Commits

Branch selector

 master

User selector

All users

All time

Commit History

Commits on Dec 13, 2021

• update gitignore

  StevenShi-23

  committed
  1a148a1

  Copy full SHA for 1a148a1

Commits on Dec 11, 2021

• Make XLA dump only from one host.

  Show description for 3d22fec
  

  zhangqiaorjc

  authored and

  copybara-github

  committed
  3d22fec

  Copy full SHA for 3d22fec

• Expand normalization policy options and make attention head size configurable.

  Show description for 688fa66
  

  lingvo-bot

  authored and

  copybara-github

  committed
  688fa66

  Copy full SHA for 688fa66

• TPU_REPLICATED_CORE is not meaningful in eager placement.

  Show description for a7ae57a
  

  lingvo-bot

  authored and

  copybara-github

  committed
  a7ae57a

  Copy full SHA for a7ae57a

• Read the TriviaQA text as bytes representations by calling ast.literal_eval on it first.

  Show description for 5bf5601
  

  lingvo-bot

  authored and

  copybara-github

  committed
  5bf5601

  Copy full SHA for 5bf5601

• Use model.GetVariablesDict() to obtain all variables for checkpoint saving in

  Show description for bcf61e3
  

  Jonathan Shen

  authored and

  copybara-github

  committed
  bcf61e3

  Copy full SHA for bcf61e3

• Skip UpdateProgramSchedule for MultiTaskProgramSchedule as it is not supported

  Show description for de682cc
  

  Jonathan Shen

  authored and

  copybara-github

  committed
  de682cc

  Copy full SHA for de682cc

Commits on Dec 10, 2021

• Fix spmd decoding: manually fix global batch size by a factor of process count.

  Show description for a2f1e9a
  

  lingvo-bot

  authored and

  copybara-github

  committed
  a2f1e9a

  Copy full SHA for a2f1e9a

• moe weight_split_dims_mapping and activation_split_dims_mapping

  Show description for 0c2cd81
  

  lingvo-bot

  authored and

  copybara-github

  committed
  0c2cd81

  Copy full SHA for 0c2cd81

• - Support checkpoint loading using init_from_checkpoint_rules for eager trainer.

  Show description for e71ee02
  

  lingvo-bot

  authored and

  copybara-github

  committed
  e71ee02

  Copy full SHA for e71ee02

• Add ConcatLayer.

  Show description for 20ec550
  

  Sergey Kishchenko

  authored and

  copybara-github

  committed
  20ec550

  Copy full SHA for 20ec550

• Test for eager vs graph checkpoint compatibility with multiple learners.

  Show description for 0bb6f5e
  

  Jonathan Shen

  authored and

  copybara-github

  committed
  0bb6f5e

  Copy full SHA for 0bb6f5e

• Make shape inference more robust to incomplete shape information.

  Show description for 94b3e79
  

  lingvo-bot

  authored and

  copybara-github

  committed
  94b3e79

  Copy full SHA for 94b3e79

• BertSpmdL66H12kBiggerBatch8x8x8 model.

  Show description for ab1e611
  

  lingvo-bot

  authored and

  copybara-github

  committed
  ab1e611

  Copy full SHA for ab1e611

• RoundRobinScheduler should not sort tasks.

  Show description for adb5016
  

  Jonathan Shen

  authored and

  copybara-github

  committed
  adb5016

  Copy full SHA for adb5016

• Fix eager evaler/decoders.

  Show description for 971e0bc
  

  Jonathan Shen

  authored and

  copybara-github

  committed
  971e0bc

  Copy full SHA for 971e0bc

• Fix the prefix length: it's randomly between the min set in the param and the actual seq length.

  Show description for 4fb8acd
  

  lingvo-bot

  authored and

  copybara-github

  committed
  4fb8acd

  Copy full SHA for 4fb8acd

• Add an input to read plain text data.

