The session id was replaced from `ulid` to `UniqueID` (a self
incrementing u64 in memory) in a previous PR but it's not correct.
The session id is used on CAS server logs and traces and CDN logs to
identity a related group of activity (for debugging and etc. purposes)
and it needs to be globally unique (thus using `ulid`) instead of
locally unique.
This PR should resolve all Dependabot alerts by upgrading deps and
switching out some deprecated crate for suggested alternatives, e.g.
`tempdir -> tempfile`. Supersede PR #721. Fix issue #722
The last repo restructuring didn't update several bench code that are
not compiled by default as part of "cargo build". This PR fixes those
compilation errors and warning, and adds "cargo bench --no-run" to CI
which checks compilation but doesn't actually run benchmarks.
This PR adds a full integration test suite on top of the xet session
interface that mimics the integration tests in xet_data/tests/. This one
additionally tests alternate asynchronous runtimes to ensure that the
bridge to the internal tokio runtime works correctly as well.
Currently, progress tracking is split between callback-driven and
snapshot-driven paths, making session and task wiring across xet_data,
xet_pkg, hf_xet, and git_xet harder to keep consistent. This PR moves
upload/download progress to a polling snapshot model backed by atomics.
It also switches task identifiers to a UniqueID common with the progress
tracking throughout the session APIs.
This PR also updates the rate estimation to use the lighter weight
exponentially weighted moving averages model, so this can be done at a
low level.
To preserve compatibility for existing callback consumers,
callback-oriented upload/download progress tracking APIs are moved under
xet_pkg::legacy and bridged from polling snapshots via a callback based
updaters. hf_xet and git_xet are updated to use that legacy bridge
layer, so current integrations keep working until everything is fully
switched over to the XetSession method.
## Summary
- Rewrites smoke tests to drive everything through the `hf` CLI rather
than the huggingface_hub Python API, covering the actual user-facing
surface area of hf-xet
- Moves smoke tests and diagnostic scripts into a `scripts/` directory
for cleaner repo layout
- Adds storage bucket test suite exercising the full bucket lifecycle
- Adds 50 MB and 100 MB files to repo upload/download tests
## Test matrix (14 tests, all passing)
**Repository tests** (`hf upload` / `hf download`)
- Upload single file, upload folder
- Download individual files + SHA-256 verify
- Download entire repo + SHA-256 verify
- Overwrite file and verify new content served
- Delete file and confirm absent
**Bucket tests** (`hf buckets`)
- `cp` upload / download + verify
- `sync` upload / download + verify
- Recursive list confirms expected paths
- Overwrite via `cp` + verify
- `sync --delete` removes extraneous remote files
- `rm` + confirm absent from listing
## Test plan
- [x] Run `HF_TOKEN=... ./scripts/smoke_tests/run.sh` and confirm all 14
tests pass
- [x] Run `./scripts/smoke_tests/run.sh --skip-buckets` for repo-only
path
- [x] Run with `--hf-xet-version <version>` to confirm PyPI cache bypass
works
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This PR moves some config values that were part of the data
configuration into XetConfig, specifically the compression_policy,
staging_subdir, session_dir_name, and global_dedup_query_enabled. This
also consolidates the remaining values into a single struct with
endpoint and authentication information.
Currently, the SHA-256 hash of uploaded file content is computed
internally during the upload pipeline but not surfaced to callers.
Downstream consumers — e.g. OpenDAL's Hugging Face backend — need the
SHA-256 to commit files to the Hub API.
This PR adds an optional sha256 field to XetFileInfo, the session-layer
FileMetadata, and the Python-exposed PyXetUploadInfo. The field is
populated from the already-computed hash when Sha256Policy::Compute or
Sha256Policy::Provided is used, and left None for downloads and when
Sha256Policy::Skip is used. Serde attributes (default,
skip_serializing_if) ensure backward-compatible serialisation — existing
serialised data without the field deserialises cleanly.
Needed for the functionality in
https://github.com/huggingface/xet-core/pull/642.
Currently, the Queued → Running status transition in spawned upload
tasks is unconditional — it overwrites whatever the current status is,
including Cancelled set by a concurrent abort() call. This creates a
race window: if abort() sets Cancelled between the semaphore acquisition
and the status write, the task overwrites it with Running, then the
completion guard (if matches!(*s, TaskStatus::Running)) passes and sets
Completed. The result is a task that was aborted but reports Completed.
This PR makes the Queued → Running transition conditional, matching the
already-guarded Running → Completed/Failed transition. If the status is
no longer Queued when the task starts, it bails early with
SessionError::Aborted. This closes the race window — all three status
transitions are now properly guarded against concurrent abort().
This was observed as a flaky test failure on Windows CI
(test_abort_while_state_lock_held_skips_state_update_but_drains_tasks).
