Improve numeric stability for LayerNorm on CUDA #40308
Closed
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
Improve numeric stability for LayerNorm on CUDA. This diff will slightly decrease the performance of LayerNorm but increase the numerical stability especially when the variance is small. Example. M = 1024, N = 32768 input = torch.randn(M, N, dtype=torch.float, device="cuda") * 0.01 + 100.0 Previously in this case, the computed variance is [0. , 0.000977, 0. , ..., 0. , 0. , 0. ] After this diff, the variance will be [1.002436e-04, 9.923753e-05, 1.014664e-04, ..., 1.015593e-04, 1.000437e-04, 1.009089e-04] This is is close to 1e-4 as we expected. The max relative difference compare to double result is 5.69883169e-05. Differential Revision: [D21993664](https://our.internmc.facebook.com/intern/diff/D21993664/) [ghstack-poisoned]
xiaomengy
added a commit
that referenced
this pull request
Jun 19, 2020
Improve numeric stability for LayerNorm on CUDA. This diff will slightly decrease the performance of LayerNorm but increase the numerical stability especially when the variance is small. Example. M = 1024, N = 32768 input = torch.randn(M, N, dtype=torch.float, device="cuda") * 0.01 + 100.0 Previously in this case, the computed variance is [0. , 0.000977, 0. , ..., 0. , 0. , 0. ] After this diff, the variance will be [1.002436e-04, 9.923753e-05, 1.014664e-04, ..., 1.015593e-04, 1.000437e-04, 1.009089e-04] This is is close to 1e-4 as we expected. The max relative difference compare to double result is 5.69883169e-05. Differential Revision: [D21993664](https://our.internmc.facebook.com/intern/diff/D21993664/) ghstack-source-id: 106260530 Pull Request resolved: #40308
Contributor
Author
|
link #40302 |
Improve numeric stability for LayerNorm on CUDA. This diff will slightly decrease the performance of LayerNorm but increase the numerical stability especially when the variance is small. Example. M = 1024, N = 32768 input = torch.randn(M, N, dtype=torch.float, device="cuda") * 0.01 + 100.0 Previously in this case, the computed variance is [0. , 0.000977, 0. , ..., 0. , 0. , 0. ] After this diff, the variance will be [1.002436e-04, 9.923753e-05, 1.014664e-04, ..., 1.015593e-04, 1.000437e-04, 1.009089e-04] This is is close to 1e-4 as we expected. The max relative difference compare to double result is 5.69883169e-05. The average running time for F.layer_norm with input size [4096, 32768] on one devgpu before this PR is about 5.18ms, and after this PR is about 5.35ms. So the influence on performance should be limited. Differential Revision: [D21993664](https://our.internmc.facebook.com/intern/diff/D21993664/) [ghstack-poisoned]
xiaomengy
added a commit
that referenced
this pull request
Jun 20, 2020
Pull Request resolved: #40308 Improve numeric stability for LayerNorm on CUDA. This diff will slightly decrease the performance of LayerNorm but increase the numerical stability especially when the variance is small. Example. M = 1024, N = 32768 input = torch.randn(M, N, dtype=torch.float, device="cuda") * 0.01 + 100.0 Previously in this case, the computed variance is [0. , 0.000977, 0. , ..., 0. , 0. , 0. ] After this diff, the variance will be [1.002436e-04, 9.923753e-05, 1.014664e-04, ..., 1.015593e-04, 1.000437e-04, 1.009089e-04] This is is close to 1e-4 as we expected. The max relative difference compare to double result is 5.69883169e-05. ghstack-source-id: 106291482 Differential Revision: [D21993664](https://our.internmc.facebook.com/intern/diff/D21993664/)
Collaborator
|
Can you please profile performance with sizes from this script? I'm concerned about using int64 for computation, it might slow down the kernels considerably. |
