327d07c44a
[autoparallel] add conv metainfo class for auto parallel ( #1796 )
...
* [fx] metainfo class for auto parallel
* [fx] add unit test for linear metainfo
* [fx] fix bwd param for linear
* [fx] modify unit test
* [fx] modify unit test
* [fx] modify import
* [fx] modify import
* [fx] modify import
* [fx] move meta profiler to auto parallel
* [fx] add conv metainfo class
* [fx] restore profiler
* [fx] restore meta profiler
* [autoparallel] modify unit test
* [fx] modify unit test
2022-11-07 16:15:35 +08:00
501a9e9cd2
[hotfix] polish flash attention ( #1802 )
2022-11-07 14:30:22 +08:00
c248800359
[kernel] skip tests of flash_attn and triton when they are not available ( #1798 )
2022-11-07 13:41:13 +08:00
e34e850a4c
[autoparallel]add essential CommActions for broadcast oprands ( #1793 )
2022-11-04 18:36:42 +08:00
05ce3d369f
[fx] Add linear metainfo class for auto parallel ( #1783 )
...
* [fx] metainfo class for auto parallel
* [fx] add unit test for linear metainfo
* [fx] fix bwd param for linear
* [fx] modify unit test
* [fx] modify unit test
* [fx] modify import
* [fx] modify import
* [fx] modify import
* [fx] move meta profiler to auto parallel
2022-11-04 10:55:09 +08:00
2c4c7b3618
[autoparallel] add getattr handler ( #1767 )
...
* [autoparallel] add getattr haandler
* polish code
* add extra processes for Parameters
* add unit test for param resharding cost
* add docstring and polish test
2022-11-03 12:31:33 +08:00
c6a1a62636
[hotfix] fix zero's incompatibility with checkpoint in torch-1.12 ( #1786 )
...
* [hotfix] fix zero's incompatibility with checkpoint in torch-1.12
* [zero] add cpu shard init
* [zero] add tiny example test
* [colo_tensor] fix bugs for torch-1.11
2022-11-02 16:11:34 +08:00
32c1b843a9
skip torchrec unittests if not installed ( #1790 )
2022-11-02 14:44:32 +08:00
0b8161fab8
updated tp layers
2022-11-02 12:19:38 +08:00
e859380bf7
[fx] support module with bias addition ( #1780 )
...
* [autoparallel] refactor tracer to fix bias addition issue
* [fx] support module with bias addition
* create bias_addition_module
* refactor file structure
* polish code
* fix unit test
2022-11-01 22:53:51 +08:00
f3f19a5c47
[autoparallel] added matmul handler ( #1763 )
...
* [autoparallel] added matmul handler
* polish code
2022-11-01 15:14:53 +08:00
27de252334
[autoparallel] fix conv handler numerical test ( #1771 )
2022-11-01 10:43:44 +08:00
1e88811c7a
[autoparallel] move ckpt solvers to autoparallel folder / refactor code ( #1764 )
...
* [autoparallel] first move.
* [autoparallel] add solver rotor.
* [autoparallel] add ckpt solvers.
* [autoparallel] modify codegen.
* [fx] fix annotation in test.
* [fx] remove check.
* [autoparallel] polish docstring.
* [fx] refactor MetaTensor.
2022-11-01 10:43:15 +08:00
a4d1f59c78
[autoparallel] add numerical test for handlers ( #1769 )
2022-10-28 10:59:59 +08:00
b0f7c8bde8
[autoparallel] update CommSpec to CommActions ( #1768 )
...
* [autoparallel] update CommSpec to CommActions
* polish code
2022-10-28 09:57:43 +08:00
b4cc59b61e
[autoparallel] add numerical test for node strategies ( #1760 )
...
* [autoparallel] add numerical test for node strategies
* polish code
* polish code
2022-10-27 10:42:54 +08:00
25952b67d7
[feat] add flash attention ( #1762 )
2022-10-26 16:15:52 +08:00
0584654c79
[fx] refactor memory utils and extend shard utils. ( #1754 )
...
* [fx] change memory.py to memory_utils.py.
* [fx] add shard utils.
* [fx] fix import.
* [fx] check code style.
* [fx] add comment.
* [autoparallel] first move.
* [fx] add time computations.
2022-10-26 14:24:41 +08:00
314d8c497f
[autoparallel] refactor the runtime apply pass and add docstring to passes ( #1757 )
...
