fix: stop HuggingFace rate-limit failures from breaking voice engine downloads (#683)#704
fix: stop HuggingFace rate-limit failures from breaking voice engine downloads (#683)#704shreeraman96 wants to merge 1 commit into
Conversation
Clearing the provider cache and re-preparing on a transient 429 or network failure forces a full re-download, which burns more of HuggingFace's fixed per-IP API window and turns one rate limit into a self-reinforcing failure loop. Classify rate-limit/HTML/network errors (walking the underlying-error chain) and rethrow them instead, leaving cached files resumable.
Greptile SummaryPrevents transient Hugging Face rate-limit and network errors from triggering destructive cache clearing by classifying wrapped download errors and rethrowing them so partial model downloads remain resumable.
|
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 63fac7d836
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| if let hfError = candidate as? DownloadUtils.HuggingFaceDownloadError { | ||
| switch hfError { | ||
| case .rateLimited, .htmlErrorResponse: |
There was a problem hiding this comment.
In the Intel build path, ASRService.swift does not import FluidAudio because the import at the top is behind #if arch(arm64), and the repo provides Intel stubs for FluidAudio-backed providers so Whisper/Apple Speech can still build. This new unconditional reference to DownloadUtils.HuggingFaceDownloadError therefore leaves DownloadUtils out of scope for x86_64 macOS builds, breaking the Intel support path; wrap this branch in the same architecture/import guard or avoid the FluidAudio type outside arm64 code.
Useful? React with 👍 / 👎.
1 participant
Description
ASRService.prepareProviderWithRecoverywhen the first prepare attempt failed due to HuggingFace rate limiting (rateLimited,htmlErrorResponse) or a network error, rethrowing instead so cached model files stay resumable.NSUnderlyingErrorKeychain (depth-capped) so errors reboxed by provider layers are still recognized; cancellation is excluded and keeps its existing handling.Why: clearing the provider cache on a transient 429 forces a full re-download, which burns more of HuggingFace's fixed 500-requests/5-minutes per-IP API window and turns one rate limit into a self-reinforcing failure loop where every engine's download fails ("Rate limited while listing files"). Root fix lands in FluidAudio (altic-dev/FluidAudio#2 — single recursive listing, header-aware retries, no cache wipe on rate limits); this PR removes the app-level amplifier. The app picks up the FluidAudio fix automatically once #2 merges into the pinned
B/cohere-coreml-asrbranch.Repro evidence (offline stub HuggingFace server, real Parakeet v3 tree): before the fixes, one download gesture cost 41 listing API calls, any 429 failed instantly with the exact issue error, and a mid-flight failure wiped the cache and re-listed (a 45-call budget that fit one full pass still blew up at call #46). After the fixes: 1 listing call, downloaded file set byte-identical (23/23), sustained 429s produce header-paced retries then a clean error with the cache intact, and corrupt-model recovery still works. Full tables in altic-dev/FluidAudio#2.
Type of Change
Related Issue or Discussion
Closes #683
Testing
swiftlint --strict --config .swiftlint.yml Sources(0 violations on the changed file)swiftformat --config .swiftformat Sources(changed file already fails formatter lint onmain; not reformatting ~3k unrelated lines in this PR)DownloadUtilsRateLimitTests15/15 pass; app built via./build.sh unsigned; behavior verified end-to-end with an offline stub HuggingFace server (details in fix: eliminate HuggingFace rate-limit failures in model downloads FluidAudio#2)Screenshots / Video
Notes
Companion root-cause PR: altic-dev/FluidAudio#2. This change is safe standalone (it only narrows a destructive retry), but the user-facing rate-limit failures in #683 fully disappear when both land.
The same dependency bump also fixes #679 (progress stuck at 3%, then completes all at once). Root cause, verified empirically: the FluidAudio revision pinned by v1.6.5 (
edee7154) attaches a progress delegate but downloads via the asyncsession.download(for:)convenience, which never deliversdidWriteData— so progress only ticks at file boundaries and freezes for the entire multi-hundred-MB encoder transfer. FluidAudio6af7925(merged toB/cohere-coreml-asrtwo days after the 1.6.5 pin) rewrites this with a delegate-backed download task. Measured with an offline stub (30 MB file): pinned revision = 23 progress events / 9 distinct values for a whole repo download; current branch incl. FluidAudio#2 = 1,142 events / 73 distinct values (continuous). Re-pinning FluidAudio to the branch head once FluidAudio#2 merges resolves both #683 and #679.