[Bugfix] Add stream synchronization before the scatter operation #73
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
chang-l
changed the title
chang-l
force-pushed
the
sync-scatter-stream
branch
from
097434c
to
8ba6cdf
Compare
linhu-nv
added
bug
|
/ok to test |
rapids-bot bot
pushed a commit
that referenced
this pull request
Nov 22, 2024
Migrated from: rapidsai/wholegraph#229 This PR is to add gather/scatter support 1D tensor on python level, as WholeGraph should support basic indexing operations for both 1D (array) and 2D (matrix) wholememory tensors. Without this PR, if with 1D wholememory tensor, gather/scatter op does not work, e.g., https://github.com/rapidsai/wholegraph/blob/0efba33835d6e4e104b5d7101a91e0ea55a6ca53/python/pylibwholegraph/pylibwholegraph/torch/tensor.py#L89 To test, run ``` pytest --cache-clear --import-mode=append tests/wholegraph_torch/ops/test_wholegraph_gather_scatter.py -s ``` **Remaining issue:** On my local test with single GPU, the test can pass. For multiGPU setup, gather op works fine, but 1D scatter seems not working as it would crash at: https://github.com/rapidsai/wholegraph/blob/2e963b98aa6027c300d60e839010d3dd8ca422eb/python/pylibwholegraph/pylibwholegraph/tests/wholegraph_torch/ops/test_wholegraph_gather_scatter.py#L108 with incorrect scatter outputs: `Indices where allclose fails: tensor([0., 0., 0., ..., 0., 0., 0.]) tensor([ 1435., 1439., 1443., ..., 257703., 257707., 257711.]) ` This would work if this bugfix is merged: #73 cc. @linhu-nv Authors: - Chang Liu (https://github.com/chang-l) Approvers: - https://github.com/linhu-nv - Alex Barghi (https://github.com/alexbarghi-nv) URL: #74
|
/ok to test |
|
/merge |
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 to address the issue from this PR: rapidsai/wholegraph#229, and it's only for the last scatter operation before the Python interface (not for all internal
scatter_funccalls)Since the output of the scatter operation could be on the host (e.g., when emb_device = 'cpu'), it is necessary to perform synchronization internally. This ensures users do not need to explicitly synchronize the compute stream before accessing the host memory.
Unlike the gather operation, where the output is always in device memory, host side synchronization is unnecessary.