[LoweringContext] Support an optimized parameter mapping for SPMD #8453
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
rpsilva-aws
force-pushed
the
rpsilva_lc_mapping
branch
from
8b8b652 to
f898629
Compare
|
Note: Conflicts with #8418. Will handle rebase order based on merge timing. |
…rpsilva_lc_mapping
…rpsilva_lc_mapping
…rpsilva_lc_mapping
tengyifei
approved these changes
Dec 5, 2024
|
Replacing with #8460 to fix the merge commit, wrong local branch. |
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Outdated suggestions cannot be applied.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.
Currently, the existing parameter mapping for the lowering context is not well suited for SPMD. In case of large models, it will cause a large synchronous bottleneck when transferring all device data to the host. This is caused by each ReplicateShardedData computation that gathers and reassembles each sharded data across multiple devices. This is by design, since it is expected to collect all parameters regardless of their allocation.
In this PR, we introduce a new mapping that does not invoke the sharded replication, but instead uses references to the device data. This is generally sufficient and preferred in most cases, where the user only wants to access the validate parameters (those that are not returned as
-1fromtensor_parameter_id, as 'fake' parameters).