tests: Fix flaky test for NLLB-MoE #22880
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
|
The documentation is not available anymore as the PR was closed or merged. |
|
cc @ydshieh |
amyeroberts
approved these changes
Apr 21, 2023
| The biggest difference is the way the tokens are routed. NLLB-MoE uses a `top-2-gate` which means that blah blah blah blah. | ||
| In SwitchTransformers, once the masks are computed for each experts, we just index the current hidden_states with the routing mask, and feed the | ||
| correct tokens to the expert. However here, the implementation varies a lot as the fairseq repository used a different approach. | ||
| The biggest difference is the way the tokens are routed. NLLB-MoE uses a `top-2-gate` which means that for each input, only the top two experts are selected based on the |
There was a problem hiding this comment.
Thanks a lot for this 😅
ArthurZucker
approved these changes
Apr 21, 2023
| The biggest difference is the way the tokens are routed. NLLB-MoE uses a `top-2-gate` which means that blah blah blah blah. | ||
| In SwitchTransformers, once the masks are computed for each experts, we just index the current hidden_states with the routing mask, and feed the | ||
| correct tokens to the expert. However here, the implementation varies a lot as the fairseq repository used a different approach. | ||
| The biggest difference is the way the tokens are routed. NLLB-MoE uses a `top-2-gate` which means that for each input, only the top two experts are selected based on the |
There was a problem hiding this comment.
Thanks a lot for this 😅
Comment on lines
47
to
49
| highest predicted probabilities from the gating network, and the remaining experts are ignored. In SwitchTransformers, once the masks are computed for each expert, | ||
| we just index the current hidden_states with the routing mask, and feed the correct tokens to the expert. However here, the implementation varies a lot as the fairseq | ||
| repository used a different approach. |
There was a problem hiding this comment.
Suggested change
| highest predicted probabilities from the gating network, and the remaining experts are ignored. In SwitchTransformers, once the masks are computed for each expert, | |
| we just index the current hidden_states with the routing mask, and feed the correct tokens to the expert. However here, the implementation varies a lot as the fairseq | |
| repository used a different approach. | |
| highest predicted probabilities from the gating network, and the remaining experts are ignored. In `SwitchTransformers`, only the top-1 probabilities are computes, which means that tokens have less probability of being forwarded. Moreover, if a token is not routed to any expert, `SwitchTransformers` still adds its unmodified hidden states (kind of like a residual connexion) while they are masked in `NLLB`'s top-2 routing mecanism. |
|
Happy to! |
connor-henderson
force-pushed
the
fix-flaky-nllb_moe-test
branch
from
5e25bd3
to
0834d2a
Compare
novice03
pushed a commit
to novice03/transformers
that referenced
this pull request
Jun 23, 2023
* add test update and docs edits * docs edit suggestion
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.
What does this PR do?
Fixes #22464 (and added some light docs edits I happened to notice)
From my comment in the issue:
Looked into this and I think the flakiness is caused by the natural variability in the sparse MoE layers. Specifically that when they calculate which experts to use in the gating logic, they’re computing probabilities imperfectly for two different sets of inputs: one with prior inputs concatenated with the past key values and one with just the past key values.
The test usually passes cause magnitude of the difference is usually likely to be small. Notably, when the vocab size is increased this pass rate goes up (and vice versa) since the increased representational capacity can help the model make more accurate decisions about which experts to use for each input. For example, increasing the vocab size in the config from its current 99 to 999 increases the pass rate from ~80% to ~95%.
I think this flakiness is inherent in the sparse layers, but if I understand right the point of the test is to check the decoder uses the past properly, so I edited the test to use dense layers and moved to the rtol down to 1e-3 to be in line with the other models’ version of this check. Wrote a loop to test this on a 1000 passes and they all passed.
Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.
@ArthurZucker, @amyeroberts