# Suite34 1.0.1 — CUDA workspace warning fix

## Observed problem and scope

The supplied `suite34_ipmvcl(2).log` contains 36,866 repetitions of the same
CuBLAS deterministic-workspace warning (23,776 forward and 13,090 backward).
The companion pasted excerpt shows the same cause. Neither attachment contains
a traceback or a completed final report. Refit-only validation NaNs are expected:
there is no separate validation set during the complete-history refit.

Fix startup configuration before PyTorch/CUDA is initialized, retain full epoch
and research reporting, and exercise deterministic forward/backward in the GPU
environment check. Do not hide warnings globally, change the model, retune the
experiment, or bypass run/artifact integrity checks.

## Upgrade behavior

The CUDA execution configuration and package version are recorded in the run
signature. Version 1.0.0 results cannot be resumed under 1.0.1. Keep the old
results; start this version with a separate output directory and log. A running
old Python process cannot acquire the fix from a changed file or a later shell
export. Old results remain usable as results of the original execution, but the
missing deterministic workspace cannot be repaired retroactively.

CUDA workspace values `:4096:8` (default) and `:16:8` are accepted. Explicit
invalid or empty values fail with an actionable startup message. The supported
entry points configure the workspace before importing PyTorch. For embedded use
after CUDA has already been initialized, restart the Python process first.

Reference: https://docs.pytorch.org/docs/2.7/generated/torch.use_deterministic_algorithms.html
