feat: support for mcore optimizer (to enable MoE) #380
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
terrykong
force-pushed
the
tk/dpo-moe-fix
branch
2 times, most recently
from
56163e1 to
4fd8963
Compare
terrykong
changed the title
|
The DPO dataset changes should stand on their own, but are needed to test the mcore opt changes for moe. If moe issues take too long to resolve, I'll break this up. |
terrykong
force-pushed
the
tk/dpo-moe-fix
branch
3 times, most recently
from
9a48bf1 to
b9cf184
Compare
terrykong
changed the title
terrykong
changed the title
This was referenced Nov 7, 2024
terrykong
force-pushed
the
tk/dpo-moe-fix
branch
from
e3c9c88 to
c18de8d
Compare
terrykong
force-pushed
the
tk/dpo-moe-fix
branch
7 times, most recently
from
b1bca36 to
ccce0b3
Compare
terrykong
force-pushed
the
tk/dpo-moe-fix
branch
from
ccce0b3 to
7e93e37
Compare
ashors1
previously approved these changes
Nov 22, 2024
terrykong
added
Run CICD
terrykong
force-pushed
the
tk/dpo-moe-fix
branch
3 times, most recently
from
e897fd7 to
eab96f2
Compare
Signed-off-by: Terry Kong <[email protected]> moe test is all2all Signed-off-by: Terry Kong <[email protected]> other params Signed-off-by: Terry Kong <[email protected]> fix peft mixtral Signed-off-by: Terry Kong <[email protected]> dockerfile bump to be on dev Signed-off-by: Terry Kong <[email protected]> just take dockerfile on dev Signed-off-by: Terry Kong <[email protected]>
terrykong
force-pushed
the
tk/dpo-moe-fix
branch
from
eab96f2 to
1c98a08
Compare
Signed-off-by: Terry Kong <[email protected]>
Signed-off-by: Terry Kong <[email protected]>
terrykong
added
Run CICD
gshennvm
reviewed
Jan 10, 2025
| # from multiple simultaneous NCCL calls | ||
| ptl_model._optimizer._finish_bucket_grad_sync() | ||
| # Mcore DistOpt handles this, so we don't have to | ||
| if not ptl_model.use_mcore_dist_optim: |
There was a problem hiding this comment.
how do you feel about dropping support for non mcore dist optim? are they equivalent to apex now?
There was a problem hiding this comment.
yea, i want to do that in a follow up PR (would help our build times immensely). This just adds the feature without breaking apex
gshennvm
reviewed
Jan 10, 2025
gshennvm
reviewed
Jan 10, 2025
| @@ -150,7 +150,7 @@ def train_single_step(self, batch): | |||
| grad_norm = grad_norm.item() if torch.is_tensor(grad_norm) else grad_norm | |||
| lr = self.optimizer.param_groups[0]["lr"] | |||
|
|
|||
| self.optimizer.step() | |||
| self.optimizer.step(closure=None) | |||
There was a problem hiding this comment.
do we have to specify the closure now? i thought this was optional?
There was a problem hiding this comment.
for mcore dist opt it's required, so i just set it everywhere
Signed-off-by: Terry Kong <[email protected]>
terrykong
added
Run CICD
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.
What does this PR do ?
Rebase stack
Changelog
Usage
# Add a code snippet demonstrating how to use thisBefore your PR is "Ready for review"
Pre checks:
Checklist when contributing a new algorithm
max_steps=-1andvalidation?Additional Information