feat: unify local and Cloud Run deployment, merge demo improvements#25
Merged
Conversation
…emove 4bit auto-override, fix contained segments
…e/unbias-plus into unbias-pretrained
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
…e/unbias-plus into unbias-pretrained
Adds a `cache-model` job that runs Cloud Build to download the model from HuggingFace to local disk and uploads blobs + GCS FUSE symlink objects directly — bypassing GCS FUSE append-mode writes, which caused the startup probe to fail on the SFT-Instruct revision. `deploy-vllm` now gates on both `build-vllm` and `cache-model`, ensuring weights are present before the Cloud Run revision is created. The cache step is a no-op when large blobs already exist in GCS.
The path filter limited code_checks to .py / .ipynb / lockfile / config changes, so PRs touching only docs, templates, CSS, or assets bypassed pre-commit entirely. Remove the filter so every push and PR runs the full hook suite — pre-commit's per-hook scoping already handles which files each hook touches.
This commit introduces a new `training` directory containing scripts for fine-tuning the Qwen3 model for bias detection, including `train_sft.py` and `train_grpo.py`. It also adds a `sanity_check.py` script for testing model inference, along with a README for usage instructions. New tests for the training scripts are included in the `tests/training` directory, ensuring coverage for the functionality. Additionally, the `pyproject.toml` file is updated to include the new test paths.
… scripts and update mypy configuration This commit cleans up the training scripts by removing unnecessary type ignore comments for imports that are now resolved. Additionally, the `pyproject.toml` file is updated to include a new mypy path and overrides for handling heavy GPU/training-only dependencies, ensuring better type checking across environments.
This commit adds a new section to the main README outlining the training process for the Qwen3-8B model, including details on the fine-tuning stages and dependencies. Additionally, the training README is updated to include information about the dataset used, its schema, and instructions for preparing the data for training.
…e/unbias-plus into unbias-pretrained-update # Please enter a commit message to explain why this merge is necessary, # especially if it merges an updated upstream into a topic branch. # # Lines starting with '#' will be ignored, and an empty message aborts # the commit.
…on to 3.8.1 for security advisory compliance, and update transformers version constraint to <6.0.0. Adjust unsloth tag to March-2026 in uv.lock.
aravind-3105
force-pushed
the
unbias-pretrained
branch
from
97f08e3 to
8cf030a
Compare
…e/unbias-plus into unbias-pretrained-update # Please enter a commit message to explain why this merge is necessary, # especially if it merges an updated upstream into a topic branch. # # Lines starting with '#' will be ignored, and an empty message aborts # the commit.
Unbias pretrained updated site
Replace blob-count heuristic with a git ls-remote / refs/main SHA comparison so stale weights are detected and cleared automatically. Also fixes a tr '/' '--' bug that produced single-dash paths instead of the double-dash format used by the HuggingFace cache layout.
- Bump setup-uv to v8.1.0 across all workflows (from main) - Keep CVE/PYSEC ignore list from unbias-pretrained (with justification comments) - Regenerate uv.lock from merged pyproject.toml
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Unifies local and Cloud Run deployment into a single branch. Replaces the separate
deploybranch.Type of Change
Changes Made
api.py— unified vLLM + local transformer modes viaVLLM_BASE_URLenv varscript.js— added file upload, copy button, cold-start retry logic, 5000 char limitDockerfile,nginx.conf,entrypoint.sh,deploy-demo.yml)0.10.12across all workflowspyproject.tomlto omittrainextra from default lockfile resolutionTesting
Deployment Notes
Set
VLLM_BASE_URLenv var to switch to Cloud Run vLLM mode. Without it, runs local transformer as before.