Check the prompt or conditioning file before planning a generation

[ischencheng]

Change

generate reads --prompt-file and --conditioning only after GPU detection, planning and artifacts.mkdir(exist_ok=False). A mistyped path, an empty prompt or a non-UTF-8 file fails there and leaves <out>.artifacts behind. After fixing the path, every retry with the same --out stops at

FileExistsError: [Errno 17] File exists: '/tmp/fv/out/fox.artifacts'

until the directory is deleted by hand.

This checks the input right after argument parsing. The prompt is read as UTF-8 and must be nonempty (same message as the encoder worker and the Mac path). --conditioning must be a file. Nothing is created when the check fails. exist_ok=False and the artifacts kept from real attempts are unchanged. The check runs before the macOS dispatch, so the Mac path, which creates <out>.artifacts first, gets it too.

Validation

Linux on a Modal L4 (sm89), driver 580.95.05, torch 2.13.0+cu130, Python 3.12, pip install --no-deps -e .. Compared main 5f26105 with this branch. The container has no models and did not run setup, so retries end at the encoder child. What matters here is whether a retry gets past <out>.artifacts.

Each step runs python -m freevideo_engine generate ... --cache <stub fp8 manifest> --width 768 --height 448 --seconds 2:

Step	main	this PR
--prompt-file promt.txt (typo)	FileNotFoundError after 4.3 s (detection + planning), leaves fox.artifacts/	FileNotFoundError after 0.16 s, nothing created
retry with prompt.txt, same --out	FileExistsError: .../fox.artifacts	gets past it, encoder starts
--prompt-file empty.txt	encoder child starts, leaves blank.artifacts/	ValueError: A nonempty prompt is required after 0.14 s
retry, same --out	FileExistsError: .../blank.artifacts	gets past it, encoder starts
--conditioning conditioning.pt (missing)	FileNotFoundError after 4.0 s, leaves cond.artifacts/	FileNotFoundError: Conditioning file not found after 0.14 s
retry, same --out	FileExistsError: .../cond.artifacts	gets past it, encoder starts

In this container the encoder child then stops with "Worker memory monitoring is incomplete", because the gVisor sandbox has no /proc/self/smaps_rollup. Both branches behave the same there. Reading encode_worker.py, on a normal host the empty prompt would stop at the worker's nonempty check instead, still after <out>.artifacts exists.

Also ran locally (macOS): python scripts/check_installation.py, bash -n setup.sh test.sh freevideo scripts/bootstrap_linux.sh, ./freevideo --help and generate --help. A GBK-encoded prompt now fails at once with UnicodeDecodeError.

Not run: a full generation, and the Mac app path. On macOS I only checked that the new check fires before dispatch.

Repro

mkdir -p /tmp/fv/cache /tmp/fv/out && cd /tmp/fv
echo '{"precision": "fp8"}' > cache/manifest.json
echo 'A red fox runs through fresh snow at dawn.' > prompt.txt
gen() { python -m freevideo_engine generate "$@" --cache cache --width 768 --height 448 --seconds 2; }
gen --prompt-file promt.txt --out out/fox.mp4    # typo
gen --prompt-file prompt.txt --out out/fox.mp4   # main: FileExistsError .../fox.artifacts