* [autoparallel] refactor the runtime apply pass and add doc string to passes
* fix unit test
* polish
2022-10-25 14:32:22 +08:00
f9a613d660
[autoparallel] added binary elementwise node handler ( #1758 )
...
* [autoparallel] added binary elementwise node handler
* polish code
2022-10-25 14:32:01 +08:00
d2fc067231
[autoparallel] fix param hook issue in transform pass ( #1755 )
2022-10-24 13:13:38 +08:00
262652c8bc
[autoparallel] added addbmm handler ( #1751 )
2022-10-21 18:55:48 +08:00
980ed21723
[autoparallel] shard param and buffer as expected ( #1753 )
...
* [autoparallel] shard param and buffer as expected
* fix unit test issue
2022-10-21 15:45:13 +08:00
cdb7d5e7d2
[hotfix] autoparallel unit test ( #1752 )
2022-10-20 19:51:38 +08:00
a4ce180e85
[autoparallel] add sequential order to communication actions ( #1735 )
2022-10-20 18:48:18 +08:00
b893342f95
[fx] test tracer on diffuser modules. ( #1750 )
...
* [fx] test tracer on diffuser modules.
* [fx] shorter seq_len.
* Update requirements-test.txt
2022-10-20 18:25:05 +08:00
b80b6eaa88
[autoparallel] recovered skipped test cases ( #1748 )
2022-10-20 16:37:33 +08:00
474111ecb5
[autoparallel] fixed wrong sharding strategy in conv handler ( #1747 )
...
* [autoparallel] fixed wrong sharding strategy in conv handler
* polish code
2022-10-20 16:12:39 +08:00
8b8937d901
[autoparallel] fixed wrong generated strategy for dot op ( #1746 )
...
* [autoparallel] fixed wrong generated strategy for dot op
* polish code
2022-10-20 15:18:16 +08:00
88a79814fb
[autoparallel] handled illegal strategy in node handler ( #1743 )
...
* [autoparallel] handled illegal strategy in node handler
* polish code
2022-10-19 17:08:52 +08:00
30874f1692
[fx/profiler] debug the fx.profiler / add an example test script for fx.profiler ( #1730 )
...
* [fx/profiler] add test.
* [fx] fix file names.
* [fx] add docstring and comment.
* [fx] polish profiler.py.
* [fx] fix import errors.
* [fx] fix profiler.
* [fx] fix names.
2022-10-19 14:24:51 +08:00
eee84908d4
[autoparallel] handled illegal sharding strategy ( #1728 )
...
* [autoparallel] handled illegal sharding strategy
* polish code
2022-10-19 12:53:06 +08:00
cbe9a4cb45
[NFC] polish tests/test_layers/test_3d/test_3d.py code style ( #1740 )
2022-10-19 12:20:51 +08:00
912eb58ea0
[NFC] polish tests/test_layers/test_3d/checks_3d/common.py code style ( #1733 )
2022-10-19 12:20:51 +08:00
754aa7c81f
[NFC] polish tests/test_layers/test_3d/checks_3d/check_layer_3d.py code style ( #1731 )
2022-10-19 12:20:51 +08:00
ff373a11eb
[NFC] polish tests/test_layers/test_sequence/checks_seq/check_layer_seq.py code style ( #1723 )
2022-10-19 12:20:51 +08:00
b38efe4e8a
[NFC] polish test_2p5d/checks_2p5d/check_operation_2p5d.py code style ( #1718 )
2022-10-19 12:20:51 +08:00
f6389d0813
[NFC] polish tests/test_layers/test_2d/checks_2d/check_operation_2d.py code style ( #1715 )
2022-10-19 12:20:51 +08:00
f69f9bf223
[zero] add chunk init function for users ( #1729 )
...
* add chunk manager init function
* fix unit tests
* add comment
* add flush=True
2022-10-18 16:31:22 +08:00
393f594051
[fx/meta/rpc] move _meta_registration.py to fx folder / register fx functions with compatibility checks / remove color debug ( #1710 )
...
* [fx] move meta registration
* [fx] fix tests.
* [fx] fix test.
* [fx] fix.
* [meta] refactor meta registration.py.
* [fx] add compatibility descriptions.
* [fx] polish import.
* [fx] add a decorator.
* [fx] fix tests.
* [fx] remove print.
* [fx] edit raise error.
* [fx] edit raise error.
* [fx] add type hint.
* [fx] fix import in experimental.
* [rpc] remove color debug.
