Commit Graph

269 Commits (af952673f758c71126b27de8b32bdf5df8f74b69)

Author SHA1 Message Date
ver217 8f7ce94b8e
[hotfix] fix auto tensor placement policy (#753) 2022-04-14 12:04:45 +08:00
HELSON 84c6700b2a
[zero] refactor memstats_collector (#746) 2022-04-14 12:01:12 +08:00
Jiarui Fang 3d7dc46d33
[zero] use factory pattern for tensor_placement_policy (#752) 2022-04-14 11:07:29 +08:00
ver217 4b048a8728
fix prepare grads in sharded optim (#749) 2022-04-13 22:36:11 +08:00
ver217 e396bb71f2
[zero] add tensor placement policies (#743)
* add tensor placement policies

* polish comments

* polish comments

* update moe unit tests
2022-04-13 15:00:48 +08:00
HELSON 22c4b88d56
[zero] refactor ShardedParamV2 for convenience (#742) 2022-04-13 14:54:26 +08:00
ver217 e6212f56cd
[hotfix] fix memory leak in backward of sharded model (#741) 2022-04-13 09:59:05 +08:00
Jiarui Fang 7db3ccc79b
[hotfix] remove duplicated param register to stateful tensor manager (#728) 2022-04-12 13:55:25 +08:00
Jiarui Fang 4d90a7b513
[refactor] zero directory (#724) 2022-04-11 23:13:02 +08:00
Jiarui Fang 193dc8dacb
[refactor] refactor the memory utils (#715) 2022-04-11 16:47:57 +08:00
HELSON dbd96fe90a
[zero] check whether gradients have inf and nan in gpu (#712) 2022-04-11 15:40:13 +08:00
ver217 715b86eadd
[hotfix] fix stm cuda model data size (#710) 2022-04-11 15:10:39 +08:00
HELSON a9b8300d54
[zero] improve adaptability for not-shard parameters (#708)
* adapt post grad hooks for not-shard parameters
* adapt optimizer for not-shard parameters
* offload gradients for not-replicated parameters
2022-04-11 13:38:51 +08:00
ver217 ab8c6b4a0e
[zero] refactor memstats collector (#706)
* refactor memstats collector

* fix disposable

* polish code
2022-04-11 10:46:08 +08:00
HELSON ee112fe1da
[zero] adapt zero hooks for unsharded module (#699) 2022-04-08 20:23:26 +08:00
ver217 3c9cd5bb5e
[zero] stateful tensor manager (#687)
* [WIP] stateful tensor manager

* add eviction strategy

* polish code

* polish code

* polish comment

* add unit test

* fix sampler bug

* polish code

* fix max sampling cnt resetting bug

* fix sampler bug

* polish code

* fix bug

* fix unit test

Co-authored-by: jiaruifang <fangjiarui123@gmail.com>
2022-04-08 17:51:34 +08:00
HELSON d7ecaf362b
[zero] fix init bugs in zero context (#686)
* adapt model weight initialization for methods in Pytorch nn.init
2022-04-07 17:38:45 +08:00
Jiarui Fang 59bf2dc590
[zero] initialize a stateful tensor manager (#614) 2022-04-06 16:18:49 +08:00
HELSON 17e73e62cc
[hotfix] fix bugs for unsharded parameters when restore data (#664) 2022-04-03 22:02:11 +08:00
Jiarui Fang 0aab52301e
[hotfix] fix a bug in model data stats tracing (#655) 2022-04-03 21:48:06 +08:00
Jiarui Fang 036404ca8a
Revert "[zero] polish init context (#645)" (#657) 2022-04-02 18:30:06 +08:00
Jiarui Fang 67b4928244
[zero] polish init context (#645) 2022-04-02 15:52:04 +08:00
HELSON 055fbf5be6
[zero] adapt zero for unsharded paramters (Optimizer part) (#601) 2022-04-01 20:10:47 +08:00
ver217 0ef8819c67
polish docstring of zero (#612) 2022-04-01 14:50:56 +08:00
ver217 9bee119104
[hotfix] fix sharded optim zero grad (#604)
* fix sharded optim zero grad

* polish comments
2022-04-01 12:41:20 +08:00
Jiarui Fang e956d93ac2
[refactor] memory utils (#577) 2022-04-01 09:22:33 +08:00
HELSON e6d50ec107
[zero] adapt zero for unsharded parameters (#561)
* support existing sharded and unsharded parameters in zero