Improve numeric stability for LayerNorm on CUDA. This diff will slightly decrease the performance of LayerNorm but increase the numerical stability especially when the variance is small. Example. M = 1024, N = 32768 input = torch.randn(M, N, dtype=torch.float, device="cuda") * 0.01 + 100.0 Previously in this case, the computed variance is [0. , 0.000977, 0. , ..., 0. , 0. , 0. ] After this diff, the variance will be [1.002436e-04, 9.923753e-05, 1.014664e-04, ..., 1.015593e-04, 1.000437e-04, 1.009089e-04] This is is close to 1e-4 as we expected. The max relative difference compare to double result is 5.69883169e-05. The average running time for F.layer_norm with input size [4096, 32768] on one devgpu before this PR is about 5.18ms, and after this PR is about 5.35ms. So the influence on performance should be limited. Differential Revision: [D21993664](https://our.internmc.facebook.com/intern/diff/D21993664/) [ghstack-poisoned]
Improve numeric stability for LayerNorm on CUDA. This diff will slightly decrease the performance of LayerNorm but increase the numerical stability especially when the variance is small. Example. M = 1024, N = 32768 input = torch.randn(M, N, dtype=torch.float, device="cuda") * 0.01 + 100.0 Previously in this case, the computed variance is [0. , 0.000977, 0. , ..., 0. , 0. , 0. ] After this diff, the variance will be [1.002436e-04, 9.923753e-05, 1.014664e-04, ..., 1.015593e-04, 1.000437e-04, 1.009089e-04] This is is close to 1e-4 as we expected. The max relative difference compare to double result is 5.69883169e-05. The average running time for F.layer_norm with input size [4096, 32768] on one devgpu before this PR is about 5.18ms, and after this PR is about 5.35ms. So the influence on performance should be limited. Differential Revision: [D21993664](https://our.internmc.facebook.com/intern/diff/D21993664/) [ghstack-poisoned]
xiaomengy
added a commit
that referenced
this pull request
Aug 14, 2020
Pull Request resolved: #40308 Improve numeric stability for LayerNorm on CUDA. This diff will slightly decrease the performance of LayerNorm but increase the numerical stability especially when the variance is small. Example. M = 1024, N = 32768 input = torch.randn(M, N, dtype=torch.float, device="cuda") * 0.01 + 100.0 Previously in this case, the computed variance is [0. , 0.000977, 0. , ..., 0. , 0. , 0. ] After this diff, the variance will be [1.002436e-04, 9.923753e-05, 1.014664e-04, ..., 1.015593e-04, 1.000437e-04, 1.009089e-04] This is is close to 1e-4 as we expected. The max relative difference compare to double result is 5.69883169e-05. ghstack-source-id: 109969199 Differential Revision: [D21993664](https://our.internmc.facebook.com/intern/diff/D21993664/)
💊 CI failures summary and remediationsAs of commit a118b98 (more details on the Dr. CI page): 💚 💚 Looks good so far! There are no failures yet. 💚 💚 This comment was automatically generated by Dr. CI (expand for details).Follow this link to opt-out of these comments for your Pull Requests.Please report bugs/suggestions on the GitHub issue tracker or post in the (internal) Dr. CI Users group. This comment has been revised 16 times. |
Contributor
Author
I tried to change the index summation to int32 type, but I didn't see great difference. On one V100 machine with input size [4096, 65536], the average running time of int32 and int64 is 4.375ms vs 4.384ms. |
Improve numeric stability for LayerNorm on CUDA. This diff will slightly decrease the performance of LayerNorm but increase the numerical stability especially when the variance is small. Example. M = 1024, N = 32768 input = torch.randn(M, N, dtype=torch.float, device="cuda") * 0.01 + 100.0 Previously in this case, the computed variance is [0. , 0.000977, 0. , ..., 0. , 0. , 0. ] After this diff, the variance will be [1.002436e-04, 9.923753e-05, 1.014664e-04, ..., 1.015593e-04, 1.000437e-04, 1.009089e-04] This is is close to 1e-4 as we expected. The max relative difference compare to double result is 5.69883169e-05. The average running time for F.layer_norm with input size [4096, 32768] on one devgpu before this PR is about 5.18ms, and after this PR is about 5.35ms. So the influence on performance should be limited. Differential Revision: [D21993664](https://our.internmc.facebook.com/intern/diff/D21993664/) [ghstack-poisoned]
