8000 Commits · sadikneipp/paxml · GitHub

More Web Proxy on the site http://driver.im/

Commits

Commits on Jul 28, 2023

Fix type signature for SeqIO input config fixtures.
The paxml Authors
committed
Pax-specific codegen and config normalization: Wrap NestedMap in a helper function, to reduce custom objects in the config.
The paxml Authors
committed
Add support for passthrough features to LanguageModelFeatures.
The paxml Authors
committed
Enabling OCDBT with simply calling `PyTreeCheckpointHandler(use_ocdbt=True)`. Users no longer need to call create_coordinator_server_and_context() & register_standard_handlers_with_options().

ChromeHearts
authored and
pax authors
committed
Add a test for write_post_init_model_hparams_file in trainer_lib.

sagunb
authored and
pax authors
committed
Pax-specific codegen: remove config parameter from experiment_fixture() for
The paxml Authors
committed
Update requirements files.

laurentes
authored and
pax authors
committed
Add FSDP style sharding for 175B model.

wang12tao
authored and
pax authors
committed
Pax-specific codegen: emit docstrings for classes and methods, to silence lint errors.
The paxml Authors
committed
Minor paxml style nits
The paxml Authors
committed

Commits on Jul 27, 2023

Options to remove eval / decoder datasets, if those are not needed for downstream usage and just add verbosity to the baseline config.
The paxml Authors
committed
Internal change.

mlbileschi
authored and
pax authors
committed
Add return type annotations for new but non-auto_config functions.
The paxml Authors
committed
Add option to rescale update by variable initial scale.
The paxml Authors
committed
Internal change.

mlbileschi
authored and
pax authors
committed

Commits on Jul 26, 2023

Update Learners and MultiLearner classes to use Optimizer protocol.

ishark
authored and
pax authors
committed
Use new import_manager_wrapper utility module (See change in https://github.com/google/fiddle)
The paxml Authors
committed
Pax-specific codegen: Use the new baseline experiment class.
The paxml Authors
committed
Pax-specific codegen: Add library for base class of generated baselines. (This is useful outside of the codegen.)
The paxml Authors
committed
Pax-specific codegen: Use fiddle validators to emit a nicer error message when
The paxml Authors
committed
Add task.decode.profiler_capture_step to enable starting the profiler at a later decoding step to avoid possibly only capturing the traced pass before the first decoding step.

TalSchuster
authored and
pax authors
committed

Commits on Jul 25, 2023

Pax-specific codegen: modularize codegen.py a bit
The paxml Authors
committed

Commits on Jul 24, 2023

Add 33B and 175B benchmark for A100.

wang12tao
authored and
pax authors
committed
Add document for PGLE workflow for paxml GPU users.

wang12tao
authored and
pax authors
committed
Fix SeqIO import by enabling SeqIO extensions.
The paxml Authors
committed
Update Pax DPO implementation and test.
The paxml Authors
committed
Enable evaluation and decoding on the final training step.
The paxml Authors
committed
Config normalization: convert dataclasses to configs, and also apply defaults removal to datasets, not just the task.
The paxml Authors
committed

Commits on Jul 21, 2023

Pax-specific codegen: Use dataclasses.field(default_factory=) for mutable high-level settings.
The paxml Authors
committed
57AE
Pax-specific codegen: Add boilerplate stanza methods to configs.
The paxml Authors
committed
Add full cache updates for n-grammer

vlad17
authored and
pax authors
committed
Generate experiment diffs from baselines.
The paxml Authors
committed
Correctly pass the `old` config to the model sharding fiddler generation, which makes the output significantly less verbose. Also uses the new short namer in case intermediate variables are needed.
The paxml Authors
committed
Add support for multiple output layers in ghostnorm generic wrapper
The paxml Authors
committed

Commits on Jul 20, 2023

Add more tests for TrainStateMetadata creation in trainer_lib.

sagunb
authored and
pax authors
committed

0