Young Jin Kim
9fa166be88
Online quantization ( #847 )
...
* Enable on-line packing/quantization
* Add half precision min/max quantization for model weights
* Change default quantization of B matrix to min/max, revert a false commit for AggregateAll
* Fixed missing half quantization
* Fix quantization range for A
* Set all default values for the quantize range to 0.f
* Use 7 bits clip for the weight matrix quantization to avoid an overflow of VPMADDUBSW
2021-05-24 22:51:37 -07:00
Marcin Junczys-Dowmunt
3133a9b27b
resolve conflict
2021-05-24 11:19:20 -07:00
Marcin Junczys-Dowmunt
84a20f65a1
Merge branch 'master' into pmaster
2021-05-24 11:17:53 -07:00
Marcin Junczys-Dowmunt
8b818b7c07
Avoid Ampere misaligment issue
2021-05-17 13:25:13 -07:00
Roman Grundkiewicz
fe74576dc3
Update VERSION
2021-05-04 12:36:37 +01:00
Nikolay Bogoychev
379212b75c
Enable compute86 where supported ( #863 )
...
* Enable compute86 where supported
2021-05-04 12:36:10 +01:00
Hieu Hoang
7faebf77ca
use args
2021-04-30 23:59:05 +00:00
Hieu Hoang
4500221cfc
don't define BLAS_FOUND
2021-04-30 23:50:14 +00:00
Hieu Hoang
ab2afff23a
old-style iter. For gcc 5
2021-04-30 23:10:17 +00:00
Hieu Hoang
077734331f
debug
2021-04-30 21:55:49 +00:00
Hieu Hoang
86d7e30254
getFactorMasksMultiDim
2021-04-30 21:36:05 +00:00
Hieu Hoang
82db7abf8b
start using only expr
2021-04-30 20:50:20 +00:00
Hieu Hoang
560bdbdfdc
warnings
2021-04-30 20:18:51 +00:00
Hieu Hoang
b8153bba8e
warnings
2021-04-30 12:02:15 -07:00
Hieu Hoang
1e62a16bc1
warnings
2021-04-30 10:53:18 -07:00
Hieu Hoang
84d498756b
warnings
2021-04-30 09:52:56 -07:00
Hieu Hoang
5be82498ae
virtual destructor
2021-04-30 00:13:31 -07:00
Hieu Hoang
1672201450
use origBatchIdx
2021-04-29 15:33:49 -07:00
Hieu Hoang
daf853e7aa
batch idx nearly there
2021-04-29 13:43:31 -07:00
Hieu Hoang
1a3e5ab58e
debug
2021-04-29 13:01:03 -07:00
Hieu Hoang
947301a817
lsh runs but crap output
2021-04-29 12:54:35 -07:00
Hieu Hoang
1784da0585
start lsh
2021-04-29 12:22:45 -07:00
Hieu Hoang
e518fc9666
cachedShortWt_ works
2021-04-29 11:59:25 -07:00
Hieu Hoang
d41353eeb7
get ready for cachedShortWt_
2021-04-29 11:24:00 -07:00
Hieu Hoang
5225331767
cachedShortLemmaEt_ works
2021-04-29 02:22:26 -07:00
Hieu Hoang
6b2b7d1188
factor mask
2021-04-29 00:44:30 -07:00
Hieu Hoang
f41acb1aa8
start factor mask
2021-04-29 00:31:01 -07:00
Hieu Hoang
67fe82f840
start broadcast
2021-04-29 00:08:21 -07:00
Hieu Hoang
592854f571
move cache variables into shortlist class
2021-04-28 23:56:25 -07:00
Hieu Hoang
909df372d1
restart
2021-04-28 23:40:00 -07:00
Kenneth Heafield
36b4b69d7b
Remove unused memoized_ variable ( #852 )
2021-04-28 13:28:50 +01:00
Roman Grundkiewicz
49e379bba5
Merged PR 18612: Early stopping on first, all, or any validation metrics
...
Adds `--early-stopping-on first|all|any` allowing to decide if early stopping should take into account only first, all, or any validation metrics.
