Commit Graph

541 Commits

Author SHA1 Message Date
i-robot bf03f0e030
!31252 Implementation of element wise parallel ops
Merge pull request !31252 from liuluobin/element_wise_ops
2022-03-16 01:32:29 +00:00
liuluobin f13d342986 Implementation of element wise parallel ops 2022-03-15 19:24:17 +08:00
yangzhenzhang 1f98ffb79c adafactor parallel skip handle reshape 2022-03-15 09:18:19 +08:00
i-robot dcb5cd670c
!30953 Dynamic Weight Decay
Merge pull request !30953 from wanyiming/dynamic_wd
2022-03-14 06:03:24 +00:00
lilei 690c58ebcf modify virtualdataset bug for master 2022-03-13 10:49:15 +08:00
i-robot c2a5cc1486
!31040 Produce parallel operators for Argmin/max, SquareSumAll and UnsortedSegmentProd
Merge pull request !31040 from Bert0108/reduce_operators_arg
2022-03-11 06:15:56 +00:00
i-robot 216e7c6a92
!31041 add check for conv2d
Merge pull request !31041 from yangzhenzhang/add-check-for-conv2d
2022-03-11 01:20:52 +00:00
yangzhenzhang c00d29f223 rebase 2022-03-10 19:17:06 +08:00
liuluobin 8f045d02e3 Fix a bug where ROIAlign and CropAndResize distributed op do not support GPU 2022-03-10 17:31:36 +08:00
Bert0108 bfc5e4345c add distributed operators for argmax/min sqauresumall and unsortedsetmentprod 2022-03-10 14:44:37 +08:00
wanyiming a124ec4de7 add dynamic_decay 2022-03-10 11:02:27 +08:00
i-robot 335ef1c270
!30459 Add ut validate function for parallel
Merge pull request !30459 from liuluobin/ut_master
2022-03-07 15:23:14 +00:00
liuluobin b797a410cc Add validate function for parallel ut 2022-03-07 16:19:52 +08:00
Bert0108 dfc92f1791 add distributed parallel operators for reduceall and reduceprod 2022-03-07 15:40:11 +08:00
yao_yf b60e54e0d5 support not only power of 2 2022-03-06 18:08:46 +08:00
wangjun 46612fabfb add st for shard 2022-03-01 15:32:37 +08:00
i-robot ad9757ccf0
!30661 [Auto parallel] [MoE] Fix an error of configuring MoE parallel
Merge pull request !30661 from Xiaoda/124-moe-changes
2022-03-01 02:26:37 +00:00
Xiaoda Zhang 81e5abe580 fix an error of configuring parallel 2022-02-28 17:03:53 +08:00
huangxinjing 896daee845 [AutoParallel]Fix insert error for the mirror 2022-02-26 17:54:42 +08:00
yangzhenzhang 43e6e16da3 check platform for resizebilinear 2022-02-26 09:11:32 +08:00
i-robot 0341d96dd6
!30469 add shard function to support part of the graph executed in auto_parallel under pynative mode
Merge pull request !30469 from wangjun/0223_pp
2022-02-25 06:52:24 +00:00
i-robot cfe0f76d2b
!30491 ut for allgather fusion
Merge pull request !30491 from jiahongQian/master
2022-02-25 01:30:24 +00:00
wangjun 24d448239c add pynative_parallel 2022-02-24 21:07:51 +08:00
i-robot 981eae461a
!30118 自动优化器并行特性
Merge pull request !30118 from zhuyuxiao/I4S85V
2022-02-24 11:09:44 +00:00
jiahongQian 25f57505bf ut for allgather fusion 2022-02-24 17:40:32 +08:00
i-robot bbcfbce9e0
!29997 [Auto parallel] [MoE] Support data_parallel + expert_parallel
Merge pull request !29997 from Xiaoda/124-moe-changes
2022-02-24 09:23:47 +00:00
zhuyuxiao d0e0e305d3 good 2022-02-24 16:10:18 +08:00
i-robot f2130e7434
!30483 [AutoParallel]Pipeline Automatic detection Opt
Merge pull request !30483 from lichen/pipeline_opt_detection
2022-02-24 03:40:15 +00:00
yao_yf e21f878e14 adasum ut fix 2022-02-24 09:31:04 +08:00
Xiaoda Zhang b714451937 implementing expert_parallel+data_parallel in MoE:
1) extending _Linear's input as 4-dimension tensor: [outer_batch, expert_dim, -1, hidden], and _Liner's BatchMatMul becomes BatchMatMul(4_dim_tensor, 3_dim_tensor);
2) configuring the _Linear's BatchMatMul sharding strategy as [[dp, ep, 1, 1], [ep, 1, mp]];
3) introducing a new parameter 'expert_parallel' in TransformerOpParallelConfig, creating a new class MoEParallelConfig to include 'data_parallel', 'model_parallel' and 'expert_parallel';
4) changing parallel config for FeedForward, TransformerEncoderLayer, TransformerDecoderLayer.
