Staging validation: Studio CI fix bundle (pydantic + llama.cpp + change-password)#141
Staging validation: Studio CI fix bundle (pydantic + llama.cpp + change-password)#141danielhanchen wants to merge 2 commits into
Conversation
Three independent CI regressions across maintainer PRs: 1. Windows --no-torch install: pydantic + pydantic-core drift to incompatible versions under `uv pip install --no-deps -r no-torch-runtime.txt`, tripping pydantic's _ensure_pydantic_core_version() check on the next import. Resolve pydantic WITH deps in a separate pip call so pip picks a compatible (pydantic, pydantic-core) pair. pydantic's transitive deps are torch-free. Drop pydantic from no-torch-runtime.txt and delete the now-redundant `Patch Studio venv with full typer / pydantic dep trees` workaround in the four Windows smoke YAMLs. (PR unslothai#5733 + unslothai#5734) 2. Linux Studio Update CI: upstream llama.cpp b9261+ split each binary's entry code into a paired `libllama-<binary>-impl.so` shared library. `llama-server` and `llama-quantize` NEEDED-link against `libllama-server-impl.so` / `libllama-quantize-impl.so` with RUNPATH `$ORIGIN`, so the prebuilt overlay must copy those alongside the binaries. Add `libllama-*-impl.so*` to the Linux runtime patterns and lock the pattern in test_rocm_support.TestRuntimePatterns. 3. Mac Studio UI Chat test: change-password submit button was clicked while disabled because the gate only checked new + confirm password length. Gate also on current_password.length >= 8 and mirror the gate in the submit handler so Enter / autofill cannot bypass it. (PR unslothai#5738) Staging-fork-only validation workflows (Linux + Windows + macOS) exercise each fix end-to-end without burning credits on the main repo. The four pre-existing Studio workflows are removed from the staging fork so we stay below the 5-concurrent-Windows-runner cap.
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request addresses dependency versioning issues by installing pydantic with its dependencies separately to ensure compatibility with pydantic-core across different platforms. It also updates the prebuilt installer to include new llama.cpp shared libraries and enhances password validation in the frontend authentication form. Feedback was provided to refine the error message in the authentication form to better distinguish between bootstrap failures and missing user input for the current password.
| if (currentPassword.length < 8) { | ||
| setError( | ||
| currentPassword | ||
| ? "Current password must be at least 8 characters." | ||
| : "Unable to initialize setup. Reload the page and try again.", | ||
| ); | ||
| return; | ||
| } |
There was a problem hiding this comment.
The error message "Unable to initialize setup. Reload the page and try again." is shown when the current password is empty in change-password mode. While this might be appropriate if the bootstrap process failed, it is confusing for a user who simply left the "Current password" field empty when it is visible (!hasBootstrapPassword). Consider providing a more descriptive validation error when the field is user-facing.
if (currentPassword.length < 8) {
setError(
currentPassword
? "Current password must be at least 8 characters."
: hasBootstrapPassword
? "Unable to initialize setup. Reload the page and try again."
: "Current password is required.",
);
return;
}
…ller overlays the prebuilt
1 participant
Staging-fork validation PR for a maintainer-side CI fix bundle. Not for upstream merge — this lands the same diff that would go to unslothai/unsloth main, against the throwaway staging fork, with a minimal workflow set so we can verify on real GitHub-hosted Ubuntu / macOS / Windows runners without touching the main repo's CI budget.
Bundled fixes
Windows --no-torch install: pydantic-core mismatch. Every Studio Windows smoke fails at
unsloth studio setupwithSystemError: The installed pydantic-core version (2.47.0) is incompatible with the current pydantic version, which requires 2.46.4.uv pip install --no-deps -r no-torch-runtime.txtresolvespydanticandpydantic-coreindependently from latest. Resolve pydantic WITH deps in a focused pip call so pip pins pydantic-core to the version pydantic's metadata requires; pydantic's transitive deps are torch-free. Drop pydantic from no-torch-runtime.txt and remove the four redundantPatch Studio venv with full typer / pydantic dep treesworkarounds.Linux Studio Update CI: llama.cpp falls back to source build. Upstream b9261+ split each binary's entry code into a paired
libllama-<binary>-impl.soshared library.llama-serveris NEEDED-linked againstlibllama-server-impl.sowith RUNPATH\$ORIGIN. Without copying the impl libs alongside the binaries, ldd reports them missing, preflight rejects, the installer falls back to source build, andstudio-update-smokeannotatessetup.sh idempotency regressed. Addlibllama-*-impl.so*to the Linux runtime overlay pattern and lock it in TestRuntimePatterns.Mac Studio UI Chat: change-password submit gated on a stale state. Playwright's submit click was racing the React state update for the current-password field, hitting a disabled ``. Tighten the disable gate to also require `currentPassword.length >= 8` and mirror the same check in the submit handler so Enter / autofill cannot bypass.
Staging-fork CI surface
This branch removes all 26 pre-existing workflows on the staging fork and adds 3 narrow ones:
Each pins `runs-on:` to a single OS, uses `concurrency.cancel-in-progress: true`. Triggered on `push: branches: [ci/general-fix-2026-05-23]` with `paths:` filters scoped to the install scripts and frontend auth form. No GPU-dependent steps; the existing CUDA-spoof harness handles import-time CUDA assertions if any path lands there.
Plan after this turns green
Once these three jobs are green, open the same diff against unslothai/unsloth main as a single consolidated PR and close unslothai#5733, unslothai#5734, unslothai#5738 in favor of it. Then re-base the open maintainer PR backlog so the dependent Windows jobs unblock automatically.