Enable multi-device computation in runtime #1716
Merged
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
wooseokTT
requested review from
jnie-TT,
kmabeeTT,
AleksKnezevic,
pilkicTT,
sdjordjevicTT,
svuckovicTT,
mtopalovicTT,
nobradovictt,
jserbedzijaTT,
azecevicTT,
nsmithtt and
tapspatel
as code owners
wooseokTT
force-pushed
the
wooseok/enable_multidevice_runtime
branch
from
90ce1bf to
2a566ed
Compare
Closed
sdjordjevicTT
approved these changes
Jan 9, 2025
| @@ -275,7 +275,8 @@ class TTNNLayoutDPSOperandsRewriter | |||
| LogicalResult matchAndRewrite(DestinationStyleOpInterface op, | |||
| PatternRewriter &rewriter) const final { | |||
| // To layout op is a special case, we don't want to rewrite it | |||
| if (mlir::isa<ttir::ToLayoutOp>(op.getOperation())) { | |||
| if (mlir::isa<ttir::ToLayoutOp>(op.getOperation()) || | |||
| mlir::isa<ttir::MeshShardOp>(op.getOperation())) { | |||
There was a problem hiding this comment.
Any comment on why MeshShardOp is a special one in this regard?
There was a problem hiding this comment.
TTNN mesh shard APIs are currently CPU only operations.. So, by enforcing tensors to be located in system memory, we can ensure (1) a tensor can be sharded into multi-device storage in cpu side, and (2) later tiled and transferred to device to individual devices.
tapspatel
approved these changes
Jan 9, 2025
wooseokTT
force-pushed
the
wooseok/enable_multidevice_runtime
branch
2 times, most recently
from
2fd39a1 to
992997d
Compare
nsmithtt
approved these changes
Jan 10, 2025
wooseokTT
force-pushed
the
wooseok/enable_multidevice_runtime
branch
2 times, most recently
from
efddf68 to
f5bec77
Compare
jnie-TT
reviewed
Jan 14, 2025
jnie-TT
reviewed
Jan 14, 2025
jnie-TT
reviewed
Jan 14, 2025
jnie-TT
reviewed
Jan 14, 2025
jnie-TT
approved these changes
Jan 14, 2025
* Allow ttnn runtime operations including reduce_scatter, mesh_shard, and all_gather * Force mesh_shard ops to use system memory because they are host-side operations * Use strongly-typed sharding options for mesh_shard ops * Add Silicon multi-device test cases * Fix bug in determining axis of all_reduce when converting from stableHLO * Fix typo in ttnn workaround pass
wooseokTT
force-pushed
the
wooseok/enable_multidevice_runtime
branch
from
f5bec77 to
3e57fd9
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.
Enable multi-device computation in runtime