M1T3 — Explore numeric range and error¶
Run the seeded experiment:
To inspect a real tensor, save a tensor or state dictionary with torch.save and pass --tensor-file path.pt. The run writes results/latest.json and results/format-error.png.
The FP-like conversions are software emulations for studying range and mantissa loss. They are not throughput benchmarks and do not imply that the current GPU supports an FP8 kernel. Success means explaining at least one reversal where a nominally wider format has worse task-relevant behavior because of range, scaling, or outliers.