Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[js/webgpu] Fix Tanh explosion #19201

Merged
merged 3 commits into from
Jan 25, 2024
Merged

[js/webgpu] Fix Tanh explosion #19201

merged 3 commits into from
Jan 25, 2024

Conversation

hujiajie
Copy link
Contributor

Description

$$\tanh(x)=\frac{e^x-e^{-x}}{e^x+e^{-x}}= \left\{ \begin{array}{cc} -\frac{1-e^{-2\cdot(-x)}}{1+e^{-2\cdot(-x)}}, & x<0 \\\ 0, & x=0 \\\ \frac{1-e^{-2x}}{1+e^{-2x}}, & x>0 \end{array} \right.$$

Motivation and Context

On some platforms,
$$\tanh(1000)=\frac{e^{1000}-e^{-1000}}{e^{1000}+e^{-1000}}$$ would produce NaN instead of 0.999... or 1 (imagine $e^{1000}=\infty$ and $\frac{\infty}{\infty}$ explodes).

@hujiajie
[js/webgpu] Fix Tanh explosion
aa85bdc 
### Description
```math
\tanh(x)=\frac{e^x-e^{-x}}{e^x+e^{-x}}=
\left\{
\begin{array}{cc}
-\frac{1-e^{-2\cdot(-x)}}{1+e^{-2\cdot(-x)}}, & x<0 \\
0, & x=0 \\
\frac{1-e^{-2x}}{1+e^{-2x}}, & x>0
\end{array}
\right.
```

### Motivation and Context
On some platforms,
$$\tanh(1000)=\frac{e^{1000}-e^{-1000}}{e^{1000}+e^{-1000}}$$ would
produce NaN instead of 0.999... or 1 (imagine $e^{1000}=\infty$ and
$\frac{\infty}{\infty}$ explodes).
@guschmue guschmue added the ep:WebGPU ort-web webgpu provider label Jan 19, 2024
@gyagp
Copy link

gyagp commented Jan 22, 2024

Related WebGPU issue has been filed as gpuweb/gpuweb#4458. Zhaoming Jiang will bring this to Working Group discussion, and hopefully this will be fixed in browser implementation.
Test cases here are still valuable.

guschmue
guschmue previously approved these changes Jan 23, 2024
View reviewed changes
@guschmue
Copy link
Contributor

so we want to merge it until fixed in chromium ?

@hujiajie
[js/webgpu] Fix Tanh explosion
d781d4f 
### Description
```math
\tanh(x)=\frac{e^x-e^{-x}}{e^x+e^{-x}}=
\left\{
\begin{array}{cc}
-\frac{1-e^{-2\cdot(-x)}}{1+e^{-2\cdot(-x)}}, & x<0 \\
0, & x=0 \\
\frac{1-e^{-2x}}{1+e^{-2x}}, & x>0
\end{array}
\right.
```

### Motivation and Context
On some platforms,
$$\tanh(1000)=\frac{e^{1000}-e^{-1000}}{e^{1000}+e^{-1000}}$$ would
produce NaN instead of 0.999... or 1 (imagine $e^{1000}=\infty$ and
$\frac{\infty}{\infty}$ explodes).
@hujiajie
Merge remote-tracking branch 'origin/main' into main
1495f69
@hujiajie hujiajie requested a review from fs-eire January 24, 2024 07:10
@gyagp
Copy link

gyagp commented Jan 24, 2024

According to the discussion in WebGPU Working Group this morning, they'd postpone the related spec change to Milestone 2, but we have no idea when Milestone 1 will be released. This means we may not see spec changes in a few months.
We'll try to fix this in Chromium, even if spec doesn't want to tighten the condition now, but it may still take some time. Even if Chromium has the fix, the workaround here is also useful as other browsers may not do related changes before related changes in spec.

@guschmue
Copy link
Contributor

/azp run ONNX Runtime Web CI Pipeline

@azure-pipelines Azure Pipelines
Copy link

Azure Pipelines successfully started running 1 pipeline(s).