* add unitest for moe-zero model init

* polish moe gradient handler
2022-03-31 18:34:11 +08:00
ver217 7c6c427db1
[zero] trace states of fp16/32 grad and fp32 param (#571) 2022-03-31 16:26:54 +08:00
Jiarui Fang 7675366fce
[polish] rename col_attr -> colo_attr (#558) 2022-03-31 12:25:45 +08:00
ver217 014bac0c49
[zero] hijack p.grad in sharded model (#554)
* hijack p.grad in sharded model

* polish comments

* polish comments
2022-03-30 18:14:50 +08:00
Jiarui Fang f552b11294
[zero] label state for param fp16 and grad (#551) 2022-03-30 15:57:46 +08:00
Jiarui Fang 214da761d4
[zero] add stateful tensor (#549) 2022-03-30 13:51:37 +08:00
Jiarui Fang 107b99ddb1
[zero] dump memory stats for sharded model (#548) 2022-03-30 09:38:44 +08:00
HELSON 8c90d4df54
[zero] add zero context manager to change config during initialization (#546) 2022-03-29 17:57:59 +08:00
Jiarui Fang 53b1b6e340
[zero] non model data tracing (#545) 2022-03-29 15:45:48 +08:00
ver217 fb841dd5c5
[zero] optimize grad offload (#539)
* optimize grad offload

* polish code

* polish code
2022-03-29 12:48:00 +08:00
ver217 1f90a3b129
[zero] polish ZeroInitContext (#540) 2022-03-29 09:09:04 +08:00
Jiarui Fang c11ff81b15
[zero] get memory usage of sharded optim v2. (#542) 2022-03-29 09:08:18 +08:00
HELSON a30e2b4c24
[zero] adapt for no-leaf module in zero (#535)
only process module's own parameters in Zero context

add zero hooks for all modules that contrain parameters

gather parameters only belonging to module itself
2022-03-28 17:42:18 +08:00
Jiarui Fang 705f56107c
[zero] refactor model data tracing (#537) 2022-03-28 16:38:18 +08:00
Jiarui Fang a590ed0ba3
[zero] improve the accuracy of get_memory_usage of sharded param (#538) 2022-03-28 16:19:19 +08:00
Jiarui Fang 37cb70feec
[zero] get memory usage for sharded param (#536) 2022-03-28 15:01:21 +08:00
Jiarui Fang 05e33b2578
[zero] fix grad offload (#528)
* [zero] fix grad offload

* polish code
2022-03-25 18:23:25 +08:00
Jiarui Fang 8d8c5407c0
[zero] refactor model data tracing (#522) 2022-03-25 18:03:32 +08:00
Jiarui Fang 4d322b79da
[refactor] remove old zero code (#517) 2022-03-25 14:54:39 +08:00
Jiarui Fang 920c5889a7
[zero] add colo move inline (#521) 2022-03-25 14:02:55 +08:00
Jiarui Fang 0bebda6ea5
[zero] fix init device bug in zero init context unittest (#516) 2022-03-25 12:24:18 +08:00
Jiarui Fang 7ef3507ace
[zero] show model data cuda memory usage after zero context init. (#515) 2022-03-25 11:23:35 +08:00
ver217 a2e61d61d4
[zero] zero init ctx enable rm_torch_payload_on_the_fly (#512)
* enable rm_torch_payload_on_the_fly

* polish docstr
2022-03-24 23:44:00 +08:00
Jiarui Fang bca0c49a9d
[zero] use colo model data api in optimv2 (#511) 2022-03-24 17:19:34 +08:00
Jiarui Fang 0035b7be07
[memory] add model data tensor moving api (#503) 2022-03-24 14:29:41 +08:00
ver217 9ec1ce6ab1
[zero] sharded model support the reuse of fp16 shard (#495)
* sharded model supports reuse fp16 shard

* rename variable

* polish code

* polish code

* polish code
2022-03-23 14:59:59 +08:00
ver217 c4c02424f3
[zero] sharded model manages ophooks individually (#492) 2022-03-22 17:33:20 +08:00
ver217 a9ecb4b244
[zero] polish sharded optimizer v2 (#490) 2022-03-22 15:53:48 +08:00
ver217 62b0a8d644
[zero] sharded optim support hybrid cpu adam (#486)
* sharded optim support hybrid cpu adam

* update unit test

* polish docstring
2022-03-22 14:56:59 +08:00
Jiarui Fang b334822163
[zero] polish sharded param name (#484)
* [zero] polish sharded param name

