Files
polyscribe/docs/development.md

90 lines
3.5 KiB
Markdown

# Development
This document describes how to build, test, and run PolyScribe locally, and how to work with models during development.
Prerequisites
- Rust toolchain via rustup (recommended).
- ffmpeg installed and on PATH.
- For GPU builds: appropriate toolkits and libraries installed (CUDA/ROCm/Vulkan).
Rust toolchain
- Target stable Rust.
- Recommended:
- rustup install stable
- rustup default stable
Dependency pinning
- We pin whisper-rs (git dependency) to a known-good commit in Cargo.toml for reproducibility.
- To bump it, resolve/test the desired commit locally, then run:
- cargo update -p whisper-rs --precise 135b60b85a15714862806b6ea9f76abec38156f1
Replace the SHA with the desired commit and update the rev in Cargo.toml accordingly.
Build
- CPU-only (default):
- cargo build
- cargo test
- Enable GPU features at build time:
- CUDA: cargo build --features gpu-cuda
- HIP: cargo build --features gpu-hip
- Vulkan: cargo build --features gpu-vulkan
Run locally
- Development model directory defaults to ./models when built in debug mode.
- You can override:
- POLYSCRIBE_MODELS_DIR=/path/to/models
- WHISPER_MODEL=/path/to/model.bin (forces a specific file)
- Example run:
- cargo run -- -v -o output samples/example.mp3
Models during development
- Interactive downloader:
- cargo run -- --download-models
- Non-interactive update (checks sizes/hashes, downloads if missing):
- cargo run -- --update-models --no-interaction -q
Tests
- Run all tests:
- cargo test
- The test suite includes CLI-oriented integration tests and unit tests. Some tests simulate GPU detection using env vars (POLYSCRIBE_TEST_FORCE_*). Do not rely on these flags in production code.
Examples check (no network, non-interactive)
- To quickly validate that example scripts are wired correctly (no prompts, quiet, exit 0), run:
- make examples-check
- What it does:
- Iterates over examples/*.sh
- Forces execution with --no-interaction and -q via a wrapper
- Uses a stubbed BIN that performs no network access and exits successfully
- Redirects stdin from /dev/null to ensure no prompts
- This is intended for CI smoke checks and local verification; it does not actually download models or transcribe audio.
Clippy
- Run lint checks and treat warnings as errors:
- cargo clippy --all-targets -- -D warnings
- Common warnings can often be fixed by simplifying code, removing unused imports, and following idiomatic patterns.
Code layout
- src/lib.rs: core library surface and re-exports
- backend: runtime backend selection and transcription glue
- models: model discovery, manifest, download/update logic
- src/main.rs: CLI parsing, I/O orchestration, logging, and workflows
- tests/: integration tests
Adding a feature
- Find the closest existing module (backend/models/main) and add a small, focused unit test.
- Keep user-facing changes documented in README/docs and update CHANGELOG.md.
- Prefer small functions with clear responsibilities; avoid exposing unnecessary items publicly.
- Follow existing logging style (elog!/wlog!/ilog!/dlog!).
Running the model downloader
- Interactive:
- cargo run -- --download-models
- Non-interactive suggestions for CI:
- POLYSCRIBE_MODELS_DIR=$PWD/models \
cargo run -- --update-models --no-interaction -q
Env var examples for local testing
- Use a local copy of models and a specific model file:
- export POLYSCRIBE_MODELS_DIR=$PWD/models
- export WHISPER_MODEL=$PWD/models/large-v3-turbo-q8_0.bin
- Test manifests/offline copies are handled internally. For full offline runs, pre-populate models/ with the desired .bin files.