Improve numeric stability for LayerNorm on CUDA. This diff will slightly decrease the performance of LayerNorm but increase the numerical stability especially when the variance is small. Example. M = 1024, N = 32768 input = torch.randn(M, N, dtype=torch.float, device="cuda") * 0.01 + 100.0 Previously in this case, the computed variance is [0. , 0.000977, 0. , ..., 0. , 0. , 0. ] After this diff, the variance will be [1.002436e-04, 9.923753e-05, 1.014664e-04, ..., 1.015593e-04, 1.000437e-04, 1.009089e-04] This is is close to 1e-4 as we expected. The max relative difference compare to double result is 5.69883169e-05. The average running time for F.layer_norm with input size [4096, 32768] on one devgpu before this PR is about 5.18ms, and after this PR is about 5.35ms. So the influence on performance should be limited. Differential Revision: [D21993664](https://our.internmc.facebook.com/intern/diff/D21993664/) [ghstack-poisoned]
xiaomengy
added a commit
that referenced
this pull request
Aug 15, 2020
Pull Request resolved: #40308 Improve numeric stability for LayerNorm on CUDA. This diff will slightly decrease the performance of LayerNorm but increase the numerical stability especially when the variance is small. Example. M = 1024, N = 32768 input = torch.randn(M, N, dtype=torch.float, device="cuda") * 0.01 + 100.0 Previously in this case, the computed variance is [0. , 0.000977, 0. , ..., 0. , 0. , 0. ] After this diff, the variance will be [1.002436e-04, 9.923753e-05, 1.014664e-04, ..., 1.015593e-04, 1.000437e-04, 1.009089e-04] This is is close to 1e-4 as we expected. The max relative difference compare to double result is 5.69883169e-05. ghstack-source-id: 110005133 Differential Revision: [D21993664](https://our.internmc.facebook.com/intern/diff/D21993664/)
Improve numeric stability for LayerNorm on CUDA. This diff will slightly decrease the performance of LayerNorm but increase the numerical stability especially when the variance is small. Example. M = 1024, N = 32768 input = torch.randn(M, N, dtype=torch.float, device="cuda") * 0.01 + 100.0 Previously in this case, the computed variance is [0. , 0.000977, 0. , ..., 0. , 0. , 0. ] After this diff, the variance will be [1.002436e-04, 9.923753e-05, 1.014664e-04, ..., 1.015593e-04, 1.000437e-04, 1.009089e-04] This is is close to 1e-4 as we expected. The max relative difference compare to double result is 5.69883169e-05. The average running time for F.layer_norm with input size [4096, 32768] on one devgpu before this PR is about 5.18ms, and after this PR is about 5.35ms. So the influence on performance should be limited. Differential Revision: [D21993664](https://our.internmc.facebook.com/intern/diff/D21993664/) [ghstack-poisoned]
xiaomengy
added a commit
that referenced
this pull request
Sep 13, 2020
Pull Request resolved: #40308 Improve numeric stability for LayerNorm on CUDA. This diff will slightly decrease the performance of LayerNorm but increase the numerical stability especially when the variance is small. Example. M = 1024, N = 32768 input = torch.randn(M, N, dtype=torch.float, device="cuda") * 0.01 + 100.0 Previously in this case, the computed variance is [0. , 0.000977, 0. , ..., 0. , 0. , 0. ] After this diff, the variance will be [1.002436e-04, 9.923753e-05, 1.014664e-04, ..., 1.015593e-04, 1.000437e-04, 1.009089e-04] This is is close to 1e-4 as we expected. The max relative difference compare to double result is 5.69883169e-05. ghstack-source-id: 111954875 Differential Revision: [D21993664](https://our.internmc.facebook.com/intern/diff/D21993664/)
Improve numeric stability for LayerNorm on CUDA. This diff will slightly decrease the performance of LayerNorm but increase the numerical stability especially when the variance is small. Example. M = 1024, N = 32768 input = torch.randn(M, N, dtype=torch.float, device="cuda") * 0.01 + 100.0 Previously in this case, the computed variance is [0. , 0.000977, 0. , ..., 0. , 0. , 0. ] After this diff, the variance will be [1.002436e-04, 9.923753e-05, 1.014664e-04, ..., 1.015593e-04, 1.000437e-04, 1.009089e-04] This is is close to 1e-4 as we expected. The max relative difference compare to double result is 5.69883169e-05. The average running time for F.layer_norm with input size [4096, 32768] on one devgpu before this PR is about 5.18ms, and after this PR is about 5.35ms. So the influence on performance should be limited. Differential Revision: [D21993664](https://our.internmc.facebook.com/intern/diff/D21993664/) [ghstack-poisoned]
