Pipeline Setup Checklist

• Pick task and model ID; pin a revision for reproducibility.
• Move compute to GPU if available; consider 8-bit/4-bit for memory limits.
• Batch inputs and enable fast tokenizers.
• Configure truncation, padding, and max length to match task.
• Log model name, revision, seeds, and decoding parameters for audits.

Attention Sanity Checks

• Confirm correct masks (padding vs causal) for your use case.
• Inspect sequence lengths; avoid quadratic blow-ups when unnecessary.
• Prefer optimized attention kernels (e.g., FlashAttention where supported).
• Verify positional settings for long-context workflows.

Text Generation Defaults

• Deterministic baseline: greedy or low-temp beam (2–4 beams).
• Balanced creativity: top-p 0.9–0.95, temperature 0.7–0.9.
• Reduce loops: small repetition penalty or no-repeat n-gram 2–3.
• Use stop sequences and max tokens appropriate to the task.