Sharded SDPAFwdOp #2565
Merged
Sharded SDPAFwdOp #2565
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
Priya2698
reviewed
Jul 10, 2024
Priya2698
reviewed
Jul 10, 2024
csrc/ir/nodes.cpp
Outdated
| // Add back the device dim axis for outputs with a head dimension. | ||
| if (handle_device_dim) { | ||
| output = output.unsqueeze(0); | ||
| log_sumexp = log_sumexp.unsqueeze(0); |
There was a problem hiding this comment.
Why do we need the DID axis in log_sumexp?
There was a problem hiding this comment.
Good question, we only need it for the backwards, so we don't need to track the DID axis. I'll remove it!
wujingyue
approved these changes
Jul 10, 2024
csrc/root_domain_map.cpp
Outdated
| } | ||
| } | ||
| // Map D from any input (query/key/value) to output, logsumexp only. |
There was a problem hiding this comment.
Still needed given https://github.com/NVIDIA/Fuser/pull/2565/files#r1672758030?
|
!build |
|
!build |
|
Note: failing tests are unrelated and caused by |
jacobhinkle
pushed a commit
that referenced
this pull request
Jul 19, 2024
Temporarily add support for a sharded forward scaled dot product attention. Currently, we only support DID parallelization on the logical domain which requires us to split and parallelize an axis at the logical level see #2563. This is a hack until we support DID parallelization on the loop domain after which this PR can be reverted. Restrictions: 1. q,k,v inputs are manually sharded _before_ the SDPAFwdOp is created. We cannot rely on sharding propagation or sharding after the Fusion is created, because the dimension checks are called when the op is created. 2. Only the head dimension is sharded and all inputs and outputs have either a sharded head dimension or unshaded. 3. DID axis is the outermost axis. This is because during evaluation if we see 5 dimensions, it is assumed the first is the DID axis and is appropriately squeezed from the inputs and unsqueezed onto the outputs.
cowanmeg
added a commit
that referenced
this pull request
Aug 23, 2024
Adds temporary support for sharded backwards scaled dot product attention. Until #2563 is completed. Similar to #2565 Similar restrictions: 1. All necessary sharded inputs are manually sharded before the SDPABwdOp is created. We cannot rely on sharding propagation or sharding after the Fusion is created, because the dimension checks are called when the op is created. 2. Only the head dimension is sharded and all inputs and outputs have either a sharded head dimension or unshaded. 3. DID axis is the outermost axis. This is because during evaluation if we see 5 dimensions, it is assumed the first is the DID axis and is appropriately squeezed from the inputs and unsqueezed onto the outputs.
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.
Temporarily add support for a sharded forward scaled dot product attention.
Currently, we only support DID parallelization on the logical domain which requires us to split and parallelize an axis at the logical level see #2563. This is a hack until we support DID parallelization on the loop domain after which this PR can be reverted.
Restrictions: