Add a workaround for stochastic rounding for AMD GPUs #3908
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
Summary: X-link: facebookresearch/FBGEMM#997 This diff contains a workaround for the stochastic rounding issue for the AMD GPUs. Problem: `quantize_store` calls `nearest_rounding_vector` instead of `stochastic_rounding_vector` when stochastic rounding is used because the `StochasticRoundingRNGState` pointer is a nullptr (https://fburl.com/code/kna14icj) We found that the `WeightRow` constructor also gets a null `StochasticRoundingRNGState` pointer (https://fburl.com/code/vyq53lia) When `WeightRow` is instantiated, we confirm that `stochastic_rounding` is true. `WeightRow` should receive `&state`, but instead it receives a nullptr. (https://fburl.com/code/o3kxgt4z) We suspect that the compiler might have optimized out the `StochasticRoundingRNGState` since it is only passed to `WeightRow` and not utilized anywhere else in the caller kernel. Workaround: We move the `StochasticRoundingRNGState` storage inside the `WeightRow` struct and pass a boolean to the `WeightRow` constructor instead. Differential Revision: D72201618
|
This pull request was exported from Phabricator. Differential Revision: D72201618 |
✅ Deploy Preview for pytorch-fbgemm-docs ready!
To edit notification comments on pull requests, go to your Netlify site configuration. |
|
This pull request has been merged in 8093e7d. |
q10
pushed a commit
to q10/FBGEMM
that referenced
this pull request
Apr 10, 2025
Summary: X-link: pytorch#3908 Pull Request resolved: facebookresearch/FBGEMM#997 This diff contains a workaround for the stochastic rounding issue for the AMD GPUs. Problem: `quantize_store` calls `nearest_rounding_vector` instead of `stochastic_rounding_vector` when stochastic rounding is used because the `StochasticRoundingRNGState` pointer is a nullptr (https://fburl.com/code/kna14icj) We found that the `WeightRow` constructor also gets a null `StochasticRoundingRNGState` pointer (https://fburl.com/code/vyq53lia) When `WeightRow` is instantiated, we confirm that `stochastic_rounding` is true. `WeightRow` should receive `&state`, but instead it receives a nullptr. (https://fburl.com/code/o3kxgt4z) We suspect that the compiler might have optimized out the `StochasticRoundingRNGState` since it is only passed to `WeightRow` and not utilized anywhere else in the caller kernel. Workaround: We move the `StochasticRoundingRNGState` storage inside the `WeightRow` struct and pass a boolean to the `WeightRow` constructor instead. Reviewed By: q10, yinbinm, jianyuh, xw285cornell, yoyoyocmu, joebos Differential Revision: D72201618 fbshipit-source-id: a2bc7f004ac5183c84eb0501ada6d848ebca17e1
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.
Summary:
X-link: https://github.com/facebookresearch/FBGEMM/pull/997
This diff contains a workaround for the stochastic rounding issue for
the AMD GPUs.
Problem:
quantize_storecallsnearest_rounding_vectorinstead ofstochastic_rounding_vectorwhen stochastic rounding is used becausethe
StochasticRoundingRNGStatepointer is a nullptr(https://fburl.com/code/kna14icj)
We found that the
WeightRowconstructor also gets a nullStochasticRoundingRNGStatepointer (https://fburl.com/code/vyq53lia)When
WeightRowis instantiated, we confirm thatstochastic_roundingistrue.
WeightRowshould receive&state, but instead it receives anullptr. (https://fburl.com/code/o3kxgt4z)
We suspect that the compiler might have optimized out the
StochasticRoundingRNGStatesince it is only passed toWeightRowand not utilized anywhere else in the caller kernel.
Workaround:
We move the
StochasticRoundingRNGStatestorage inside theWeightRowstruct and pass a boolean to theWeightRowconstructorinstead.
Differential Revision: D72201618