* [meta] fix naming.
2022-10-18 10:44:23 +08:00
e8d8eda5e7
[autoparallel] moved tests to test_tensor_shard ( #1713 )
2022-10-17 13:54:20 +08:00
845ff4a47a
[autoparallel] resnet block runtime apply ( #1709 )
...
* [autoparallel] resnet block runtime apply
* seperate buffer and parameter in MemoryCost
* polish code
* add comments and todos
* fix test issue
2022-10-17 13:37:38 +08:00
22a115406b
[autoparallel] fixed broken node handler tests ( #1708 )
2022-10-14 18:25:59 +08:00
1468e4bcfc
[zero] add constant placement policy ( #1705 )
...
* fixes memory leak when paramter is in fp16 in ZeroDDP init.
* bans chunk releasement in CUDA. Only when a chunk is about to offload, it is allowed to release.
* adds a constant placement policy. With it, users can allocate a reserved caching memory space for parameters.
2022-10-14 17:53:16 +08:00
6c331a5a09
[autoparallel] refactored the autoparallel module for organization ( #1706 )
...
* [autoparallel] refactored the autoparallel module for organization
* polish code
2022-10-14 13:27:00 +08:00
91cd34e6e0
[unittest] added doc for the pytest wrapper ( #1704 )
2022-10-14 10:56:17 +08:00
451cd72dea
[autoparallel] adapt runtime passes ( #1703 )
...
* [autoparallel] adapt runtime passes v2
* polish code
2022-10-14 10:14:07 +08:00
21962e1593
[embedding] rename FreqAwareEmbedding -> CachedEmbedding ( #1699 )
2022-10-13 22:22:27 +08:00
0e52f3d3d5
[unittest] supported condititonal testing based on env var ( #1701 )
...
polish code
2022-10-13 19:38:45 +08:00
8283e95db3
[autoparallel] collated all deprecated files ( #1700 )
...
* [autoparallel] collated all deprecated files
* polish code
2022-10-13 18:24:11 +08:00
81f7530ee7
[autoparallel] adapt solver and CostGraph with new handler ( #1695 )
...
* [autoparallel] adapt solver and CostGraph with new handler
* fix test issue
2022-10-13 14:04:15 +08:00
42b882ef06
[autoparallel] add output handler and placeholder handler ( #1694 )
...
* [autoparallel] add output handler and placeholder handler
* Delete test_solver_with_resnet.py
* fix test bugs
2022-10-13 13:42:36 +08:00
56088e6d98
[autoparallel] add pooling handler ( #1690 )
...
* [autoparallel] add pooling handler
* polish code
2022-10-13 13:42:13 +08:00
319d654f79
[autoparallel] where_handler_v2 ( #1688 )
...
* where generator
* [autoparallel] where_handler_v2
2022-10-13 11:02:22 +08:00
31d2f03d27
[autoparallel] fix C version rotor inconsistency ( #1691 )
2022-10-12 15:21:58 +08:00
4973157ad7
[autoparallel] added sharding spec conversion for linear handler ( #1687 )
2022-10-12 11:16:18 +08:00
af718e83f2
[autoparallel] add reshape handler v2 and fix some previous bug ( #1683 )
2022-10-11 18:12:59 +08:00
3dd6994427
[fx/profiler] assigned UUID to each unrecorded tensor/ improved performance on GPT-2 ( #1679 )
...
* [fx/profiler] modify data_ptr into uuid for all tensors.
* [fx] modify uuid.
* [fx/profiler] tune performance on GPT-2.
* [fx] updates.
* [fx] debug.
* [fx] debug.
* [fx] cuda.
2022-10-11 11:03:35 +08:00
517b63939a
[autoparallel] add unary element wise handler v2 ( #1674 )
2022-10-09 17:30:42 +08:00
f6c6a932b8
[autoparallel] add following node generator ( #1673 )
...
* [autoparallel] add following node generator
* polish code
* polish code
* update name of arguments
2022-10-09 14:49:18 +08:00
52fda88796
[autoparallel] add layer norm handler v2 ( #1671 )
...
* [autoparallel] add layer norm handler v2
* polish code
* polish code
2022-10-09 14:23:22 +08:00
b28991dd0a
[feature] A new ZeRO implementation ( #1644 )
2022-10-09 09:18:51 +08:00
1df98d5b66
[autoparallel] add rotor C version ( #1658 )
...