Feature request: https://github.com/marian-nmt/marian-dev/issues/850
Regression tests: https://github.com/marian-nmt/marian-regression-tests/pull/79
2021-04-26 11:51:43 +00:00
Marcin Junczys-Dowmunt
1c8ee95a54
update version
2021-04-21 05:14:36 +00:00
Marcin Junczys-Dowmunt
309bd748ab
Merge branch 'master' of github.com:marian-nmt/marian-dev into pmaster
2021-04-21 05:13:58 +00:00
Marcin Junczys-Dowmunt
3e51ff3872
fix depth-scaling in FFN
2021-04-20 15:50:53 +00:00
Kenneth Heafield
bb6092da2b
Compute tensor size using integers ( #851 )
2021-04-14 08:48:51 -07:00
Marcin Junczys-Dowmunt
8a53b761d5
update version
2021-04-11 04:30:35 +00:00
Marcin Junczys-Dowmunt
ed29048004
Merge branch 'master' of vs-ssh.visualstudio.com:v3/machinetranslation/Marian/marian-dev
2021-04-11 04:29:46 +00:00
Marcin Junczys-Dowmunt
ea55722372
Merge branch 'pmaster'
2021-04-11 04:29:17 +00:00
huangjq0617
a7c3a0b2ef
fix beam_search ABORT when enable openmp and OMP_NUM_THREADS > 1 ( #767 )
2021-04-10 21:28:04 -07:00
Martin Junczys-Dowmunt
caddad90cd
Merged PR 18505: RMSNorm on GPU
...
Support for RMSNorm as drop-in replace for LayerNorm from _Biao Zhang; Rico Sennrich (2019). Root Mean Square Layer Normalization_. Enabled in Transformer model via `--transformer-postprocess dar` instead of `dan`.
2021-04-10 15:28:38 +00:00
Roman Grundkiewicz
a05124176d
Merged PR 18531: Install GCC in Azure pipelines
...
This fixes Azure pipelines after recent changes in Azure-hosted runners removing GCC 8 and older on some Ubuntu images. GCC is now installed explicitly via `apt-get`.
2021-04-09 18:44:11 +00:00
Marcin Junczys-Dowmunt
6435c6f1ce
synced with public master
2021-04-09 16:12:34 +00:00
Marcin Junczys-Dowmunt
fdf9fe7d4a
Update VERSION
2021-04-09 09:03:39 -07:00
Marcin Junczys-Dowmunt
be65065623
Allow to choose fine-grained CPU intrinsics on as CMake options ( #849 )
...
* allow to choose fine-grained CPU intrinsics on as CMake options
* inform user that e.g. -DCOMPILE_AVX2=off will be ignored with -march=native if there is compiler support
2021-04-09 09:02:34 -07:00
Marcin Junczys-Dowmunt
a17ee300f4
Create VERSION
2021-04-08 21:48:01 -07:00
rhenry-nv
fddd0e0661
Adds better Affine support for GPUs when using CUDA 11. Introduces a new bias addition kernel for CUDA < 11 ( #778 )
...
Co-authored-by: Marcin Junczys-Dowmunt <marcinjd@microsoft.com>
2021-04-08 21:46:27 -07:00
Roman Grundkiewicz
0223ce90b1
Fix Ubuntu GitHub checks ( #848 )
...
* Change ubuntu-latest to ubuntu-18.04
* Install gcc/g++
2021-04-08 18:41:15 +01:00
Marcin Junczys-Dowmunt
bfa6180033
Revert "remove TC_MALLOC from optional dependencies ( #840 )"
...
This reverts commit 096c48e51c
.
2021-04-08 07:30:38 +00:00
Rohit Jain
4408e88a94
Merged PR 18366: Fix generation of special control characters for default vocabulary
...
This PR extends the --allow-special feature to default vocabulary items as well. If the default vocabulary is provided with symbols ostensibly generated from the SentencePiece Byte Fallback mechanism, we suppress the control characters from that list.
2021-03-30 21:43:06 +00:00