Add ONNX parsing for SimplifiedLayerNormalization #3129
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
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## develop #3129 +/- ##
========================================
Coverage 92.26% 92.26%
========================================
Files 499 500 +1
Lines 20020 20048 +28
========================================
+ Hits 18471 18497 +26
- Misses 1549 1551 +2 ☔ View full report in Codecov by Sentry. |
Check results before merge 🔆 |
🔴bert_large_uncased_fp16: FAILED: MIGraphX is not within tolerance - check verbose output |
turneram
requested review from
umangyadav,
TedThemistokleous and
shivadbhavsar
and removed request for
TedThemistokleous
|
Where is this spec for this operator? |
CharlieL7
reviewed
Jun 11, 2024
| auto rms = info.add_instruction(make_op("reduce_mean", {{"axes", {axis}}}), x_sq); | ||
| auto mean = rms; | ||
| epsilon = | ||
| (x_dtype == migraphx::shape::half_type and std::abs(epsilon) < 1e-7) ? 1e-7 : epsilon; |
There was a problem hiding this comment.
Why are we limiting the epsilon for half type? It looks like a user input
There was a problem hiding this comment.
That is how we handle epsilon in our regular LayerNorm parser, so I did the same here.
There isn't actually one for SimplifiedLayerNormalization, but this is the spec for SkipSimplifiedLayerNormalization, which is just add + SLN. That spec does include an optional bias input, but neither of the ORT implementations utilize it, so I omitted it from ours. |
The equation is the same as RMS LayerNorm. |
TedThemistokleous
added
onnxruntime
umangyadav
reviewed
Jul 23, 2024
Comment on lines
+33
to
+48
| std::vector<half> x{half{0.8}, | ||
| half{-0.5}, | ||
| half{0.0}, | ||
| half{1.0}, | ||
| half{0.5}, | ||
| half{0.2}, | ||
| half{0.3}, | ||
| half{-0.6}, | ||
| half{10.0}, | ||
| half{-1.0}, | ||
| half{0.0}, | ||
| half{1.0}, | ||
| half{1.2}, | ||
| half{3.2}, | ||
| half{-4.1}, | ||
| half{5.3}}; |
There was a problem hiding this comment.
You shouldn't require casting all the elements to half. Same applies to all other places.
umangyadav
reviewed
Jul 23, 2024
| auto result = info.add_common_op("mul", x, rrms); | ||
| result = info.add_common_op("mul", result, scale); | ||
|
|
||
| return {result, mean, rrms}; |
There was a problem hiding this comment.
Is this being matched with LayerNorm kernel on the GPU target ?
TedThemistokleous
approved these changes
Aug 2, 2024
umangyadav
approved these changes
Aug 8, 2024
TedThemistokleous
pushed a commit
that referenced
this pull request
Aug 13, 2024
* Add simplified_layer_normalization
TedThemistokleous
pushed a commit
that referenced
this pull request
Aug 21, 2024
* Add simplified_layer_normalization
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.
No description provided.