[mlir][mesh] Shardingcontrol #102598
Merged
[mlir][mesh] Shardingcontrol #102598
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
Adding halo_sizes and shard_dims_sizes to sharding. First spmdization of halo annotated sharding
fschlimb
requested review from
dcaballe,
nicolasvasilache and
rengolin
as code owners
/llvm-project/mlir/lib/Dialect/Mesh/IR/MeshOps.cpp:182:16:
error: unused variable 'inAxis' [-Werror,-Wunused-variable]
auto inAxis = dimSz % inShape.size();
^
The cleanest structure is probably at the Func dilaect extensions. It is not critical to make it like that now as everything in Mesh already depends on the Tensor dialect. But in general the mesh dialect itself and the general transforms like sharding propagation and spmdization should also only depend on |
sogartar
approved these changes
Aug 12, 2024
|
LLVM Buildbot has detected a new failure on builder Full details are available at: https://lab.llvm.org/buildbot/#/builders/35/builds/1822 Here is the relevant piece of the build log for the reference: |
rengolin
pushed a commit
that referenced
this pull request
Aug 21, 2024
…ion lib (#104913) Follow-up to #102598 : as discussed, move tensor sharding implementation into separate tensor extension lib. @sogartar @yaochengji, could you take a look at this PR?
cjdb
pushed a commit
to cjdb/llvm-project
that referenced
this pull request
Aug 23, 2024
…ion lib (llvm#104913) Follow-up to llvm#102598 : as discussed, move tensor sharding implementation into separate tensor extension lib. @sogartar @yaochengji, could you take a look at this PR?
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Outdated suggestions cannot be applied.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.
This is a fixed copy of #98145 (necessary after it got reverted).
@sogartar @yaochengji
This PR adds the following to #98145:
UpdateHaloOpaccepts amemref(instead of a tensor) and not returning a result to clarify its inplace-semanticsUpdateHaloOpacceptssplit_axisto allow multiple mesh-axes per tensor/memref-axis (similar tomesh.sharding)Shardinginterfacefor tensor operation (tensor.emptyfor now) moved from the tensor library to the mesh interface library.spmdizeuses features frommeshdialect. @rengolin agreed thattensorshould not depend onmeshso this functionality cannot live in atensors lib. The unfulfilled dependency caused the issues leading to reverting [mlir][mesh] adding shard-size control #98145. Such cases are generally possible and might lead to re-considering the current structure (like for tosa ops).Replacing
#mesh.shardingattribute with operationmesh.shardinghalo_sizesandsharded_dims_sizesmesh::MeshShardingWhat previously was
is now
and allows additional annotations to control the shard sizes:
mesh.shardop accepts additional optional attributeforce, useful for halo updatestensor.emptyreacting onsharded_dims_sizesandhalo_sizesin the shardingmesh.update_haloas a spmdized target for shardings withhalo_sizes