This PR introduces V2 multirange URL fetching for xorbs, but optionally
splits the multirange requests into multiple single-range requests that
can be executed in parallel. This allows the reconstruction process to
generate full multirange presigned URLs, but the client effectively
performs the retrieval stage as a sequence of parallel single-range
queries.
The config variable `client.enable_multirange_fetching` controls this
behavior; by default it is set to false due to the current observed
slowness of fetching multiranged URLs.
---------
Co-authored-by: Adrien <adrien@huggingface.co>
Currently, `UploadCommitSync` and `DownloadGroupSync` are thin wrappers
around `UploadCommit` and `DownloadGroup` that delegate every method
through `external_run_async_task`. This means two types, two sets of doc
comments, and two test suites covering the same underlying behavior.
This PR removes the separate sync types and adds `_blocking` suffixed
methods directly on `UploadCommit` and `DownloadGroup`. The session
factory methods `new_upload_commit_blocking()` and
`new_download_group_blocking()` now return the same types as their async
counterparts, and the entire `xet_session::sync` module is deleted (~680
lines removed).
This also fixes a minor bug: `UploadCommitSync::upload_from_path` did
not call `std::path::absolute()` on the file path before dispatching,
unlike the async version. The new `upload_from_path_blocking` includes
the `std::path::absolute()` call, matching the async version's behavior.
This PR adds in a feature flag, "python" to the xet_runtime package such
that when compiled, the XetConfig struct is built to have python getters
and setters. This integrates the handling of the config struct directly
into the XetConfig struct and the macros used to register the config
values, making the handling of values in the python bindings seamless.
Stacking on top of https://github.com/huggingface/xet-core/pull/694,
this updates both the async and sync APIs: update
`XetSession::UploadCommit` and `XetSession::UploadCommitSync` pub APIs
to explicitly take a `sha256: Sha256Policy` to control whether to
compute and embed a sha256 for that file in the upload info.
Resolves XET-898
---------
Co-authored-by: Hoyt Koepke <hoytak@huggingface.co>
It's a bit annoying to try to ensure our CDN routing is correct. Logging
the URL domain for the first fetch term download to the debug logs.
Please don't hesitate to recommend alternative approaches.
`XetSession` always created its own tokio runtime via
`XetRuntime::new_with_config`, and calling `external_run_async_task`
panics when already inside a tokio context. This blocked embedding the
session in async Rust frameworks.
Core strategy:
- `RuntimeMode` enum —
`Owned` (session created its own thread pool via
`XetSessionBuilder::build` or `XetSessionBuilder::build_async` when
outside tokio context. Both `_blocking` and async methods are supported.
Async methods use an internal `bridge_to_owned` bridge that routes
futures onto the owned thread pool, so they work from any executor
(tokio, smol, async-std))
vs
`External` (session wraps a caller-supplied tokio handle via
`XetSessionBuilder::with_tokio_handle` or
`XetSessionBuilder::build_async` when inside qualified tokio context.
Only async methods may be called; `_blocking` methods return
`SessionError::WrongRuntimeMode`. No second thread pool is created).
- `XetRuntime::bridge_to_owned` — a new bridge that routes a future onto
the owned tokio thread pool from any executor (smol, async-std,
futures::executor, non-qualified tokio runtime) by delivering the result
via a `tokio::sync::oneshot` channel that can be polled by any async
executor.
- Async public API — `UploadCommit` and `DownloadGroup` methods
(`upload_from_path`, `upload_bytes`, `upload_file`, `commit`, `finish`)
are now async fn. Factory methods `XetSession::new_upload_commit` and
`new_download_group` are async.
Example:
```
let session = XetSessionBuilder::new().build_async().await?;
// Upload
let commit = session.new_upload_commit().await?;
let handle = commit.upload_from_path("file.bin".into()).await?;
let results = commit.commit().await?;
// Download
let group = session.new_download_group().await?;
let info = XetFileInfo {
hash: ...,
file_size: ...,
};
let dl_handle = group.download_file_to_path(info, "out/file.bin".into())?;
let finish_results = group.finish().await?;
```
- Sync wrappers — New `UploadCommitSync` / `DownloadGroupSync` in
`xet_session/sync/` expose a fully blocking API for sync Rust and Python
(PyO3) callers. Returned by `new_upload_commit_blocking()` and
`new_download_group_blocking()`.
Example:
```
let session = XetSessionBuilder::new().build()?;
// Upload
let commit = session.new_upload_commit_blocking()?;
let handle = commit.upload_from_path("file.bin".into())?;
let results = commit.commit()?;
let m = results.values().next().unwrap().as_ref().as_ref().unwrap();
// Download
let group = session.new_download_group_blocking()?;
let info = XetFileInfo {
hash: ...,
file_size: ...,
};
let dl_handle = group.download_file_to_path(info, "out/file.bin".into())?;
let finish_results = group.finish()?;
```
Additional fixes: `download_file_to_path` and `upload_from_path` now
canonicalize paths with `std::path::absolute` before enqueuing; task
status is only overwritten when still `Running`, preventing a race with
concurrent abort().
