Add manual sharding API for SPMD #2
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Summary: This pull request makes SPMD support the manual sharding type via a new private API called: _mark_manual_sharding. I don't expect users will need to call this function explicitly. Besides adding support for the sharding annotation, we also need to define the behavior of the data shards. For data, the current behavior is error out. Test Plan: PJRT_DEVICE=TPU python test/spmd/test_xla_sharding.py -v -k test__mark_manual_sharding
Summary: This pull request supports SPMDFullToShardShape which is a custom op that opens a region for non-partitioned graph in SPMD program. It will stop SPMD auto sharding and partition in that region and therefore allows manual sharding like cc ops. To implement it, this pull request expands CustomSharding node to accept a new type. To be notice, the output shape of the op needs to be the shard shape of the input, and the node needs to have manual sharding annotation. Test Plan: PJRT_DEVICE=TPU python test/spmd/test_xla_sharding.py -v -k test_spmd_full_to_shard_shape
Summary: This pull request enables SPMDShardToFullShape. The trickiest part is how to get the full shape, and here is a couple of options: 1. Bookkeeping the shape full shape that enters SPMDFullToShardShape. This is not selected given the output could be created on the fly. 2. Constructing the full shape from the local shard and the sharding spec. This is not selected given there is no way to deal with the padding. We can't examine the data during the tracing time. 3. Let users pass the full shape in. This is selected because it's just the most sounded path. Tes Plan: PJRT_DEVICE=TPU python test/spmd/test_xla_sharding.py -v -k test_manual_sharding_e2e -k test_spmd_shard_to_full_shape
Summary: This pull request introduces: 1. enable_manual_sharding: which starts the manual sharding region. 2. disable_manual_sharding: which disable the manual sharding region. Test Plan: PJRT_DEVICE=TPU python test/spmd/test_xla_sharding.py -v -k test_manual_sharding_api_e2e
|
|
baoleai
approved these changes
Aug 6, 2024
yitongh
pushed a commit
that referenced
this pull request
Aug 8, 2024
Support Disc as backend Co-authored-by: yancey.yx <yancey.yx@antfin.com> Co-authored-by: wangang.wa <wangang.wa@alibaba-inc.com>
yitongh
pushed a commit
that referenced
this pull request
Aug 8, 2024
Support Disc as backend Co-authored-by: yancey.yx <yancey.yx@antfin.com> Co-authored-by: wangang.wa <wangang.wa@alibaba-inc.com>
anw90
pushed a commit
that referenced
this pull request
Oct 11, 2024
* build with BladeDISC (#8) * [to #53687860] feat: DISC client header, implement DISCComputation and DISCData POC implement in : https://code.alibaba-inc.com/torchx/xla/codereview/14984824 Link: https://code.alibaba-inc.com/torchx/xla/codereview/14987956 * Disc computation (#2) Support Disc as backend Co-authored-by: yancey.yx <yancey.yx@antfin.com> Co-authored-by: wangang.wa <wangang.wa@alibaba-inc.com> * add bazel flag to disable disc backend (#23) * add flag to disable disc backend in bazel workspace * support disc debug mode to dump mhlo and logs (#25) support disc backend debug mode to dump DISC compilation logs * support flash attention in disc (pytorch#34) * fix disc flag when complie python (pytorch#39) * fix bazel flag when complie python * fix lint. * support bf16 on disc backend (pytorch#40) add float-norm pass to support bf16 amp training * Support Flash Attention 2.5.6 for disc backend (#4) * fix build failed with NCCL (#5) * fix build failed on nccl * using nccl hdrs * Use the value of DISC_DEVICE as the device type of disc backend (#8) * change the device type of disc to cuda to make amp work properly * Use the value of DISC_DEVICE as the device type of disc backend * disable compilation of DISC by default (#15) --------- Co-authored-by: Yan Xu <yancey1989@gmail.com> Co-authored-by: wenting.swt <wenting.swt@alibaba-inc.com> Co-authored-by: Dalong <yuanxiulong.yxl@alibaba-inc.com> Co-authored-by: Baole Ai <baole.abl@alibaba-inc.com> Co-authored-by: Yan Xu <yancey.yx@alibaba-inc.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
cherry-pick from pytorch#6915, pytorch#6922, pytorch#6925, pytorch#6931