@guschmue
Copy link
Contributor

/azp run Linux CPU CI Pipeline,Linux CPU Minimal Build E2E CI Pipeline,Linux GPU CI Pipeline,Linux GPU TensorRT CI Pipeline,Linux OpenVINO CI Pipeline,Linux QNN CI Pipeline,MacOS CI Pipeline,Windows ARM64 QNN CI Pipeline,Windows CPU CI Pipeline

@guschmue
Copy link
Contributor

/azp run Windows GPU CI Pipeline,Windows GPU TensorRT CI Pipeline,onnxruntime-binary-size-checks-ci-pipeline,orttraining-linux-ci-pipeline,orttraining-linux-gpu-ci-pipeline,orttraining-ortmodule-distributed,Windows x64 QNN CI Pipeline

@azure-pipelines Azure Pipelines
Copy link

Azure Pipelines could not run because the pipeline triggers exclude this branch/path.

@azure-pipelines Azure Pipelines
Copy link

Azure Pipelines successfully started running 1 pipeline(s).

@guschmue guschmue merged commit 5b06505 into microsoft:main Jan 25, 2024
38 checks passed
fs-eire pushed a commit that referenced this pull request Mar 15, 2024
@hujiajie @fs-eire
[js/webgpu] Fix Tanh explosion (#19201)
b03a1c5 
### Description
```math
\tanh(x)=\frac{e^x-e^{-x}}{e^x+e^{-x}}=
\left\{
\begin{array}{cc}
-\frac{1-e^{-2\cdot(-x)}}{1+e^{-2\cdot(-x)}}, & x<0 \\
0, & x=0 \\
\frac{1-e^{-2x}}{1+e^{-2x}}, & x>0
\end{array}
\right.
```

### Motivation and Context
On some platforms,
$$\tanh(1000)=\frac{e^{1000}-e^{-1000}}{e^{1000}+e^{-1000}}$$ would
produce NaN instead of 0.999... or 1 (imagine $e^{1000}=\infty$ and
$\frac{\infty}{\infty}$ explodes).
YUNQIUGUO pushed a commit that referenced this pull request Mar 29, 2024
@fs-eire @qjia7
@axinging @satyajandhyala @Honry @hujiajie @guschmue @martholomew @dependabot @segevfiner @ibelem
1.17.3 cherry-picks for ORT Web changes (#19926)
45ff957 
### Description
This PR is a preview of cherry-picks for ort-web to `rel-1.17.3` based
on `rel-1.17.2`.

<details>

<summary>Changes of ort-web to cherry-pick</summary>

The following commits are from main branch.