* [autoparallel] add rotor c version
* [fx] remove metainfoprop in rotor solver
* [autoparallel] modify C
code format
* [autoparallel] remove build.py
* [autoparallel] fix C extension build
* [autoparallel] add C solver consistency test
* [autoparallel] remove some unused imports
* [autoparallel] refactor rotor solver code
* [autoparallel] replace print with colossalai logger
* [autoparallel] ranks fixed
2022-10-03 17:13:30 +08:00
11ec070e53
[hotfix]unit test ( #1670 )
2022-09-29 12:49:28 +08:00
a60024e77a
[autoparallel] added utils for broadcast operation ( #1665 )
...
* [autoparallel] added utils for broadcast operation
* polish code
2022-09-29 11:22:29 +08:00
3f068d1409
[autoparallel] update CommSpec ( #1667 )
2022-09-29 11:20:59 +08:00
746f8f979d
[autoparallel] add batch norm handler v2 ( #1666 )
2022-09-29 11:02:49 +08:00
9708638ded
[pipeline/pytree] add pytree to process args and kwargs | provide `data_process_func` to process args and kwargs after forward ( #1642 )
...
* [pipeline/tuning] improve dispatch performance both time and space cost
* [pipeline/converge] add interface for testing convergence
* [NFC] polish colossalai/utils/multi_tensor_apply/multi_tensor_apply.py code style
* Update PipelineBase.py
* [pipeline/chimera] reconstruct PipelineBase and Worker to support more feasible custom schedule | finish Chimera
* [pipeline/chimera] test chimera | fix bug of initializing
* [pipeline/pytree] add pytree to process args and kwargs | provide to process args and kwargs after forward
2022-09-29 10:58:58 +08:00
3a4d6f63a8
[autoparallel] added node handler for bmm ( #1655 )
2022-09-28 11:32:16 +08:00
095854477f
[autoparallel] add conv handler v2 ( #1663 )
2022-09-28 11:24:59 +08:00
1e7816a460
[autoparallel] adapt solver with gpt ( #1653 )
2022-09-28 11:17:26 +08:00
30e50c8b4a
[autoparallel] implemented all matmul strategy generator ( #1650 )
2022-09-27 12:06:25 +08:00
03978aad45
[autoparallel] change the following nodes strategies generation logic ( #1636 )
...
* [autoparallel] change the following nodes strategies generation logic
* fix unit test
2022-09-27 11:20:52 +08:00
59f100510a
[autoparallel] where handler ( #1651 )
...
* [autoparallel] where handler
* fix unit test
2022-09-27 11:20:43 +08:00
5d0fdb9cb4
[fx] fix offload codegen test ( #1648 )
...
* [fx] fix offload codegen test
* [fx] modify typing
2022-09-27 10:25:27 +08:00
45b39a692a
[autoparallel] implemented linear projection strategy generator ( #1639 )
2022-09-26 16:58:14 +08:00
154d3ef432
[fix] fixed the collective pattern name for consistency ( #1649 )
...
* [fix] fixed the collective pattern name for consistency
* polish code
2022-09-26 16:39:37 +08:00
b2b2a4af98
[autoparallel] adapt solver with mlp ( #1638 )
2022-09-26 15:26:14 +08:00
c5d39215f6
Revert "[feature] new zero implementation ( #1623 )" ( #1643 )
...
This reverts commit 5be118f405
2022-09-26 10:06:03 +08:00
5be118f405
[feature] new zero implementation ( #1623 )
2022-09-24 19:58:18 +08:00
95c35f73bd
[moe] initialize MoE groups by ProcessGroup ( #1640 )
2022-09-23 17:20:41 +08:00
a088022efc
[moe] fix moe bugs ( #1633 )
2022-09-23 15:33:57 +08:00
702dbc5288
[tensor] use communication autograd func ( #1617 )
...
* [tensor] use communication autograd func
* change all to all comm spec info
* rename pattern and distinguish fwd/bwd
* polish code
2022-09-23 13:31:15 +08:00
0c703189b9
[autoparallel] add layernorm handler ( #1629 )
2022-09-23 12:00:25 +08:00
bf77d3ab65
[autoparallel] recover the merged node strategy index ( #1613 )
2022-09-23 11:52:42 +08:00
d6b01feb66
[fx] Modify offload codegen ( #1618 )
...