2022-02-24 09:24:27 +08:00
wangshengnan12@huawei.com acbefd80ea pipeline_opt_detection 2022-02-24 09:23:17 +08:00
i-robot 81260a2319
!30466 takedown test_auto_parallel_adasum.py to ensure stability, again
Merge pull request !30466 from yanghaoran/master
2022-02-23 09:15:58 +00:00
i-robot 14393503b7
!30431 allreduce allgather fusion
Merge pull request !30431 from jiahongQian/master
2022-02-23 08:52:40 +00:00
yanghaoran 71d6b7d506 takedown test_auto_parallel_adasum.py to ensure stability, again 2022-02-23 16:42:13 +08:00
i-robot 2e8eac8341
!30367 auto_parallel_adasum_support_data_parallel
Merge pull request !30367 from yao_yf/auto_parallel_adasum_support_data_parallel
2022-02-23 06:50:26 +00:00
jiahongQian 8a2151d8bb allgather reducescatter fusion 2022-02-23 10:47:30 +08:00
i-robot 5bee7156b9
!30369 add_virtualdataset_ut
Merge pull request !30369 from lilei/add_virtualdataset_ut
2022-02-23 02:21:34 +00:00
yao_yf 19236b1a70 auto parallel adasum support data parallel and hybrid parallel 2022-02-22 20:14:00 +08:00
huangxinjing 092ba035e3 Add global norm parallel support 2022-02-22 16:08:53 +08:00
yanghaoran bfe139b662 takedonw test_auto_parallel_adasum.py to ensure gate stability 2022-02-22 14:51:03 +08:00
lilei bc62e24d94 add_virtualdataset_ut 2022-02-22 11:38:10 +08:00
i-robot 94c8c6355c
!30294 auto_parallel_adasum_checks_and_ut.
Merge pull request !30294 from yao_yf/auto_parallel_adasum_checks_and_ut
2022-02-22 01:10:24 +00:00
huangxinjing 5e325ac336 [AUTO_PARALLEL]Fix insert nodes error 2022-02-21 16:39:50 +08:00
i-robot 7386612515
!29820 moe_topk routing
Merge pull request !29820 from wangshengnan123/moe_topk_routing
2022-02-21 06:21:55 +00:00
yao_yf 4b79d4c425 auto parallel adasum uts and checks 2022-02-21 10:41:09 +08:00
wangshengnan123 7322426648 top_k routing 2022-02-18 14:10:22 +08:00
i-robot 48d4f34576
!30167 Add UT case for 'Convolution+Transformer' structure
Merge pull request !30167 from Bert0108/ut_conformer
2022-02-17 07:43:58 +00:00
Bert0108 25a9c73a08 add ut case for monitoring the conformer structure 2022-02-17 15:08:01 +08:00
liuluobin b0b79ef8fb Add backend check for RandomChoiceWithMask 2022-02-16 18:56:57 +08:00
i-robot d5100190c9
!29612 Impementatin of CropAndResizeInfo and ROIAlignInfo.
Merge pull request !29612 from liuluobin/crop_resize_master
2022-02-15 11:36:06 +00:00