xiaomengy
added a commit
that referenced
this pull request
Sep 14, 2020
Pull Request resolved: #40308 Improve numeric stability for LayerNorm on CUDA. This diff will slightly decrease the performance of LayerNorm but increase the numerical stability especially when the variance is small. Example. M = 1024, N = 32768 input = torch.randn(M, N, dtype=torch.float, device="cuda") * 0.01 + 100.0 Previously in this case, the computed variance is [0. , 0.000977, 0. , ..., 0. , 0. , 0. ] After this diff, the variance will be [1.002436e-04, 9.923753e-05, 1.014664e-04, ..., 1.015593e-04, 1.000437e-04, 1.009089e-04] This is is close to 1e-4 as we expected. The max relative difference compare to double result is 5.69883169e-05. ghstack-source-id: 111964204 Differential Revision: [D21993664](https://our.internmc.facebook.com/intern/diff/D21993664/)
Improve numeric stability for LayerNorm on CUDA. This diff will slightly decrease the performance of LayerNorm but increase the numerical stability especially when the variance is small. Example. M = 1024, N = 32768 input = torch.randn(M, N, dtype=torch.float, device="cuda") * 0.01 + 100.0 Previously in this case, the computed variance is [0. , 0.000977, 0. , ..., 0. , 0. , 0. ] After this diff, the variance will be [1.002436e-04, 9.923753e-05, 1.014664e-04, ..., 1.015593e-04, 1.000437e-04, 1.009089e-04] This is is close to 1e-4 as we expected. The max relative difference compare to double result is 5.69883169e-05. The average running time for F.layer_norm with input size [4096, 32768] on one devgpu before this PR is about 5.18ms, and after this PR is about 5.35ms. So the influence on performance should be limited. Differential Revision: [D21993664](https://our.internmc.facebook.com/intern/diff/D21993664/) [ghstack-poisoned]
xiaomengy
added a commit
that referenced
this pull request
Sep 24, 2020
Pull Request resolved: #40308 Improve numeric stability for LayerNorm on CUDA. This diff will slightly decrease the performance of LayerNorm but increase the numerical stability especially when the variance is small. Example. M = 1024, N = 32768 input = torch.randn(M, N, dtype=torch.float, device="cuda") * 0.01 + 100.0 Previously in this case, the computed variance is [0. , 0.000977, 0. , ..., 0. , 0. , 0. ] After this diff, the variance will be [1.002436e-04, 9.923753e-05, 1.014664e-04, ..., 1.015593e-04, 1.000437e-04, 1.009089e-04] This is is close to 1e-4 as we expected. The max relative difference compare to double result is 5.69883169e-05. ghstack-source-id: 112836419 Differential Revision: [D21993664](https://our.internmc.facebook.com/intern/diff/D21993664/)
Codecov Report
@@ Coverage Diff @@
## gh/BIT-silence/4/base #40308 +/- ##
=========================================================
- Coverage 68.06% 68.05% -0.01%
=========================================================
Files 393 393
Lines 50918 50918
=========================================================
- Hits 34655 34654 -1
- Misses 16263 16264 +1
Continue to review full report at Codecov.
|
Contributor
Author
|
#59987 should fix the issue. |
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.
Stack from ghstack:
Improve numeric stability for LayerNorm on CUDA.
This diff will slightly decrease the performance of LayerNorm but increase the numerical stability especially when the variance is small.
Example.
M = 1024, N = 32768
input = torch.randn(M, N, dtype=torch.float, device="cuda") * 0.01 + 100.0
Previously in this case, the computed variance is
[0. , 0.000977, 0. , ..., 0. , 0. , 0. ]
After this diff, the variance will be
[1.002436e-04, 9.923753e-05, 1.014664e-04, ..., 1.015593e-04, 1.000437e-04, 1.009089e-04]
This is is close to 1e-4 as we expected. The max relative difference compare to double result is 5.69883169e-05.
The average running time for F.layer_norm with input size [4096, 32768] on one devgpu before this PR is about 5.18ms, and after this PR is about 5.35ms. So the influence on performance should be limited.
Differential Revision: D21993664