial `.chunked.part` file and
235    /// its per-chunk progress sidecar are preserved, and a subsequent call
236    /// for the same file resumes from each chunk's last checkpoint via
237    /// `Range` requests — provided the upstream etag still matches. Raise
238    /// this when downloading large files on slow links (e.g.
239    /// `Duration::from_secs(1800)` for 5–15 GiB files at home-broadband
240    /// speeds).
241    #[must_use]
242    pub fn timeout_per_file(mut self, duration: Duration) -> Self {
243        self.timeout_per_file = Some(duration);
244        self
245    }
246
247    /// Sets the maximum total time for the entire download operation.
248    ///
249    /// Defaults to no limit when not set. If the total download time
250    /// exceeds this duration, remaining files are skipped and a
251    /// [`FetchError::Timeout`] is returned.
252    ///
253    /// Independent of [`timeout_per_file`](Self::timeout_per_file): the
254    /// per-file budget bounds any single transfer, while this bounds the
255    /// whole batch (including retries between transfers). Files that
256    /// completed before the total elapsed are kept; an interrupted
257    /// in-flight chunked transfer leaves its partial + sidecar on disk
258    /// for resume on a future call.
259    #[must_use]
260    pub fn timeout_total(mut self, duration: Duration) -> Self {
261        self.timeout_total = Some(duration);
262        self
263    }
264
265    /// Sets the maximum number of retry attempts per file.
266    ///
267    /// Defaults to 3. Set to 0 to disable retries.
268    /// Uses exponential backoff with jitter (base 300ms, cap 10s).
269    #[must_use]
270    pub fn max_retries(mut self, retries: u32) -> Self {
271        self.max_retries = Some(retries);
272        self
273    }
274
275    /// Enables or disables SHA256 checksum verification after download.
276    ///
277    /// When enabled, downloaded files are verified against the SHA256 hash
278    /// from `HuggingFace` LFS metadata. Files without LFS metadata (small
279    /// config files stored directly in git) are skipped.
280    ///
281    /// Defaults to `true`.
282    #[must_use]
283    pub fn verify_checksums(mut self, verify: bool) -> Self {
284        self.verify_checksums = Some(verify);
285        self
286    }
287
288    /// Sets the minimum file size (in bytes) for chunked parallel download.
289    ///
290    /// Files at or above this threshold are downloaded using multiple HTTP
291    /// Range connections in parallel. Files below use the standard single
292    /// connection. Set to `u64::MAX` to disable chunked downloads entirely.
293    ///
294    /// When omitted, the download plan optimizer auto-tunes this value
295    /// based on file size distribution. Falls back to 100 MiB
296    /// (104\_857\_600 bytes) if no plan recommendation is available.
297    #[must_use]
298    pub fn chunk_threshold(mut self, bytes: u64) -> Self {
299        self.chunk_threshold = Some(bytes);
300        self
301    }
302
303    /// Sets the number of parallel HTTP connections per large file.
304    ///
305    /// Only applies to files at or above `chunk_threshold`. When omitted,
306    /// the download plan optimizer auto-tunes this value based on file size
307    /// distribution. Falls back to 8 if no plan recommendation is available.
308    #[must_use]
309    pub fn connections_per_file(mut self, connections: usize) -> Self {
310        self.connections_per_file = Some(connections);
311        self
312    }
313
314    /// Sets a progress callback invoked for each progress event.
315    #[must_use]
316    pub fn on_progress<F>(mut self, callback: F) -> Self
317    where
318        F: Fn(&ProgressEvent) + Send + Sync + 'static,
319    {
320        self.on_progress = Some(Arc::new(callback));
321        self
322    }
323
324    /// Creates a `tokio::sync::watch` channel for async progress consumption.
325    ///
326    /// Returns `(self, receiver)`. The receiver yields the latest [`ProgressEvent`]
327    /// via `.changed().await` + `.borrow()`. Only the most recent event is retained.
328    ///
329    /// The channel is initialized with [`ProgressEvent::default()`] (all zeros,
330    /// empty filename). Use `.changed().await` rather than eager `.borrow()` to
331    /// avoid observing this sentinel value before the first real event.
