Place XLAShard data on CPU on creation #5032
Merged
+14
−19
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
JackCaoG
reviewed
May 18, 2023
jonb377
force-pushed
the
jonbolin-cpu-shard
branch
from
c777c1a to
6811b8e
Compare
JackCaoG
approved these changes
May 19, 2023
yeounoh
reviewed
May 20, 2023
| # Shards on the devices are materialized/available after the lazy | ||
| # execution of the SPMDPartitioned HLO graph. Each XLAShard points | ||
| # to torch.Tensor (xla::device_data). The shards represent a snapshot | ||
| # to torch.Tensor (xla::device_data). The shards represent a snapshot on CPU |
There was a problem hiding this comment.
nit. Could we update the comment here, "execution of the partitioned HLO graph. Each XLAShard points to torch.Tensor. The shards represent a snapshot on CPU, detached from the global tensor." ?
yeounoh
approved these changes
May 20, 2023
jonb377
force-pushed
the
jonbolin-cpu-shard
branch
2 times, most recently
from
1b605d7 to
9ceadd3
Compare
jonb377
force-pushed
the
jonbolin-cpu-shard
branch
from
9ceadd3 to
d967760
Compare
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Outdated suggestions cannot be applied.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.
As a follow-up from #5016, we should place the shards on CPU by default. This makes it clear to the user that the shards will not remain up-to-date with the underlying XLAShardedTensor, and it eases interoperability of the shards' data with other tensors.
Expanding on the second point, the shards are backed by PjRtData when returned from
_get_local_shards. In SPMD mode, we expect all computation inputs to have PjRtShardedData handles, which makes the on-device shards incompatible with SPMD execution. In order to use the shards in an on-device computation, they will now need to be transferred back to the device, which implies implicit replication.