* polish code

* polish

* polish code

* polish

* polsih

* polish
2022-03-22 14:36:16 +08:00
ver217 8d3250d74b
[zero] ZeRO supports pipeline parallel (#477) 2022-03-21 16:55:37 +08:00
ver217 3cb3fc275e
zero init ctx receives a dp process group (#471) 2022-03-21 11:18:55 +08:00
ver217 fc8e6db005
[doc] Update docstring for ZeRO (#459)
* polish sharded model docstr

* polish sharded optim docstr

* polish zero docstr

* polish shard strategy docstr
2022-03-18 16:48:20 +08:00
ver217 a241f61b34
[zero] Update initialize for ZeRO (#458)
* polish code

* shard strategy receive pg in shard() / gather()

* update zero engine

* polish code
2022-03-18 16:18:31 +08:00
ver217 642846d6f9
update sharded optim and fix zero init ctx (#457) 2022-03-18 15:44:47 +08:00
Jiarui Fang e2e9f82588
Revert "[zero] update sharded optim and fix zero init ctx" (#456)
* Revert "polish code"

This reverts commit 8cf7ff08cf.

* Revert "rename variables"

This reverts commit e99af94ab8.

* Revert "remove surplus imports"

This reverts commit 46add4a5c5.

* Revert "update sharded optim and fix zero init ctx"

This reverts commit 57567ee768.
2022-03-18 15:22:43 +08:00
ver217 e99af94ab8 rename variables 2022-03-18 14:25:25 +08:00
ver217 57567ee768 update sharded optim and fix zero init ctx 2022-03-18 14:25:25 +08:00
Jiarui Fang 0fcfb1e00d
[test] make zero engine test really work (#447) 2022-03-17 17:24:25 +08:00
Jiarui Fang 237d08e7ee
[zero] hybrid cpu adam (#445) 2022-03-17 15:05:41 +08:00
Jiarui Fang 496cbb0760
[hotfix] fix initialize bug with zero (#442) 2022-03-17 13:16:22 +08:00
Jiarui Fang 640a6cd304
[refactory] refactory the initialize method for new zero design (#431) 2022-03-16 19:29:37 +08:00
ver217 fce9432f08 sync before creating empty grad 2022-03-16 14:24:09 +08:00
ver217 ea6905a898 free param.grad 2022-03-16 14:24:09 +08:00
ver217 9506a8beb2 use double buffer to handle grad 2022-03-16 14:24:09 +08:00
Jiarui Fang adebb3e041
[zero] cuda margin space for OS (#418) 2022-03-15 12:02:19 +08:00
Jiarui Fang 56bb412e72
[polish] use GLOBAL_MODEL_DATA_TRACER (#417) 2022-03-15 11:29:46 +08:00
Jiarui Fang 23ba3fc450
[zero] refactory ShardedOptimV2 init method (#416) 2022-03-15 10:45:55 +08:00
Frank Lee e79ea44247
[fp16] refactored fp16 optimizer (#392) 2022-03-15 10:05:38 +08:00
Jiarui Fang 21dc54e019
[zero] memtracer to record cuda memory usage of model data and overall system (#395) 2022-03-14 22:05:30 +08:00
Jiarui Fang 370f567e7d
[zero] new interface for ShardedOptimv2 (#406) 2022-03-14 20:48:41 +08:00
ver217 63469c0f91 polish code 2022-03-14 15:48:55 +08:00
ver217 88804aee49 add bucket tensor shard strategy 2022-03-14 14:48:32 +08:00
HELSON 7c079d9c33
[hotfix] fixed bugs in ShardStrategy and PcieProfiler (#394) 2022-03-11 18:12:46 +08:00
Jiarui Fang 3af13a2c3e [zero] polish ShardedOptimV2 unittest (#385)
* place params on cpu after zero init context

* polish code

* bucketzed cpu gpu tensor transter

* find a bug in sharded optim unittest

* add offload unittest for ShardedOptimV2.

