[OPT] Reduce update_weight peek memory in RL #1306
Open
+155
−61
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
hhaAndroid
approved these changes
Nov 28, 2025
HAOCHENYE
reviewed
Nov 28, 2025
xtuner/v1/model/base.py
Outdated
|
|
||
| # Concatenate the tensors along the FSDP shard dim | ||
| for tensors, size in zip(_fsdp_unsharded_tensor_list, origin_fsdp_size): | ||
| # special case for partition of tensors are contiguous |
Collaborator
There was a problem hiding this comment.
The comment should describe why rather than how
xtuner/v1/model/base.py
Outdated
Comment on lines
1182
to
1188
| if ( | ||
| t.untyped_storage().data_ptr() != storage.data_ptr() | ||
| or t.dtype != dtype | ||
| or t.device != device | ||
| or t.stride()[1:] != inner_stride | ||
| ): | ||
| return None |
Collaborator
There was a problem hiding this comment.
If we made it a private function, we should remove some unnecessary check
HAOCHENYE
reviewed
Nov 28, 2025
xtuner/v1/rl/base/worker.py
Outdated
| ep_mesh: DeviceMesh = model.ep_mesh | ||
| ep_size = ep_mesh.size() | ||
| ep_group = ep_mesh.get_group() | ||
| ep_rank = dist.get_rank(group=ep_group) |
Collaborator
There was a problem hiding this comment.
Add comments describing what happened here.
hhaAndroid
force-pushed
the
opt_update_weight_mem
branch
from
3bc7871 to
11fb109
Compare
CyCle1024
force-pushed
the
opt_update_weight_mem
branch
3 times, most recently
from
803c5ef to
05df05c
Compare
…k instead of local rank in pg
CyCle1024
force-pushed
the
opt_update_weight_mem
branch
from
05df05c to
61e09a0
Compare
HAOCHENYE
approved these changes
Dec 5, 2025
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 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Outdated suggestions cannot be applied.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.
Key Change:
xtuner/v1/model/base.py:_get_fused_hf_param_get_hf_paramsin RL update weights.update_weight_bucket_size_in_gbto config pack tensor size in RL update weights.test_update_weightdefault usingep_size=4.skip_load_weights == Truein RolloutConfig, lmdeploy would load weights from training workers, this would save time in large scale cluster.