Flux.jl
9ec52437 - Compare Reactant optimiser-state tests at full matmul precision

Commit
33 days ago
Compare Reactant optimiser-state tests at full matmul precision CUDA's default TF32-like matmuls put RAdam's small biases off by 1e-3 after `adjust!`, past the tolerance. At the highest precision the difference is below 1e-5. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Committer
Parents
Loading