Fix XET-891
---------
Co-authored-by: Hoyt Koepke <hoytak@huggingface.co>
Currently, the condition for increasing connection concurrency is gated
on the model predicting that a 64MB transmission will complete within 90
seconds. However, when the transmissions are primarily composed of small
packets, this can drastically overestimate the round trip, artificially
suppressing the connection concurrency.
This PR fixes this issue by also modeling the average predicted packet
size, using the 95% quantile of that (bounded by two config variables)
to predict the round trip time when considering a concurrency increase.
## Summary
- Bump hf_xet version from 1.4.1 to 1.4.2 in Cargo.toml and Cargo.lock
- Follows up on 1.4.1 release where the version bump PR was merged after
the release artifacts were built
This PR updates the interface for retrieving per-task results after
UploadCommit::commit() or DownloadGroup::finish(). The problem with the
previous interface is that commit() and finish() return a vector of
FileMetadata or DownloadResult, making it difficult for users to
associate each result with a specific task.
The new interface uses `task_id` as a strong binding bridge:
## Upload per-task result access patterns
After commit() completes, there are two equivalent ways to retrieve a
per-task FileMetadata result:
1. Lookup in the global result map:
```
let commit = session.new_upload_commit()?;
let handle = commit.upload_from_path(src)?;
let results = commit.commit()?;
let result = results.get(&handle.task_id)
```
2. Direct access from the handle:
```
let commit = session.new_upload_commit()?;
let handle = commit.upload_from_path(src)?;
commit.commit()?;
// handle.result() is populated by commit() via the shared Arc.
let result = handle.result()
```
## Download per-task result access patterns
The pattern is similar to the above.
## Why not put results in a vector in the same order as tasks are
registered to the commit instance?
After a commit instance is created, it can be cloned (since it is itself
an Arc wrapping an internal struct) and sent to different threads. When
multiple threads are registering tasks, there is no static registration
order that a program can observe upfront.
This PR creates a folder, api_changes, in which AI agents can record
updates to the API surface that could affect downstream PRs and
dependencies. This can be scanned by AI agents to reliably perform
merges or to propagate changes. See api_changes/README.md for a
description of how this should work.
This PR is a massive rearrangement of the code base into 5 packages
intended for release on cargo. The directories and corresponding
packages are:
1. xet_runtime/ — compiles into the xet-runtime package. Contains the
runtime, config, and logging management.
2. xet_core_structures/ — compiles into the xet-core-structures package.
Contains core data structures for hashing, shards, and xorbs as well as
internal data structures that depend on these.
3. xet_client/ — compiles into the xet-client package, contains client
code for remotely connecting to the Hugging Face servers.
4. xet_data/ — compiles into the xet-data package, contains the data
processing pipeline: chunking/deduplication, file reconstruction,
clean/smudge operations, and progress tracking.
5. xet_pkg/ — compiles into the hf-xet package, provides the top-level
session-based API for file upload and download with user-facing error
categorization. This is the primary package downstream dependencies
would use. This also contains a single summary error type, XetError,
that translates cleanly into python error types.
In addition, the other tools are:
- git_xet/ — the git_xet CLI binary crate (location preserved).
- hf_xet/ -- the hf_xet python package (location preserved).
- simulation/ — the simulation crate for upload scenario benchmarking.
- wasm/ -- the wasm objects.
The full description — and information for an AI agent to use to update
downstream dependencies — is at
api_changes/update_260309_package_restructure.md.
Summary of moves:
- xet_runtime: became xet_runtime::core inside xet_runtime/.
- utils: became xet_runtime::utils inside xet_runtime/.
- xet_config: became xet_runtime::config inside xet_runtime/.
- xet_logging: became xet_runtime::logging inside xet_runtime/.
- error_printer: became xet_runtime::error_printer inside xet_runtime/.
- file_utils: became xet_runtime::file_utils inside xet_runtime/.
- merklehash: became xet_core_structures::merklehash inside
xet_core_structures/.
- mdb_shard: became xet_core_structures::metadata_shard inside
xet_core_structures/.
- xorb_object: became xet_core_structures::xorb_object inside
xet_core_structures/.
- cas_client: became xet_client::cas_client inside xet_client/.
- hub_client: became xet_client::hub_client inside xet_client/.
- cas_types: became xet_client::cas_types inside xet_client/.
- chunk_cache: became xet_client::chunk_cache inside xet_client/.
- data: became xet_data::processing inside xet_data/.
- deduplication: became xet_data::deduplication inside xet_data/.
