[KeyanHu-git]

Problem and result

When HEAD and the one-byte range probe both fail to provide a length, download() treats an early EOF as completion even if Strata already has the file's published size. A short file is renamed out of .part, and verification either stops with a size error or, without a published hash, marks the fragment as finished.

Use the published size for transfer completion as well as verification. The existing downloader then resumes an interrupted transfer, keeps partial bytes when retries run out, and publishes the file only after receiving the expected length.

Related discussion

#1915. Version 0.2 added a range probe, which covers missing HEAD length when that probe succeeds. This fix covers the remaining case where both probes lack a size. The original ModelScope URL supplied Content-Length when checked on October 22; the regression below reproduces the remaining condition with a local HTTP server.

Approach

After the transport probes, fall back to the repository pin's size, or ModelScope's published size when there is no pin. Reuse the existing metadata lookup for hash verification.

HEAD size → range-probe size → published size → existing download/resume logic
                                                    ↓
                                          size check → rename → hash check

The production diff moves the metadata lookup and adds a two-line fallback. Serial retries, parallel transfers, source identity and checksum verification retain their existing implementations.

Validation

Base: 011c9e60 (0.2.2 / engine 0.1.47). Final revision: 7527e00f. Python 3.12 on macOS and Windows.

python -m unittest tools.test_setup_download_length tools.test_setup_download -v

Passed: all 8 focused tests on both systems. They cover early EOF with a pin, metadata with and without a hash, exhausted retries preserving .part, parallel transfer after a failed initial probe, an already complete .part, a transport/pin size conflict and a wrong hash. Six of these fail on the unmodified base; the size-conflict and hash checks pass on both.

The early-EOF fixture supplies 262,144 of 2,633,785 bytes before closing. The patched serial transfer resumes at byte 262,144 and produces the exact original bytes. Exhausting retries preserves those 262,144 bytes without a final file or .done mark.

Failed: the combined suite passes 36/37 on both macOS and Windows. The shared redirect test counts the first piece twice ([0, 0, 1, …, 40] instead of [0, 1, …, 40]). It uses a successful HEAD length, so this fallback is not entered. The production code for that path is unchanged. The same 37 tests on unmodified main report 4 failures and 3 errors on both systems, including the six focused regressions above.

Not run: another multi-gigabyte transfer through the original reporter's network. All transfer tests above use real HTTP requests to a local server; published metadata is supplied by the fixture.

Limits

If neither transport probes nor published metadata provide a size, the existing unknown-length behavior remains. The fallback does not replace hash verification or change which model files Strata accepts.

[KeyanHu-git]

Hi @Niko1221, could you take a look at this fix for #1915 when you have a chance? The early-EOF regression cases reproduce the issue on main and pass with the patch on Windows.

I also submitted #2288 for the GGUF validation issues in #1919. I’d appreciate your thoughts on its handling of unknown tensor sizes: it reports that completeness cannot be verified and asks the user to keep the file, instead of counting the unknown size as zero.

Thanks for your time!

[Niko1221]

Thanks for the pull request.

We're resetting the tracker for 0.2.0-beta. There are more than 1,300 open issues and pull requests, too many for us to review properly, so we're closing everything that isn't being worked on right now. Closed here does not mean rejected.

To reopen it takes about two minutes:

1. Rebase on main if it no longer applies, and re-run your checks on 0.2.0-beta.
2. Start the title with its area: [Engine], [GUI], [TUI], [Benchmark], [Docs] or [Other].
3. Fill in the pull request template: the problem and result, and the validation (what you ran, on which revision, what you saw).
4. Comment /reopen. A check runs and either reopens the pull request or lists what is still missing.

The rules are in CONTRIBUTING.md.

[KeyanHu-git]

/reopen

[github-actions]

Reopened, thank you. A maintainer will pick it up from here. If it conflicts with main, rebase it first; that speeds up review.

[Niko1221]

Thanks. Confirmed: with no length from either probe, an early EOF is taken as complete. Under review for 0.2.7.

Release: https://github.com/Niko1221/Strata/releases/tag/v0.2.7