* polish code and make it more robust
2022-03-11 15:50:28 +08:00
Jiarui Fang 272ebfb57d [bug] shard param during initializing the ShardedModelV2 (#381) 2022-03-11 15:50:28 +08:00
Jiarui Fang b5f43acee3 [zero] find miss code (#378) 2022-03-11 15:50:28 +08:00
Jiarui Fang 6b6002962a [zero] zero init context collect numel of model (#375) 2022-03-11 15:50:28 +08:00
jiaruifang d9217e1960 Revert "[zero] bucketized tensor cpu gpu copy (#368)"
This reverts commit bef05489b6.
2022-03-11 15:50:28 +08:00
Jiarui Fang 00670c870e [zero] bucketized tensor cpu gpu copy (#368) 2022-03-11 15:50:28 +08:00
Jiarui Fang 44e4891f57 [zero] able to place params on cpu after zero init context (#365)
* place params on cpu after zero init context

* polish code
2022-03-11 15:50:28 +08:00
ver217 253e54d98a fix grad shape 2022-03-11 15:50:28 +08:00
Jiarui Fang ea2872073f [zero] global model data memory tracer (#360) 2022-03-11 15:50:28 +08:00
Jiarui Fang cb34cd384d [test] polish zero related unitest (#351) 2022-03-11 15:50:28 +08:00
ver217 d0ae0f2215 [zero] update sharded optim v2 (#334) 2022-03-11 15:50:28 +08:00
jiaruifang 5663616921 polish code 2022-03-11 15:50:28 +08:00
jiaruifang 7977422aeb add bert for unitest and sharded model is not able to pass the bert case 2022-03-11 15:50:28 +08:00
ver217 1388671699 [zero] Update sharded model v2 using sharded param v2 (#323) 2022-03-11 15:50:28 +08:00
Jiarui Fang 11bddb6e55 [zero] update zero context init with the updated test utils (#327) 2022-03-11 15:50:28 +08:00
Jiarui Fang de0468c7a8 [zero] zero init context (#321)
* add zero init context

* add more flags for zero init context
fix bug of repeated converting param to ShardedParamV2

* polish code
2022-03-11 15:50:28 +08:00
LuGY a3269de5c9 [zero] cpu adam kernel (#288)
* Added CPU Adam

* finished the cpu adam

* updated the license

* delete useless parameters, removed resnet

* modified the method off cpu adam unittest

* deleted some useless codes

* removed useless codes

Co-authored-by: ver217 <lhx0217@gmail.com>
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: jiaruifang <fangjiarui123@gmail.com>
2022-03-11 15:50:28 +08:00
Jiarui Fang 90d3aef62c [zero] yet an improved sharded param (#311) 2022-03-11 15:50:28 +08:00
Jiarui Fang c9e7d9582d [zero] polish shard strategy (#310)
* init shard param from shape tuple

* add more unitest for shard param

* add set_payload method for ShardedParam

* [zero] add shareded tensor class

* polish code

* add shard stratgy

* move shard and gather logic to shard strategy from shard tensor.

* polish code
2022-03-11 15:50:28 +08:00
ver217 3092317b80 polish code 2022-03-11 15:50:28 +08:00
ver217 36f9a74ab2 fix sharded param hook and unit test 2022-03-11 15:50:28 +08:00
ver217 001ca624dd impl shard optim v2 and add unit test 2022-03-11 15:50:28 +08:00
Jiarui Fang 74f77e314b [zero] a shard strategy in granularity of tensor (#307) 2022-03-11 15:50:28 +08:00
Jiarui Fang 80364c7686 [zero] sharded tensor (#305)
* init shard param from shape tuple

* add more unitest for shard param

* add set_payload method for ShardedParam

* [zero] add shareded tensor class

* polish code
2022-03-11 15:50:28 +08:00
ver217 b105371ace rename shared adam to sharded optim v2 2022-03-11 15:50:28 +08:00
ver217 70814dc22f fix master params dtype 2022-03-11 15:50:28 +08:00
ver217 795210dd99 add fp32 master params in sharded adam 2022-03-11 15:50:28 +08:00
ver217 a109225bc2 add sharded adam 2022-03-11 15:50:28 +08:00
Jiarui Fang e17e92c54d Polish sharded parameter (#297)
* init shard param from shape tuple

* add more unitest for shard param

* add more unittests to shareded param
2022-03-11 15:50:28 +08:00
ver217 7aef75ca42 [zero] add sharded grad and refactor grad hooks for ShardedModel (#287) 2022-03-11 15:50:28 +08:00
Frank Lee 9afb5c8b2d fixed typo in ShardParam (#294) 2022-03-11 15:50:28 +08:00
Frank Lee e17e54e32a added buffer sync to naive amp model wrapper (#291) 2022-03-11 15:50:28 +08:00
Jiarui Fang 5a560a060a Feature/zero (#279)
* add zero1 (#209)

