chore: Remove stale CUBLAS_WORKSPACE_CONFIG requirement

[ValerianRey]

Summary

CUBLAS_WORKSPACE_CONFIG=:4096:8 appeared in three places (CONTRIBUTING.md, tests/profiling/run_profiler.py, tests/profiling/speed_grad_vs_jac_vs_gram.py) but is no longer needed. This PR removes it.

History

• Initial squashed commit (June 2024): tests/unit/conftest.py called torch.use_deterministic_algorithms(True) unconditionally for all devices. CUBLAS_WORKSPACE_CONFIG=:4096:8 is required by cuBLAS when deterministic mode is enabled on CUDA, so the variable was necessary.

• Commit 83d2e046 ("Add command for cuda unit tests in CONTRIBUTING.md"): Added the CUBLAS_WORKSPACE_CONFIG=:4096:8 instruction to CONTRIBUTING.md for running unit tests on CUDA — correct at the time.

• PR feat: add autogram.Engine (batched) #387 / commit d1594301 ("feat: add autogram.Engine (batched)", Aug 2025): Changed conftest.py to only call torch.use_deterministic_algorithms(True) on CPU, with an explicit comment explaining why ("we also use GPU to benchmark algorithms, and we would rather have them use non-deterministic but faster algorithms"). The commit message even lists "Force using deterministic algorithms only on CPU" under the Testing section. However, CUBLAS_WORKSPACE_CONFIG was left behind in all three places — an oversight.

Since PR #387, CUBLAS_WORKSPACE_CONFIG has been unnecessary:

• Tests on CUDA: deterministic algorithms are not enabled → cuBLAS does not require the workspace config.
• Profiling on CUDA: same, and non-deterministic algorithms are preferable for benchmarking.
• tests/trajectories/optimize.py: does call use_deterministic_algorithms(True) unconditionally, but all its tensors are CPU-only (the script ends with .numpy() calls), so cuBLAS is never involved.

🤖 Generated with Claude Code