`o` stands for pick, and `x` stands for skip.
```
o   2e0a388 [js/webgpu] Add HardSigmoid support (#19215)
o   d226e40 [js/webgpu] set query type in onRunStart (#19202)
o   61610ff [js/webgpu] Add FusedConv clip test case (#18900)
o   a33b5bd [JS/WebGPU] Added Uniforms to SkipLayerNorm. (#18788)
o   591f90c [js/webgpu] Fix issue of timestamp query (#19258)
o   7252c6e [WebNN EP] Support WebNN async API with Asyncify (#19145)
o   5b06505 [js/webgpu] Fix Tanh explosion (#19201)
o   656ca66 [js/webgpu] Support uniforms for conv, conv transpose, conv grouped (#18753)
o   a3f0e24 [js/webgpu] Support f16 uniform (#19098)
o   9e69606 fix f16 for attention, enable slice and flatten for more types (#19262)
o   624b4e2 [js/webgpu] Remove enableShapesUniforms (#19279)
o   90883a3 [js/webgpu] Add hardSigmoid activation for fusedConv (#19233)
o   85cef0a [js/webgpu] Support capture and replay for jsep (#18989)
o   d73131c [js/webgpu] Use DataType as uniform cpu type (#19281)
o   dd1f6cc [js/webgpu] resolve codescan alert (#19343)
o   3a2ab19 [js/webgpu] Refactor createTensorShapeVariables (#18883)
o   efc17e7 [js/webgpu] Fix the undefined push error (#19366)
 x  50806a7 [js/web] support external data in npm test (#19377)
o   ccbe264 [js/webgpu] Add LeakyRelu activation for fusedConv (#19369)
o   5ff27ef [js/webgpu] support customop FastGelu (#19392)
 x  03be65e [js/web] fix types exports in package.json (#19458)
o   06269a3 [js/webgpu] allow uint8 tensors for webgpu (#19545)
o   dfeda90 [JS/WebGPU] Add MatMulNBits (#19446)
o   1b48054 [js/webgpu] Create Split indices helpers by rank, not by shape (#19554)
o   3fe2c13 [js] small fix to workaround formatter (#19400)
 x  70567a4 [js/web] use ApiTensor insteadof onnxjs Tensor in TensorResultValidator (#19358)
o   6e04e36 [js/common] upgrade tsc in common from 4.9.5 to 5.2.2 (#19317)
o   58f4921 [js] changes to allow Float16Array if any polyfill is available (#19305)
o   57d6819 [js/web] Fix fused-conv is not included in npm test (#19581)
o   ebd220b Misspelling in README.md (#19433)
o   38c3432 Bump ip from 1.1.8 to 1.1.9 in /js/react_native (#19582)
o   fe82fcc [js/webgpu] Fix Conv2DTransposeMatMul f16 compilation failure (#19596)
o   76a2a48 Bump ip from 1.1.8 to 1.1.9 in /js/react_native/e2e (#19583)
o   29b1106 [node] Switch to setImmediate to avoid starving the Node.js event loop (#19610)
o   ae3d73c [JS/WebGPU] Fix Split and Where to handle corner cases. (#19613)
o   aec2389 [js/webgpu] allows a ProgramInfo's RunData to use zero sized output (#19614)
o   bb43a0f [js/webgpu] minor fixes to make tinyllama work (#19564)
o   0edb035 [js/web] fix suite test list for zero sized tensor (#19638)
o   3cb81cd [js/common] move 'env.wasm.trace' to 'env.trace' (#19617)
o   e30618d [js/webgpu] use Headless for webgpu test by default (#19702)
o   f06164e [js/web] transfer input buffer back to caller thread (#19677)
 x  a788514 [js/web] dump debug logs for karma for diagnose purpose (#19785)
o   24b72d2 [JS/WebGPU] Preserve zero size input tensor dims. (#19737)
o   4538d31 [js/webgpu] expose a few properties in WebGPU API (#19857)
o   53de2d8 [js/webgpu] Enable GroupedConvVectorize path (#19791)
o   ed250b8 [JS/WebGPU] Optimize MatMulNBits (#19852)
 x  e771a76 [js/test] align web test runner flags with ort.env (#19790)
o   79e50ae [js/web] rewrite backend resolve to allow multiple EPs (#19735)
o   acb0df2 Fix #19931 broken Get Started link of "ONNX Runtime JavaScript API" page (#19932)
o   b29849a [js/common] fix typedoc warnings (#19933)
o   afdab62 Bump follow-redirects from 1.15.4 to 1.15.6 in /js/web (#19949)
o   28ad6c3 Bump follow-redirects from 1.15.4 to 1.15.6 in /js/node (#19951)
o   7e0d424 accumulate in fp32 for Reduce* (#19868)
o   4c6a6a3 [js/webgpu] Fix NAN caused by un-initialized buffer in instance-norm (#19387)
o   01c7aaf [js/webgpu] allow setting env.webgpu.adapter (#19940)
o   c45cff6 [js/webgpu] fix maxpool / fp16 (#19981)
```