- file_reconstruction: became xet_data::file_reconstruction inside
xet_data/.
- progress_tracking: became xet_data::progress_tracking inside
xet_data/.
- xet_session: became xet::xet_session inside xet_pkg/.
- Wasm packages (hf_xet_wasm, hf_xet_thin_wasm): moved from top-level
into wasm/; internal imports updated, public APIs unchanged.
Fixes
[XET-885](https://linear.app/xet/issue/XET-885/investigate-unsloth-upload-failure-shard-upload-timeout-on-cas)
## Summary
Shard uploads to CAS can take a long time due to server-side processing
(DynamoDB writes scale with file entry count). The default
`read_timeout(120s)` on the reqwest client kills these uploads.
**Key insight:** reqwest's per-request `RequestBuilder::timeout()` does
NOT override the client-level `read_timeout()` — they are independent
mechanisms polled as separate futures. So the original approach of using
per-request timeouts was ineffective.
**Fix:** Create a dedicated `shard_upload_http_client` on `RemoteClient`
with **no `read_timeout`**, built once at construction time and reused
for all shard uploads. All other settings (connect timeout, pool config,
auth middleware) are identical to the standard client.
## Changes
### `cas_client/src/http_client.rs`
- Added `reqwest_client_no_read_timeout()` — creates a reqwest client
with no `read_timeout`
- Added `build_auth_http_client_no_read_timeout()` — public API wrapping
it with middleware
- 4 unit tests for the new builder
### `cas_client/src/remote_client.rs`
- Added `shard_upload_http_client` field to `RemoteClient` (cfg'd out on
wasm)
- `upload_shard()` uses the pre-built no-timeout client instead of
building one per request
### `cas_client/tests/test_shard_upload_timeout.rs`
- Updated: slow server test now asserts **success** (shard uploads
should wait as long as needed)
### `xet_config/src/groups/client.rs`
- Removed `shard_read_timeout` config field (no longer needed)
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
## Summary
Fixes download stalls/deadlocks on large file reconstruction (reported
on 48.5 GB GGUF files). The root cause is a circular dependency: the
main reconstruction loop holds a buffer semaphore permit while blocking
on CAS connection permit acquisition, and xorb write locks held during
HTTP downloads cause CAS permit starvation.
### Changes
1. **Single-flight xorb downloads via `OnceCell`** (`xorb_block.rs`):
replaces `RwLock<Option<...>>` with `tokio::sync::OnceCell`. Only one
task per xorb block acquires a CAS permit and downloads the data;
concurrent callers wait on the same result without acquiring permits or
duplicating work. This eliminates duplicate downloads, prevents
double-counted transfer progress, and avoids a failing duplicate from
killing the reconstruction.
2. **Decouple CAS permit from buffer permit** (`file_term.rs`): the main
loop no longer blocks on CAS permits while holding a buffer permit. The
spawned download task delegates to `retrieve_data` which handles permit
acquisition internally via the OnceCell single-flight. This breaks the
circular dependency that causes stalls.
3. **Improve error propagation** (`sequential_writer.rs`): when the
background writer channel closes, check `RunState` for the original
error before returning a generic "channel closed" message.
### Root cause
The reconstruction pipeline has three resource pools: buffer permits
(bounded semaphore), CAS download permits (64 concurrent), and per-xorb
write locks.
Before this fix, the main loop would:
1. Acquire a **buffer permit** (blocking if buffer full)
2. Call `get_data_task()` which acquires a **CAS permit** (blocking if
pool exhausted)
3. Inside `retrieve_data()`, hold a **write lock** during the entire
HTTP download
This creates two deadlock vectors:
- **Buffer vs CAS**: buffer fills up with terms waiting for CAS permits,
but CAS permits are held by tasks blocked behind xorb write locks, and
the writer can't drain the buffer because it's waiting for those tasks
- **CAS vs write lock**: multiple tasks sharing the same xorb each hold
a CAS permit while blocked on the write lock, starving other xorbs of
permits
## Reproduction
Reliably reproducible with small buffer:
```
HF_XET_RECONSTRUCTION_DOWNLOAD_BUFFER_SIZE=64mb \
HF_XET_RECONSTRUCTION_DOWNLOAD_BUFFER_LIMIT=64mb \
python3 -c "from huggingface_hub import hf_hub_download; hf_hub_download('unsloth/Qwen3-Coder-Next-GGUF', 'Qwen3-Coder-Next-Q4_K_M.gguf', local_dir='/tmp/test')"
```
- **Before fix**: stalls at ~3.4 GB, no progression (deadlock)
- **After fix**: continuous progression, completes successfully
With default buffer (2 GB), the stall is intermittent depending on
network speed (consistently reproduced on slower connections).