  Show description for 1737e0d
  

  lingvo-bot

  authored and

  copybara-github

  committed
  1737e0d

  Copy full SHA for 1737e0d

• Minor fixes to fix and verify that SPMD decoding works with multiple Jax processes.

  Show description for 5d4d881
  

  lingvo-bot

  authored and

  copybara-github

  committed
  5d4d881

  Copy full SHA for 5d4d881

• Remove unused model_p.decoder_inputs_split_mapping and model_p.decoder_states_split_mapping.

  Show description for 7891c0a
  

  lingvo-bot

  authored and

  copybara-github

  committed
  7891c0a

  Copy full SHA for 7891c0a

• - Add a TransformerEncoderDecoder class inheriting from TransformerLm

  Show description for 483d896
  

  lingvo-bot

  authored and

  copybara-github

  committed
  483d896

  Copy full SHA for 483d896

• Avoid clearing the variable scope for optimizer apply in eager mode.

  Show description for 5094737
  

  Jonathan Shen

  authored and

  copybara-github

  committed
  5094737

  Copy full SHA for 5094737

Commits on Dec 9, 2021

• Default to wrapping optimizer.Apply in learner variable_scope.

  Show description for 0468a8d
  

  Jonathan Shen

  authored and

  copybara-github

  committed
  0468a8d

  Copy full SHA for 0468a8d

• Add support for decode_once with spmd model.

  Show description for 9ddf63a
  

  lingvo-bot

  authored and

  copybara-github

  committed
  9ddf63a

  Copy full SHA for 9ddf63a

• Bug fix

  Show description for fb11eb0
  

  lingvo-bot

  authored and

  copybara-github

  committed
  fb11eb0

  Copy full SHA for fb11eb0

• Introduce library to factorize out common model params configurations.

  Show description for 0f9c642
  

  laurentes

  authored and

  copybara-github

  committed
  0f9c642

  Copy full SHA for 0f9c642

• Upstream mesh utils to JAX core.

  Show description for 05bcb8a
  

  3 peopleauthored and

  copybara-github

  committed
  05bcb8a

  Copy full SHA for 05bcb8a

• Adds handling for zero area bounding boxes.

  Show description for 21be815
  

  lingvo-bot

  authored and

  copybara-github

  committed
  21be815

  Copy full SHA for 21be815

Commits on Dec 8, 2021

• Expose initial accumulator value for Adagrad.

  Show description for 697f305
  

  lingvo-bot

  authored and

  copybara-github

  committed
  697f305

  Copy full SHA for 697f305

• Internal change

  Show description for 6ea7773
  

  lingvo-bot

  authored and

  copybara-github

  committed
  6ea7773

  Copy full SHA for 6ea7773

• Change the default per sequence loss to be the mean across the sequence rather than the sum. This makes the range more consistent with what is typically defined as "loss" in lingvo.

  Show description for c08852d
  

  lingvo-bot

  authored and

  copybara-github

  committed
  c08852d

  Copy full SHA for c08852d

• Reset `_num_batches_produced` counter for the reset() method so that it properly resets for TFDS.

  Show description for 9c4c501
  

  lingvo-bot

  authored and

  copybara-github

  committed
  9c4c501

  Copy full SHA for 9c4c501

• Add checkpoint_to_load param for executor Programs, so that user can run oneoff eval/decode on a checkpoint, also allows flag override for oneoff runs.

  Show description for da0587d
  

  lingvo-bot

  authored and

  copybara-github

  committed
  da0587d

  Copy full SHA for da0587d

• Make sure pmap decode one step works with multi-processes by adding all_gather to get decoded outputs from all replicas.

  Show description for 2db2059
  

  lingvo-bot

  authored and

  copybara-github

  committed
  2db2059

  Copy full SHA for 2db2059

• Add conformer layer implementation

  Show description for 8b6a834
  

  lingvo-bot

  authored and

  copybara-github

  committed
  8b6a834

  Copy full SHA for 8b6a834

Pagination

Previous

Next