penzai.toolshed#
Submodules
Utility to automatically apply |
|
Basic training logic for training simple models. |
|
Utilities for gradient checkpointing / rematerialization. |
|
Utility for capturing and analyzing a small subcomputation in a larger model. |
|
Utilities for JIT compilation of Penzai models. |
|
Low-rank adaptation (LoRA, Hu et al. 2021). |
|
Helper classes for rewiring, ablating, and intervening on model activations. |
|
Patches the ipdb debugger to enable rich output (e.g. with treescope). |
|
Utilities to capture and hold intermediate values as a model runs. |
|
Utilities for working with sharded arrays and parameters in Penzai. |
|
Helpers to visualize tokens, token probabilities, and token sequences. |
|
Recursively transforms |