## Summary
- Add `ShaGenerator::Skip` variant that skips SHA-256 computation
entirely
- `ShaGenerator::finalize()` now returns `Option<Sha256>` (None when
skipped)
- `SingleFileCleaner::new()` and `FileUploadSession::start_clean()`
accept a `skip_sha256` boolean
- When skipped, no `FileMetadataExt` is included in the shard
## Context
Bucket uploads don't need SHA-256 in the shard metadata — the
`sha_index` GSI is only used for LFS pointer resolution, which doesn't
apply to buckets. Skipping SHA-256 for bucket uploads removes the main
CPU bottleneck in the upload pipeline on non-SHA-NI instances.
## Alternative: dummy SHA-256
Instead of skipping entirely, the client could send a zeroed/dummy
`FileMetadataExt`. The server would still store it but queries would
never match. This avoids the server-side schema change (xetcas PR) but
pollutes the GSI with dummy entries.
Companion PRs:
- xetcas: huggingface-internal/xetcas#498 (make `FileIdItem.sha256`
optional server-side)
This PR replaces the previous collection of scripts around setting up
docker containers with a much more nimble and lightweight set of rust
scripts and a simple, reusable proxy that can limit bandwidth and
congestion simulations. The previous scripts are rewritten to be more
nimble and use more reusable components.
New tools:
- cas_client/src/simulation/network_simulation: A lightweight,
in-process network congestion simulation proxy that lives between the
LocalServer instance and the RemoteClient instance, allowing simulation
tests to run on a network with realistic congestion conditions and a
gated bandwidth. This can be controlled dynamically through a
LocalTestServer instance.
- simulation/: A new package for collecting simulation scripts and
analyzing the results.
To run the new simulation scripts for the adaptive concurrency on
upload, compile in release mode and run one of the scripts in
`simulation/src/adaptive_concurrency/scripts/`. Docker is no longer
needed to run any of the simulations.
The old `cas_client/tests/adaptive_concurrency/` paths were removed.
This PR adds interface functions to the LocalServer class that will
allow it to become a full simulation environment for testing all the
garbage collection stages.
Currently, the async stream logic silently swallows an UnexpectedEOF,
treating it the same as an EOF. This is a bug; this PR fixes it to
propagate UnexpectedEOF while handling correct EOF as the end of the
stream.
This PR makes the use of the `cas` and `xorb` terms consistent.
Previously, "cas" (for content addressed store) could simultaneously
refer to either the remote server or the data bytes stored as a
collection of chunks. After the renames in this PR, we consistently use
`xorb` to refer to the data object and cas to refer to the remote
server.
This renames quite a few places; to aid in rebasing current work or
updating downstream dependencies, this PR includes a file
`API_UPDATES.md` that can be fed into an AI agent to quickly and
accurately perform the renaming on any downstream dependencies.
This PR introduces a new `xet_session` crate that provides a
session-based hierarchical API: Users create a XetSession to manage
runtime and configuration, then batch uploads into UploadCommit objects
and downloads into DownloadGroup objects — each of which runs transfers
in the background by the inner XetRuntime.
All pub functions are exposed as sync functions - making them easy to
use in other languages, e.g. Python, C, etc.
## Summary
- Add optional `sha256s` keyword parameter to the Python-exposed
`upload_files()` function
- Forward it to `data_client::upload_async()` which already supports it
## Context
### Double computation today
`huggingface_hub` computes SHA-256 on every file during
`CommitOperationAdd.__post_init__()` for LFS batch negotiation, then
`hf_xet` recomputes it internally because `upload_files()` doesn't
accept pre-computed hashes.
### Performance impact
This change eliminates the redundant computation entirely.
### Backward compatibility
- `sha256s` is a keyword-only parameter with default `None` — no change
for existing callers
- `data_client::upload_async()` already accepts `sha256s:
Option<Vec<String>>` since day one
- When provided, `SingleFileCleaner` uses `ShaGenerator::ProvidedValue`
and skips internal recomputation
Companion PR: huggingface/huggingface_hub#3876
## Summary
- Fix command injection vulnerability in `.github/workflows/release.yml`
(HackerOne #3581567, severity High 8.8)
- `${{ github.event.inputs.tag }}` was interpolated directly in `run:`
blocks, allowing arbitrary RCE via crafted tag input (e.g. `v0.1.0; id;
cat /etc/passwd;#`)
- Moved all 6 occurrences to `env:` variables so the value is passed as
a shell environment variable instead of being interpolated into the
script
## Jobs fixed
- `linux` — "Update version in toml" step
- `musllinux` — "Update version in toml" step
- `windows` — "Update version in toml" step
- `macos` — "Update version in toml" step
- `sdist` — "Update version in toml" step
- `github-release` — "Create GitHub Release" step (`gh release create`)
This PR adds an integrated API for streaming downloads, exposing a
DownloadStream object that is integrated with the file reconstructor. It
also uses the same memory management buffer limiting process to work
with the stream object.