* [fx] modify offload codegen
* [fx] remove repeated hook definitions
* [fx] modify offload test
2022-09-23 11:04:52 +08:00
9eae855408
[hotfix] add recompile after graph manipulatation ( #1621 )
2022-09-23 11:00:33 +08:00
d967779a32
[fx/profiler] tuned the calculation of memory estimation ( #1619 )
...
* [fx] tuned the meta info and rotor solver.
* [fx] remove import.
* [fx] remove import.
* [fx] remove import.
* [fx] tune the meta calculations.
* [fx] polish comments.
* [fx] remove assertions.
* [fx] modify test cases.
* [fx] modify test cases.
* [fx] optimize import.
* [fx
2022-09-23 10:59:47 +08:00
f7f2248771
[moe] fix MoE bugs ( #1628 )
...
* remove forced FP32 modules
* correct no_shard-contexts' positions
2022-09-22 13:56:30 +08:00
38c68b5b9a
[embedding] rollback for better FAW performance ( #1625 )
2022-09-22 11:16:25 +08:00
d925122020
[autoparallel] added new linear module handler ( #1616 )
2022-09-21 12:23:21 +08:00
170fa81095
[pipeline/chimera] test chimera | fix bug of initializing ( #1615 )
...
* [pipeline/tuning] improve dispatch performance both time and space cost
* [pipeline/converge] add interface for testing convergence
* [NFC] polish colossalai/utils/multi_tensor_apply/multi_tensor_apply.py code style
* Update PipelineBase.py
* [pipeline/chimera] reconstruct PipelineBase and Worker to support more feasible custom schedule | finish Chimera
* [pipeline/chimera] test chimera | fix bug of initializing
2022-09-20 18:00:39 +08:00
504ff1d101
[embeddings] use cache_ratio instead of cuda_row_num ( #1611 )
2022-09-20 14:33:04 +08:00
7d1bb71d5d
[fx] PoC of runtime shape consistency application ( #1607 )
...
* [fx] PoC of runtime shape consistency application
* polish code
2022-09-20 14:00:04 +08:00
47b11c432c
[autoparallel]add bcast matmul strategies ( #1605 )
2022-09-20 11:26:21 +08:00
933b6c6367
[fx] Add pofo solver ( #1608 )
...
* [fx] add pofo algorithm
* [fx] Add pofo solver
* [fx] code refactor
* [fx] fix test_linearize import
2022-09-20 11:20:48 +08:00
edc9e419ad
[pipeline/chimera] reconstruct PipelineBase and Worker to support more feasible custom schedule | finish Chimera ( #1595 )
...
* [pipeline/tuning] improve dispatch performance both time and space cost
* [pipeline/converge] add interface for testing convergence
* [NFC] polish colossalai/utils/multi_tensor_apply/multi_tensor_apply.py code style
* Update PipelineBase.py
* [pipeline/chimera] reconstruct PipelineBase and Worker to support more feasible custom schedule | finish Chimera
2022-09-19 11:44:18 +08:00
eac1b79371
[autoparallel] add bcast op handler ( #1600 )
...
* [autoparallel] add bcast op handler
* polish code
* add more BCAST FUNC OP
* polish code
* add exception handler
* polish
2022-09-16 11:33:01 +08:00
a7cda6f57d
[fx] Add offload codegen ( #1598 )
...
* [fx] add input activation offload to codegen
* [fx] modify unit test
* [fx] remove two skips in torch11
* [fx] use all_input_nodes instead of _input_nodes
2022-09-14 15:49:06 +08:00
c8e9b2ad78
[hotfix/rotor] fix variable names ( #1597 )
...
* [fx] add some comment and docstrings.
* [fx] add dataflow analysis for an autograd graph.
* add intepretation for graph analysis.
* [fx] before doing save_tensor_hooks.
* [fx] provide an accurate estimation of memory except for GPT-2.
* [fx] provide an accurate estimation of memory except for GPT-2.
* [fx] provide an accurate estimation of memory except for GPT-2.
* [fx] a very accurate version on GPT-2.
* [fx] refactor code.
* [fx] remove redundant inplace=True.
* [fx] refactor code.
* [fx] refactor code.
* [fx] refactor code.
* [fx] dive into backward memory.
* [fx] fix variable names in ckpt_solvers and unskip tests.
* [fx] commit my changes.
* [fx] restore skips.
* [fx] restore skips.
* [fx] chaange stage into phase.
* [fx] chaange stage into phase.
* [fx] chaange stage into phase.
2022-09-14 14:27:04 +08:00
Boyuan Yao