Commits

Commits on Oct 24, 2024

Retire DeepSpeed mode due to security issue with supported version (#1116)
Michael Denkowski
authored
871d986

Commits on Jun 7, 2024

Add maintenance mode notice to README. (#1114)
Michael Denkowski
authored
e42fbb3

Commits on May 22, 2024

Use `weights_only=True` when loading SockeyeModel parameters. (#1112)
Michael Denkowski
authored
b0a2cfc

Commits on Oct 17, 2023

Update url for nightly builds. (#1100)
fhieber
authored
98b4946

Commits on Oct 16, 2023

Update README.md w/ recent publications using Sockeye (#1099)
fhieber
authored
24cb8be

Commits on Sep 7, 2023

beam_search fix for running with torch.use_deterministic_algorithms(True) (#1096)
Jehovan
authored
2d80b2a

Commits on Aug 25, 2023

Refactor inference optimizations (#1094)
xingniu
authored
a59e7a2

Commits on Aug 10, 2023

Use fault tolerant symlink during training to handle temporary file system inconsistencies. (#1093)
Michael Denkowski
authored
8753d95

Commits on Mar 2, 2023

Fixes to blocking cross-attention (#1087)
xingniu
authored
4c30942

Commits on Mar 1, 2023

Two fixes to SampleK (#1086)
tdomhan
authored
01e2392

Commits on Feb 23, 2023

Add blocking cross-attention between decoder and encoded prepended tokens (#1085)
xingniu
authored
d912554

Commits on Jan 3, 2023

BFloat16 Integration/System Test Update (#1079)
Michael Denkowski
authored
13c63be

Commits on Dec 20, 2022

Enable bfloat16 inference and fix unrelated NumPy issue (#1077)
Michael Denkowski
authored
a08216c

Commits on Dec 11, 2022

Code cleanup: refactoring, type checking, and formatting (#1076)
Michael Denkowski
authored
4dba5a3

Commits on Dec 10, 2022

Add kNN-MT (#1062)

authored
b4c8427

Commits on Nov 6, 2022

Update to PyTorch 1.13 (+ 1.14 compatibility) (#1074)
fhieber
authored
288baa7

Commits on Nov 5, 2022

Add --tf32 device arg for transparent float32 acceleration (#1066)
graehl
authored
f852fbd
Update to sacrebleu 2.3.1 (#1075)
fhieber
authored
c09a168

Commits on Oct 1, 2022

Update DeepSpeed checkpoint conversion to support newer DeepSpeed versions (#1071)
Michael Denkowski
authored
be2b344

Commits on Sep 18, 2022

Change decoder softmax size logging level from info to debug. (#1068)
Michael Denkowski
authored
a3088b4

Commits on Sep 6, 2022

translator: send skip_nvs and nvs_thresh from args to ctor (#1064)
graehl
authored
7caa6b9

Commits on Aug 25, 2022

Add Support for DeepSpeed ZeRO Stage 1 (#1059)
Michael Denkowski
authored
9690498

Commits on Aug 23, 2022

Update README.md with some recent publications (#1058)
fhieber
authored
d80169d

Commits on Aug 18, 2022

[Review #1051 First] Refactoring Ahead of Adding DeepSpeed Support (#1057)
Michael Denkowski
authored
04117a0

Commits on Aug 9, 2022

Option to clamp layer outputs to min/max finite values for their dtypes (#1051)
Michael Denkowski
authored
236c81e

Commits on Aug 6, 2022

Update website with latest arxiv paper (#1056)
fhieber
authored
11a0b9d

Commits on Jul 15, 2022

Add Sockeye 3 paper citation (#1052)
Michael Denkowski
authored
a75864d

Commits on Jul 6, 2022

Support offline model quantization (#1050)
Michael Denkowski
authored
fbbf360

Commits on Jun 14, 2022

Nbest reranking options using isometric translation criteria. (#1048)
surafelml
authored
6962770

Commits on May 29, 2022

fix data type of valid_length (#1047)
hoangcuong2011
authored
7c71ff8

Commits on May 4, 2022

Neural vocabulary selection. (#1046)

tdomhan
and
Tobias Domhan
authored
94cdad7

Commits on Apr 28, 2022

Option to not reload the best training checkpoint when reducing the learning rate (#1045)
Michael Denkowski
authored
63286ff

Commits on Apr 21, 2022

Fix scoring with batches of size 1 by avoid squeezing the batch dim (#1044)
fhieber
authored
23ffd29
Checkpoint is up to date when resuming training (#1043)
Michael Denkowski
authored
c822e20

Commits on Apr 12, 2022

Ignore false-positive missing keys for traced modules (#1042)
Michael Denkowski
authored
30c3913