It also introduces cancellation support to the FileReconstructor to
ensure that tasks waiting on a long running download or semaphore wait
don't cause things to hang when an error is reported or the user drops
the stream.
Introduces a client benchmark utility to track system resource usage
(CPU, memory, disk I/O, and network I/O) of a process, so we don't need
to write scripts to capture usage stats according to different OS
standards. This becomes extremely helpful when I benchmark on Python
notebook instances, e.g. Google Colab, where system monitor is not
easily accessible or when running a separate monitor script is not easy.
# Usage #
Users can enable monitoring by setting `HF_XET_SYSTEM_MONITOR_ENABLED`
to true, set usage sample interval using
`HF_XET_SYSTEM_MONITOR_SAMPLE_INTERVAL`, this outputs metrics to the
tracing stream at `INFO` level by default. In addition, these metrics
can be redirected to a separate file by setting sample log path using
`HF_XET_SYSTEM_MONITOR_LOG_PATH`.
# Output #
The stats are output in JSON format, which can be queried using tools
like `jq`, e.g.
1. Trace of peak memory usage: `jq '.memory.peak_used_bytes'
[HF_XET_SYSTEM_MONITOR_LOG_PATH]`
2. Trace of disk write speed: `jq '.disk.average_write_speed'
[HF_XET_SYSTEM_MONITOR_LOG_PATH]`
3. Trace of network receive speed: `jq '.network.average_rx_speed'
[HF_XET_SYSTEM_MONITOR_LOG_PATH]`
Currently, the maximum number of downloaded files is fixed, regardless
of the number of downloads currently in flight. However, as the number
of downloads increases, a fixed size total could lead to waiting on
individual segments that download out-of-order or don't have enough
turnaround time to saturate the output. While writing to disk or the
download itself often becomes the bottleneck before these effects,
planned features such as streaming files and caching could be affected
by this limit. The default formula for the download buffer size now is
(2GB + 512MB * number of concurrent downloads) up to a maximum of 8GB
(these are adjustable).
This PR alleviates this by allocating an additional 512MB buffer
allocation per file, prioritized to the specific download, releasing
that capacity when the file finishes downloading. This is done using the
AdjustableSemaphore class, first introduced for the concurrent scaling,
which allows the number of total permits in a semaphore to be
incremented or decremented; on decrement, permits are discarded upon
return until the total permits is at the target number.
This PR addresses two rare but occasional test failures on windows, both
due to window's non-synchronous file system behavior.
- A race condition opening the local test database causing an error.
- Unwanted cleanup conditions in testing the log preservation can
trigger if the test execution is stretched out long enough.
- A null-termination bug in set_file_metadata that causes it to fail
silently if the memory layout is such a way that the string passed in
isn't null-terminated. This causes occasional failures in setting the
metadata time on linux.
## Summary
Wrap download progress updaters in `AggregatingProgressUpdater` to
eliminate GIL contention when Python callers provide per-file progress
callbacks.
The upload path has had this aggregation since v1.1.3 (PR #340), but the
download path was missed. Without aggregation, each XORB chunk triggers
a `spawn_blocking` + `Python::with_gil()` callback. With many concurrent
file downloads, this causes severe GIL contention — measured as a **4x
throughput reduction** (3000 MB/s → 750 MB/s on a 25 Gbps link).
The fix wraps the caller-provided `TrackingProgressUpdater` in an
`AggregatingProgressUpdater` (200ms flush interval) inside
`download_file_with_updater()`, matching the pattern already used by
`FileUploadSession`. This reduces Python callback frequency from
thousands/sec to ~5/sec per file while preserving progress bar feedback.
## Root cause
When `huggingface_hub` calls `hf_xet.download_files()`, it passes a
per-file Python callback for progress bar updates. On the Rust side,
each callback invocation goes through:
```
report_bytes_written() / report_transfer_progress()
→ tokio::spawn(register_updates())
→ spawn_blocking(Python::with_gil(callback))
```
With the detailed download progress tracking added in PR #645 (hf-xet
v1.3.0), both `report_bytes_written` and `report_transfer_progress` fire
per chunk, roughly doubling callback frequency. With 8+ concurrent file
downloads, each spawning dozens of concurrent XORB streams, the GIL
becomes a severe bottleneck.
## History
The problem has existed since xet download support was introduced, but
worsened over time:
| Version | Date | Impact |
|---------|------|--------|
| `huggingface_hub v0.30.0` / `hf-xet 0.1.x` | Mar 2025 | Moderate —
synchronous `with_gil()` per chunk, but hf_xet was an optional extra |
| `huggingface_hub v0.31.0` / `hf-xet >=1.1.0` | May 2025 | Moderate —
hf-xet became a hard dependency on x86_64/arm64 |
| `hf-xet v1.1.3` | Jun 2025 | Upload path fixed with
`AggregatingProgressUpdater` (PR #340); download path left unprotected |
| `hf-xet v1.3.0` | Feb 2026 | **Severe** — PR #645 added detailed
per-chunk progress tracking to downloads, doubling callback frequency
without aggregation |
PR #340 explicitly noted: *"each [update] has to acquire a global GIL
lock. This negatively affects the upload speed on fast connections"* —
the same problem, but only the upload side was addressed.
