Your GPU job is billing.
Is it actually training correctly?

ComputeFence is a free CLI that runs in about 30 seconds before a fine-tuning job starts and catches silent configuration failures, such as CUDA quietly falling back to CPU or checkpoints written to a disk that disappears when the pod stops.

$ pip install computefence && computefence doctor

# or, with no install:
$ uvx computefence doctor

View on GitHub →

v0.3.1 · MIT licensed · free and open source

What it catches

CUDA silently falling back to CPU, or the GPU not being visible to PyTorch
HuggingFace cache (HF_HOME) writing to ephemeral disk instead of a persistent volume
Accelerate config that doesn't match the actual GPU count
Insufficient disk headroom, and checkpoint output directories that won't persist

Built for QLoRA, LoRA, Unsloth and Axolotl fine-tuning on rented GPU pods: RunPod, Vast.ai, Lambda Labs.

What it doesn't catch

Training script correctness, learning rate issues, mid-run pod failures, export-stage out-of-memory errors, and merge integrity. It is a pre-flight check, not a monitor.

Example output

ComputeFence v0.3.1 — Pre-flight diagnostic ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 2 WARNINGS · 0 BLOCKERS · 4 PASSED Environment ✓ PyTorch 2.4.0 detected ✓ CUDA available — NVIDIA A40 (1 GPU) ✓ Accelerate config matches GPU count Storage ⚠ HF_HOME is not set — model cache will use ephemeral disk Fix: export HF_HOME=/workspace/.cache/huggingface ⚠ Checkpoint output directory is not on a persistent volume Fix: point --output-dir at an attached network volume ✓ Disk headroom: 82 GB free ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 2 warning(s) found. Review before launching. Exit codes: 0 = all passed, 1 = warnings, 2 = blockers

Price

Free and open source under the MIT licence. A design partner programme is available at $99 for three months: direct support, priority bug fixes, and work on custom checks you specify. Open an issue on GitHub or get in touch via the repository.

About