* add zero1

* add test zero1

* update zero stage 1 develop (#212)

* Implement naive zero3 (#240)

* naive zero3 works well

* add zero3 param manager

* add TODOs in comments

* add gather full param ctx

* fix sub module streams

* add offload

* fix bugs of hook and add unit tests

* fix bugs of hook and add unit tests (#252)

* add gather full param ctx

* fix sub module streams

* add offload

* fix bugs of hook and add unit tests

* polish code and add state dict hook

* fix bug

* update unit test

* refactor reconstructed zero code

* clip_grad support zero3 and add unit test

* add unit test for Zero3ParameterManager

* [WIP] initialize the shard param class

* [WIP] Yet another sharded model implementation (#274)

* [WIP] initialize the shard param class

* [WIP] Yes another implementation of shardModel. Using a better hook method.

* torch.concat -> torch.cat

* fix test_zero_level_1.py::test_zero_level_1 unitest

* remove deepspeed implementation and refactor for the reconstructed zero module

* polish zero dp unittests

Co-authored-by: ver217 <lhx0217@gmail.com>
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
2022-03-11 15:50:28 +08:00
HELSON 0f8c7f9804
Fixed docstring in colossalai (#171) 2022-01-21 10:44:30 +08:00
ver217 9ef05ed1fc
try import deepspeed when using zero (#130) 2022-01-07 17:24:57 +08:00
Frank Lee 91c327cb44
fixed zero level 3 dtype bug (#76) 2021-12-20 17:00:53 +08:00
Frank Lee 35813ed3c4
update examples and sphnix docs for the new api (#63) 2021-12-13 22:07:01 +08:00
ver217 7d3711058f
fix zero3 fp16 and add zero3 model context (#62) 2021-12-10 17:48:50 +08:00
Frank Lee da01c234e1
Develop/experiments (#59)
* Add gradient accumulation, fix lr scheduler

* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)

* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes

* fixed trainer

* Revert "fixed trainer"

This reverts commit 2e0b0b7699.

* improved consistency between trainer, engine and schedule (#23)

Co-authored-by: 1SAA <c2h214748@gmail.com>

* Split conv2d, class token, positional embedding in 2d, Fix random number in ddp
Fix convergence in cifar10, Imagenet1000

* Integrate 1d tensor parallel in Colossal-AI (#39)

* fixed 1D and 2D convergence (#38)

* optimized 2D operations

* fixed 1D ViT convergence problem

* Feature/ddp (#49)

* remove redundancy func in setup (#19) (#20)

* use env to control the language of doc (#24) (#25)

* Support TP-compatible Torch AMP and Update trainer API (#27)

* Add gradient accumulation, fix lr scheduler

* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)

* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes

* fixed trainer

* Revert "fixed trainer"

This reverts commit 2e0b0b7699.

* improved consistency between trainer, engine and schedule (#23)

Co-authored-by: 1SAA <c2h214748@gmail.com>

Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>

* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)

* add explanation for ViT example (#35) (#36)

* support torch ddp

* fix loss accumulation

* add log for ddp

* change seed

* modify timing hook

Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>

* Feature/pipeline (#40)

* remove redundancy func in setup (#19) (#20)

* use env to control the language of doc (#24) (#25)

* Support TP-compatible Torch AMP and Update trainer API (#27)

* Add gradient accumulation, fix lr scheduler

* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)

* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes

* fixed trainer

* Revert "fixed trainer"

This reverts commit 2e0b0b7699.

* improved consistency between trainer, engine and schedule (#23)

Co-authored-by: 1SAA <c2h214748@gmail.com>

Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>

* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)

* add explanation for ViT example (#35) (#36)

* optimize communication of pipeline parallel

* fix grad clip for pipeline

Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>

* optimized 3d layer to fix slow computation ; tested imagenet performance with 3d; reworked lr_scheduler config definition; fixed launch args; fixed some printing issues; simplified apis of 3d layers (#51)

* Update 2.5d layer code to get a similar accuracy on imagenet-1k dataset

* update api for better usability (#58)

update api for better usability

Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
Co-authored-by: puck_WCR <46049915+WANG-CR@users.noreply.github.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
Co-authored-by: アマデウス <kurisusnowdeng@users.noreply.github.com>
Co-authored-by: BoxiangW <45734921+BoxiangW@users.noreply.github.com>
2021-12-09 15:08:29 +08:00