## Benchmarks
Downloading 3 safetensors files (16.1 GB total) from
`Qwen/Qwen3.5-35B-A3B` on a 25 Gbps machine:
| Test | Before | After |
|------|--------|-------|
| `download_files()` with `progress_updater=None` (baseline) | 3119 MB/s
| 3119 MB/s |
| `download_files()` with per-file Python callbacks | **746 MB/s** |
**1789 MB/s** |
| `snapshot_download()` (full Python CLI path with tqdm) | ~750 MB/s |
**2395 MB/s** |
Progress callback overhead drops from **4x slowdown to <1%**.
Bumps [time](https://github.com/time-rs/time) from 0.3.44 to 0.3.47.
<details>
<summary>Release notes</summary>
<p><em>Sourced from <a
href="https://github.com/time-rs/time/releases">time's
releases</a>.</em></p>
<blockquote>
<h2>v0.3.47</h2>
<p>See the <a
href="https://github.com/time-rs/time/blob/main/CHANGELOG.md">changelog</a>
for details.</p>
<h2>v0.3.46</h2>
<p>See the <a
href="https://github.com/time-rs/time/blob/main/CHANGELOG.md">changelog</a>
for details.</p>
<h2>v0.3.45</h2>
<p>See the <a
href="https://github.com/time-rs/time/blob/main/CHANGELOG.md">changelog</a>
for details.</p>
</blockquote>
</details>
<details>
<summary>Changelog</summary>
<p><em>Sourced from <a
href="https://github.com/time-rs/time/blob/main/CHANGELOG.md">time's
changelog</a>.</em></p>
<blockquote>
<h2>0.3.47 [2026-02-05]</h2>
<h3>Security</h3>
<ul>
<li>
<p>The possibility of a stack exhaustion denial of service attack when
parsing RFC 2822 has been
eliminated. Previously, it was possible to craft input that would cause
unbounded recursion. Now,
the depth of the recursion is tracked, causing an error to be returned
if it exceeds a reasonable
limit.</p>
<p>This attack vector requires parsing user-provided input, with any
type, using the RFC 2822 format.</p>
</li>
</ul>
<h3>Compatibility</h3>
<ul>
<li>Attempting to format a value with a well-known format (i.e. RFC
3339, RFC 2822, or ISO 8601) will
error at compile time if the type being formatted does not provide
sufficient information. This
would previously fail at runtime. Similarly, attempting to format a
value with ISO 8601 that is
only configured for parsing (i.e. <code>Iso8601::PARSING</code>) will
error at compile time.</li>
</ul>
<h3>Added</h3>
<ul>
<li>Builder methods for format description modifiers, eliminating the
need for verbose initialization
when done manually.</li>
<li><code>date!(2026-W01-2)</code> is now supported. Previously, a space
was required between <code>W</code> and <code>01</code>.</li>
<li><code>[end]</code> now has a <code>trailing_input</code> modifier
which can either be <code>prohibit</code> (the default) or
<code>discard</code>. When it is <code>discard</code>, all remaining
input is ignored. Note that if there are components
after <code>[end]</code>, they will still attempt to be parsed, likely
resulting in an error.</li>
</ul>
<h3>Changed</h3>
<ul>
<li>More performance gains when parsing.</li>
</ul>
<h3>Fixed</h3>
<ul>
<li>If manually formatting a value, the number of bytes written was one
short for some components.
