[Dist Dialect] Simple MoE training in PIR #66750
Merged
Conversation
This file contains hidden or 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
pkuzyc
requested review from
Aurelius84,
changeyoung98,
cxxly,
cyber-pioneer and
xiaoguoguo626807
as code owners
|
你的PR提交成功，感谢你对开源项目的贡献! |
pkuzyc
force-pushed
the
pir_moe_pass
branch
2 times, most recently
from
089384d to
ff60be1
Compare
pkuzyc
force-pushed
the
pir_moe_pass
branch
2 times, most recently
from
45ae555 to
d1496fa
Compare
|
Sorry to inform you that d1496fa's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually. |
pkuzyc
force-pushed
the
pir_moe_pass
branch
3 times, most recently
from
24bdd87 to
bf6ddb8
Compare
pkuzyc
force-pushed
the
pir_moe_pass
branch
2 times, most recently
from
caa33b6 to
9a7205f
Compare
add sub_to_global_reshard function, the loss of PIR MoE demo is equal to dygraph auto parallel refine and rename the MoE apis
XieYunshen
approved these changes
Sep 11, 2024
zyfncg
approved these changes
Sep 11, 2024
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
6 participants
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Outdated suggestions cannot be applied.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.
PR Category
Auto Parallel
PR Types
New features
Description
Pcard-67164
Support simple MoE training in PIR:
Replace the MoE-related dist op, i.e.
moe_sub_mesh_tensorsandmoe_global_mesh_tensor([Dist Dialect] Add MoE-related api in PIR dist dialect #66462), with share_data op in remove_other_rank_op_pass.Add sub_to_global_mesh reshard function in PIR, now only supports replicated Tensor. This is needed in global_norm_clip, the l2 norm of each expert's parameter should be resharded to other ranks, e.g. reshard the l2 norm of the parameter on ProcessMesh([0]) to ProcessMesh([0,1]).
NOTE: this is a preliminary version, need further refine to make it consistent with the dygraph reshard function.
Refine the MoE-related apis for multi-dimension process mesh, e.g. mesh=[[0,1],[2,3]], one expert on ranks [0,2] and the other one on ranks [1,3].
Additoinal: add an unit test for reshape_grad when its output x_grad needs reshard, corresponding to #67729