Add all-gather coalescing for FSDP/ZeRO1 #5950
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
Also allow using reduce-scatter's scale param in FSDP. (revived #4145)
…ter tuple change without token
|
@jeffhataws let me know when you are done addressing comments, I will take another look |
JackCaoG
reviewed
Nov 30, 2023
JackCaoG
reviewed
Nov 30, 2023
JackCaoG
reviewed
Nov 30, 2023
| @@ -295,6 +295,7 @@ def __init__( | |||
| sharding_world_size: Optional[int] = None, | |||
| shard_param_on_dim_0: bool = False, | |||
| pin_layout_in_collective_ops: bool = True, | |||
| coalesce_all_gather_ops: bool = False, | |||
There was a problem hiding this comment.
Do you mind explaining the change in this file? I think coalesce_all_gather_ops is always False in our test, did you run into these issues with your own test?
There was a problem hiding this comment.
When the coalesce_all_gather_ops is True, the parameter shards are collected into a list and gathered in one all-gather coalesced command at the end (instead of all-gather one parameter at a time).
It is off by default to avoid changing existing behavior. The code is same as what we are using in our local fork.
JackCaoG
reviewed
Nov 30, 2023
| ReduceContext cc_ctx = GetReduceContext(inputs); | ||
| std::vector<xla::XlaOp> result(inputs.size()); | ||
|
|
||
| for (auto& type_ctx : cc_ctx.contexts) { |
There was a problem hiding this comment.
if you want to assume there is only one type_ctx, let's not use the for loop and GetReduceContext at all. This way we don't need to handle the token per type.
There was a problem hiding this comment.
Let me check with others on this.
JackCaoG
reviewed
Nov 30, 2023
JackCaoG
approved these changes
Dec 2, 2023
There was a problem hiding this comment.
Thanks! I think we should test allgather_coalesced using resnet on gpu to make sure we don't break it in the future. You can refer to existing test
Line 136 in 2c4983d
we can do that in a separate pr.
This was referenced Dec 5, 2023
Merged
jeffhataws
added a commit
to jeffhataws/xla
that referenced
this pull request
Dec 8, 2023
* Add all-gather and reduce-scatter coalescence support for FSDP. Also allow using reduce-scatter's scale param in FSDP. (revived pytorch#4145) * clang-format-7 and python lint fixes * Fix "SyntaxError: 'return' outside function" error * Code/test fixes to get run_tests.sh to run on CPU * Fix allgather to be compatible with openxla allgather tuple change without token * Fix reduce-scatter-coalesce to be compatible with openxla reduce-scatter tuple change without token * Separate out the reduce-scatter-coalesce changes into a separate PR * Some cleanups * Add separate BuildAllGatherCoalesced builder and AllGatherCoalesced class * Use token_handler.GetInput to capture token * Clean up * Clean up * Switch to GetOperandListWithToken naming for func GetOperandList
jeffhataws
added a commit
to jeffhataws/xla
that referenced
this pull request
Dec 11, 2023
* Add all-gather and reduce-scatter coalescence support for FSDP. Also allow using reduce-scatter's scale param in FSDP. (revived pytorch#4145) * clang-format-7 and python lint fixes * Fix "SyntaxError: 'return' outside function" error * Code/test fixes to get run_tests.sh to run on CPU * Fix allgather to be compatible with openxla allgather tuple change without token * Fix reduce-scatter-coalesce to be compatible with openxla reduce-scatter tuple change without token * Separate out the reduce-scatter-coalesce changes into a separate PR * Some cleanups * Add separate BuildAllGatherCoalesced builder and AllGatherCoalesced class * Use token_handler.GetInput to capture token * Clean up * Clean up * Switch to GetOperandListWithToken naming for func GetOperandList
chunnienc
pushed a commit
to chunnienc/xla
that referenced
this pull request
Dec 14, 2023
* Add all-gather and reduce-scatter coalescence support for FSDP. Also allow using reduce-scatter's scale param in FSDP. (revived pytorch#4145) * clang-format-7 and python lint fixes * Fix "SyntaxError: 'return' outside function" error * Code/test fixes to get run_tests.sh to run on CPU * Fix allgather to be compatible with openxla allgather tuple change without token * Fix reduce-scatter-coalesce to be compatible with openxla reduce-scatter tuple change without token * Separate out the reduce-scatter-coalesce changes into a separate PR * Some cleanups * Add separate BuildAllGatherCoalesced builder and AllGatherCoalesced class * Use token_handler.GetInput to capture token * Clean up * Clean up * Switch to GetOperandListWithToken naming for func GetOperandList
golechwierowicz
pushed a commit
that referenced
this pull request
Jan 12, 2024
* Add all-gather and reduce-scatter coalescence support for FSDP. Also allow using reduce-scatter's scale param in FSDP. (revived #4145) * clang-format-7 and python lint fixes * Fix "SyntaxError: 'return' outside function" error * Code/test fixes to get run_tests.sh to run on CPU * Fix allgather to be compatible with openxla allgather tuple change without token * Fix reduce-scatter-coalesce to be compatible with openxla reduce-scatter tuple change without token * Separate out the reduce-scatter-coalesce changes into a separate PR * Some cleanups * Add separate BuildAllGatherCoalesced builder and AllGatherCoalesced class * Use token_handler.GetInput to capture token * Clean up * Clean up * Switch to GetOperandListWithToken naming for func GetOperandList
bhavya01
pushed a commit
that referenced
this pull request
Apr 22, 2024
* Add all-gather and reduce-scatter coalescence support for FSDP. Also allow using reduce-scatter's scale param in FSDP. (revived #4145) * clang-format-7 and python lint fixes * Fix "SyntaxError: 'return' outside function" error * Code/test fixes to get run_tests.sh to run on CPU * Fix allgather to be compatible with openxla allgather tuple change without token * Fix reduce-scatter-coalesce to be compatible with openxla reduce-scatter tuple change without token * Separate out the reduce-scatter-coalesce changes into a separate PR * Some cleanups * Add separate BuildAllGatherCoalesced builder and AllGatherCoalesced class * Use token_handler.GetInput to capture token * Clean up * Clean up * Switch to GetOperandListWithToken naming for func GetOperandList
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.
This PR adds all-gather coalescence support and use that in FSDP/ZeRO1 (replacing #5624). This PR is to be used in conjunction with openxla/xla#5740 .
A separate and related PR for reduce-scatter coalescence that also enables using reduce-scatter's scale param in FSDP is #5938.
This is a revival of #4145 . Will need to address the comments.