Copy as Markdown View as Markdown

tinker.types.DimuonParams

class tinker.types.DimuonParams(StrictBase)

Hyperparameters for one Dimuon optimizer step.

The caller supplies schedules. Omitted settings use the defaults on each step; previous overrides do not carry forward. Matrix-sign iteration count and normalization epsilon are internal constants.

Fields:

• type (Literal['dimuon']) – Optimizer family for this step.
• learning_rate (float) – Learning rate applied to a direction with Frobenius norm sqrt(rank), with rank correction applied uniformly to all LoRA matrices.
• grad_clip_norm (float) – Maximum global gradient norm. Nonpositive values disable gradient clipping.
• beta1 (float) – Coefficient used for computing running averages of gradients.
• damping (float) – Positive damping that smooths the row and column preconditioner.
• rank_lr_correction_exponent (float) – Exponent for learning-rate correction relative to rank 32. Zero disables rank correction.