</details>

<details>
<summary>Cherry-pick commandlines</summary>

```sh
git cherry-pick 2e0a388
git cherry-pick d226e40
git cherry-pick 61610ff
git cherry-pick a33b5bd
git cherry-pick 591f90c
git cherry-pick 7252c6e
git cherry-pick 5b06505
git cherry-pick 656ca66
git cherry-pick a3f0e24
git cherry-pick 9e69606
git cherry-pick 624b4e2
git cherry-pick 90883a3
git cherry-pick 85cef0a  #<<<<< Note: conflicts
git cherry-pick d73131c
git cherry-pick dd1f6cc
git cherry-pick 3a2ab19
git cherry-pick efc17e7
git cherry-pick ccbe264
git cherry-pick 5ff27ef
git cherry-pick 06269a3
git cherry-pick dfeda90
git cherry-pick 1b48054
git cherry-pick 3fe2c13
git cherry-pick 6e04e36
git cherry-pick 58f4921
git cherry-pick 57d6819
git cherry-pick ebd220b
git cherry-pick 38c3432
git cherry-pick fe82fcc
git cherry-pick 76a2a48
git cherry-pick 29b1106
git cherry-pick ae3d73c
git cherry-pick aec2389
git cherry-pick bb43a0f
git cherry-pick 0edb035
git cherry-pick 3cb81cd
git cherry-pick e30618d
git cherry-pick f06164e
git cherry-pick 24b72d2
git cherry-pick 4538d31
git cherry-pick 53de2d8
git cherry-pick ed250b8
git cherry-pick 79e50ae
git cherry-pick acb0df2
git cherry-pick b29849a
git cherry-pick afdab62
git cherry-pick 28ad6c3
git cherry-pick 7e0d424
git cherry-pick 4c6a6a3
git cherry-pick 01c7aaf
git cherry-pick c45cff6
```
</details>

<details>
<summary>Cherry-pick conflicts</summary>

- 85cef0a #18989
this change is for enabling graph capture feature for JSEP, and it is
done after ROCM EP enabled graph capture feature. However, the ROCM EP
graph capture feature is not cherry-picked in rel-1.17.2.
</details>

---------

Signed-off-by: dependabot[bot] <[email protected]>
Co-authored-by: Jiajia Qin <[email protected]>
Co-authored-by: Xu Xing <[email protected]>
Co-authored-by: satyajandhyala <[email protected]>
Co-authored-by: Yang Gu <[email protected]>
Co-authored-by: Wanming Lin <[email protected]>
Co-authored-by: Jiajie Hu <[email protected]>
Co-authored-by: Guenther Schmuelling <[email protected]>
Co-authored-by: Matttttt <[email protected]>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Co-authored-by: Segev Finer <[email protected]>
Co-authored-by: Belem Zhang <[email protected]>
siweic0 pushed a commit to siweic0/onnxruntime-web that referenced this pull request May 9, 2024
@hujiajie
[js/webgpu] Fix Tanh explosion (microsoft#19201)
d39b783 
### Description
```math
\tanh(x)=\frac{e^x-e^{-x}}{e^x+e^{-x}}=
\left\{
\begin{array}{cc}
-\frac{1-e^{-2\cdot(-x)}}{1+e^{-2\cdot(-x)}}, & x<0 \\
0, & x=0 \\
\frac{1-e^{-2x}}{1+e^{-2x}}, & x>0
\end{array}
\right.
```

### Motivation and Context
On some platforms,
$$\tanh(1000)=\frac{e^{1000}-e^{-1000}}{e^{1000}+e^{-1000}}$$ would
produce NaN instead of 0.999... or 1 (imagine $e^{1000}=\infty$ and
$\frac{\infty}{\infty}$ explodes).
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@jzm-intel jzm-intel jzm-intel left review comments

@guschmue guschmue guschmue approved these changes

@fs-eire fs-eire fs-eire approved these changes

Assignees
No one assigned
Labels
ep:WebGPU ort-web webgpu provider
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
5 participants
@hujiajie @gyagp @guschmue @fs-eire @jzm-intel