332    ///
333    /// Composes with [`on_progress()`](Self::on_progress) — if a callback was
334    /// already set, both the callback and the watch channel fire for every event.
335    #[must_use]
336    pub fn progress_channel(mut self) -> (Self, crate::progress::ProgressReceiver) {
337        let (tx, rx) = tokio::sync::watch::channel(ProgressEvent::default());
338        let existing = self.on_progress.take();
339        // TRAIT_OBJECT: heterogeneous progress handlers composed with watch sender
340        self.on_progress = Some(Arc::new(move |event: &ProgressEvent| {
341            if let Some(ref cb) = existing {
342                cb(event);
343            }
344            // Skip clone + send if no receiver is listening.
345            if tx.receiver_count() > 0 {
346                // BORROW: explicit .clone() for owned ProgressEvent sent through watch channel
347                let _ = tx.send(event.clone());
348            }
349        }));
350        (self, rx)
351    }
352
353    /// Builds the [`FetchConfig`].
354    ///
355    /// # Errors
356    ///
357    /// Returns [`FetchError::InvalidPattern`] if any glob pattern is invalid.
358    pub fn build(self) -> Result<FetchConfig, FetchError> {
359        let include = build_globset(&self.include_patterns)?;
360        let exclude = build_globset(&self.exclude_patterns)?;
361
362        Ok(FetchConfig {
363            revision: self.revision,
364            token: self.token,
365            include,
366            exclude,
367            concurrency: self.concurrency.unwrap_or(4),
368            output_dir: self.output_dir,
369            timeout_per_file: self.timeout_per_file,
370            timeout_total: self.timeout_total,
371            max_retries: self.max_retries.unwrap_or(3),
372            verify_checksums: self.verify_checksums.unwrap_or(true),
373            chunk_threshold: self.chunk_threshold.unwrap_or(104_857_600),
374            connections_per_file: self.connections_per_file.unwrap_or(8).max(1),
375            on_progress: self.on_progress,
376            explicit: ExplicitSettings {
377                concurrency: self.concurrency.is_some(),
378                connections_per_file: self.connections_per_file.is_some(),
379                chunk_threshold: self.chunk_threshold.is_some(),
380            },
381        })
382    }
383}
384
385/// Common filter presets for typical download patterns.
386#[non_exhaustive]
387pub struct Filter;
388
389impl Filter {
390    /// Returns a builder pre-configured to download only `*.safetensors` files
391    /// plus common config files.
392    #[must_use]
393    pub fn safetensors() -> FetchConfigBuilder {
394        FetchConfigBuilder::default()
395            .filter("*.safetensors")
396            .filter("*.json")
397            .filter("*.txt")
398    }
399
400    /// Returns a builder pre-configured to download only GGUF files
401    /// plus common config files.
402    #[must_use]
403    pub fn gguf() -> FetchConfigBuilder {
404        FetchConfigBuilder::default()
405            .filter("*.gguf")
406            .filter("*.json")
407            .filter("*.txt")
408    }
409
410    /// Returns a builder pre-configured to download only `.npz` and `.npy`
411    /// files plus common config files. Matches NumPy-based weight repos
412    /// such as Google's `GemmaScope` transcoders (`config.yaml` + many `.npz`).
413    #[must_use]
414    pub fn npz() -> FetchConfigBuilder {
415        FetchConfigBuilder::default()
416            .filter("*.npz")
417            .filter("*.npy")
418            .filter("config.yaml")
419            .filter("*.json")
420            .filter("*.txt")
421    }
422
423    /// Returns a builder pre-configured to download only `pytorch_model*.bin`
424    /// files plus common config files.
425    #[must_use]
426    pub fn pth() -> FetchConfigBuilder {
427        FetchConfigBuilder::default()
428            .filter("pytorch_model*.bin")
429            .filter("*.json")
430            .filter("*.txt")
431    }
432
433    /// Returns a builder pre-configured to download only config files
434    /// (no model weights).
435    #[must_use]
436    pub fn config_only() -> FetchConfigBuilder {
437        FetchConfigBuilder::default()
438            .filter("*.json")
439            .filter("*.txt")
440            .filter("*.md")
441    }
442}
443
444/// Returns whether `filename` passes the given include/exclude glob filters.