This has been fixed such that the number of bytes written is always
correct.</li>
<li>The possibility of integer overflow when parsing an owned format
description has been effectively
eliminated. This would previously wrap when overflow checks were
disabled. Instead of storing the
depth as <code>u8</code>, it is stored as <code>u32</code>. This would
require multiple gigabytes of nested input to
overflow, at which point we've got other problems and trivial
mitigations are available by
downstream users.</li>
</ul>
<h2>0.3.46 [2026-01-23]</h2>
<h3>Added</h3>
<ul>
<li>All possible panics are now documented for the relevant
methods.</li>
<li>The need to use <code>#[serde(default)]</code> when using custom
<code>serde</code> formats is documented. This applies
only when deserializing an <code>Option<T></code>.</li>
<li><code>Duration::nanoseconds_i128</code> has been made public,
mirroring
<code>std::time::Duration::from_nanos_u128</code>.</li>
</ul>
<!-- raw HTML omitted -->
</blockquote>
<p>... (truncated)</p>
</details>
<details>
<summary>Commits</summary>
<ul>
<li><a
href="d5144cd287"><code>d5144cd</code></a>
v0.3.47 release</li>
<li><a
href="f6206b050f"><code>f6206b0</code></a>
Guard against integer overflow in release mode</li>
<li><a
href="1c63dc7985"><code>1c63dc7</code></a>
Avoid denial of service when parsing Rfc2822</li>
<li><a
href="5940df6e72"><code>5940df6</code></a>
Add builder methods to avoid verbose construction</li>
<li><a
href="00881a4da1"><code>00881a4</code></a>
Manually format macros everywhere</li>
<li><a
href="bb723b6d82"><code>bb723b6</code></a>
Add <code>trailing_input</code> modifier to <code>end</code></li>
<li><a
href="31c4f8e0b5"><code>31c4f8e</code></a>
Permit <code>W12</code> in <code>date!</code> macro</li>
<li><a
href="490a17bf30"><code>490a17b</code></a>
Mark error paths in well-known formats as cold</li>
<li><a
href="6cb1896a60"><code>6cb1896</code></a>
Optimize <code>Rfc2822</code> parsing</li>
<li><a
href="6d264d59c2"><code>6d264d5</code></a>
Remove erroneous <code>#[inline(never)]</code> attributes</li>
<li>Additional commits viewable in <a
href="https://github.com/time-rs/time/compare/v0.3.44...v0.3.47">compare
view</a></li>
</ul>
</details>
<br />
[](https://docs.github.com/en/github/managing-security-vulnerabilities/about-dependabot-security-updates#about-compatibility-scores)
Dependabot will resolve any conflicts with this PR as long as you don't
alter it yourself. You can also trigger a rebase manually by commenting
`@dependabot rebase`.
[//]: # (dependabot-automerge-start)
[//]: # (dependabot-automerge-end)
---
<details>
<summary>Dependabot commands and options</summary>
<br />
You can trigger Dependabot actions by commenting on this PR:
- `@dependabot rebase` will rebase this PR
- `@dependabot recreate` will recreate this PR, overwriting any edits
that have been made to it
- `@dependabot show <dependency name> ignore conditions` will show all
of the ignore conditions of the specified dependency
- `@dependabot ignore this major version` will close this PR and stop
Dependabot creating any more for this major version (unless you reopen
the PR or upgrade to it yourself)
- `@dependabot ignore this minor version` will close this PR and stop
Dependabot creating any more for this minor version (unless you reopen
the PR or upgrade to it yourself)
- `@dependabot ignore this dependency` will close this PR and stop
Dependabot creating any more for this dependency (unless you reopen the
PR or upgrade to it yourself)
You can disable automated security fix PRs for this repo from the
[Security Alerts
page](https://github.com/huggingface/xet-core/network/alerts).
</details>
Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
## Summary
Closes#588
- Add `win11-arm` runner with `aarch64-pc-windows-msvc` target to the
hf-xet Python wheel release pipeline
- Add `win11-arm` runner with `aarch64` target to the git-xet CLI
release pipeline, parameterizing the WiX installer `-arch` flag
## Test plan
- [x] Trigger a workflow_dispatch run of the Release workflow and verify
`windows` matrix includes both `x64` and `aarch64` entries
- [x] Verify ARM64 wheels and .pdb debug symbols are built and uploaded
- [ ] Trigger a workflow_dispatch run of the git-xet Release workflow
and verify ARM64 binary and MSI installer are produced
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Adding support for setting an optional `request_header` map on the
hf_xet upload and download API calls. This map is augmented with the
hf_xet user agent string and is passed along with the requests to
xetcas.
This PR also adds some unit tests for testing the map merging behavior
to `hf_xet/lib.rs` and adds support for running these with cargo test
and in github actions CI step.
Currently, the progress and dependency tracking in the upload path
requires that the total size of a file be specified at the start. This
PR changes this so that in cases where the upload is streamed and the
total size is not known, it's updated as soon as new data is processed.
Both routes now work and correctly track the file sizes.
This PR adds a FileDownloadSession struct that parallels the
FileUploadSession struct, replacing the FileDownloader. It's an
intermediate step in preparation for a session-based API that integrates
well with interfaces other than the python interface in hf_xet.
This PR adds detailed progress reporting to the download path.
- Transfer progress is reported as soon as the download streams start;
actual bytes written are reported as the reconstructed file is written
out.
- Currently, each call to download_file creates a separate progress
tracker, but this sets up for download groups with grouped download
progress tracking.
To support this, the UploadProgressStream was split into three classes;
a common StreamProgressReporter and download and upload specific
versions. This also allows us to simplify the API to RetryWrapper.
More tracking was added to the file reconstruction paths to properly
report progress.