445///
446/// A file matches when it is not excluded by any `exclude` pattern **and**
447/// either there are no `include` patterns or it matches at least one.
448#[must_use]
449pub fn file_matches(filename: &str, include: Option<&GlobSet>, exclude: Option<&GlobSet>) -> bool {
450    if let Some(exc) = exclude {
451        if exc.is_match(filename) {
452            return false;
453        }
454    }
455    if let Some(inc) = include {
456        return inc.is_match(filename);
457    }
458    true
459}
460
461fn build_globset(patterns: &[String]) -> Result<Option<GlobSet>, FetchError> {
462    if patterns.is_empty() {
463        return Ok(None);
464    }
465    let mut builder = GlobSetBuilder::new();
466    for pattern in patterns {
467        // BORROW: explicit .as_str() instead of Deref coercion
468        let glob = Glob::new(pattern.as_str()).map_err(|e| FetchError::InvalidPattern {
469            pattern: pattern.clone(),
470            reason: e.to_string(),
471        })?;
472        builder.add(glob);
473    }
474    let set = builder.build().map_err(|e| FetchError::InvalidPattern {
475        pattern: patterns.join(", "),
476        reason: e.to_string(),
477    })?;
478    Ok(Some(set))
479}
480
481/// Builds a compiled [`GlobSet`] from a list of pattern strings.
482///
483/// Returns `None` if the pattern list is empty. This is useful for callers
484/// that need glob filtering outside the download pipeline (e.g., the
485/// `list-files` subcommand).
486///
487/// # Errors
488///
489/// Returns [`FetchError::InvalidPattern`] if any pattern fails to compile.
490pub fn compile_glob_patterns(patterns: &[String]) -> Result<Option<GlobSet>, FetchError> {
491    build_globset(patterns)
492}
493
494/// Returns `true` if `s` contains glob metacharacters (`*`, `?`, `[`, `{`).
495///
496/// Useful for detecting whether a user-supplied filename should be treated
497/// as a glob pattern or an exact match.
498#[must_use]
499pub fn has_glob_chars(s: &str) -> bool {
500    s.bytes().any(|b| matches!(b, b'*' | b'?' | b'[' | b'{'))
501}
502
503#[cfg(test)]
504mod tests {
505    #![allow(clippy::panic, clippy::unwrap_used, clippy::expect_used)]
506
507    use super::*;
508
509    #[test]
510    fn test_file_matches_no_filters() {
511        assert!(file_matches("model.safetensors", None, None));
512    }
513
514    #[test]
515    fn test_file_matches_include() {
516        let include = build_globset(&["*.safetensors".to_owned()]).unwrap();
517        assert!(file_matches("model.safetensors", include.as_ref(), None));
518        assert!(!file_matches("model.bin", include.as_ref(), None));
519    }
520
521    #[test]
522    fn test_file_matches_exclude() {
523        let exclude = build_globset(&["*.bin".to_owned()]).unwrap();
524        assert!(file_matches("model.safetensors", None, exclude.as_ref()));
525        assert!(!file_matches("model.bin", None, exclude.as_ref()));
526    }
527
528    #[test]
529    fn test_exclude_overrides_include() {
530        let include = build_globset(&["*.safetensors".to_owned(), "*.bin".to_owned()]).unwrap();
531        let exclude = build_globset(&["*.bin".to_owned()]).unwrap();
532        assert!(file_matches(
533            "model.safetensors",
534            include.as_ref(),
535            exclude.as_ref()
536        ));
537        assert!(!file_matches(
538            "model.bin",
539            include.as_ref(),
540            exclude.as_ref()
541        ));
542    }
543
544    #[test]
545    fn test_has_glob_chars() {
546        assert!(has_glob_chars("*.safetensors"));
547        assert!(has_glob_chars("model-[0-9].bin"));
548        assert!(has_glob_chars("model?.bin"));
549        assert!(has_glob_chars("{a,b}.bin"));
550        assert!(!has_glob_chars("model.safetensors"));
551        assert!(!has_glob_chars("config.json"));
552        assert!(!has_glob_chars("pytorch_model.bin"));
553    }
554}