[release/2.7][ROCm][inductor] Improved fast_tanh code generation #2802
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
naromero77amd
requested review from
jataylo,
jeffdaily,
jithunnair-amd and
pruthvistony
as code owners
|
Jenkins build for 1b1fde5fcc342c2c0d3c69bf95a91501fc39b324 commit finished as FAILURE |
Author
|
I have confirmed that it resolves the reproducer in the Jira. |
This was referenced Nov 13, 2025
jataylo
reviewed
Nov 14, 2025
torch/_inductor/codegen/triton.py
Outdated
| return f"libdevice.tanh({x})" | ||
| # On ROCm, always use fast_tanhf | ||
| # Requires ROCm fork of Triton 3.3, 3.4, 3.5 or upstream Triton 3.6+ | ||
| if torch.version.hip: |
There was a problem hiding this comment.
2.7 uses 3.3 IIUC
We should at least support 3.2 in 2.7, so lets conditionalise on triton > (3,3) if 3.3 supports this.
|
Jenkins build for f416c7119ad1443bf022a37a8f3f21b201aa4bbc commit finished as FAILURE |
pruthvistony
approved these changes
Nov 17, 2025
pruthvistony
pushed a commit
that referenced
this pull request
Nov 17, 2025
In the ROCm fork of PyTorch 2.8, Inductor currently has codegen support for fast_tanhf. However, there were some NaN issues in the original Triton implementation of fast_tanhf . Upstream Triton has an improved fast_tanhf where the NaN issues are now fixed. This upstream commit has been backported to ROCm fork of Triton (see code comments). A bump in the Triton commit is also needed. Other notes: - In support of [SWDEV-560271](https://ontrack-internal.amd.com/browse/SWDEV-560271) - Triton 3.4 backport of upstream Triton commit ROCm/triton#900 - Similar to #2802, #2804 - Related to pytorch#162052
pruthvistony
pushed a commit
that referenced
this pull request
Nov 17, 2025
In the ROCm fork of PyTorch 2.9, Inductor currently has codegen support for fast_tanhf. However, there were some NaN issues in the original Triton implementation of fast_tanhf . Upstream Triton has an improved fast_tanhf where the NaN issues are now fixed. This upstream commit has been backported to ROCm fork of Triton (see code comments). A bump in the Triton commit is also needed. Other notes: - In support of [SWDEV-560271](https://ontrack-internal.amd.com/browse/SWDEV-560271) - Triton 3.5 backport of upstream Triton commit ROCm/triton#901 - Similar to #2802, #2803 - Related to pytorch#162052
jeffdaily
pushed a commit
that referenced
this pull request
Nov 17, 2025
In the ROCm fork of PyTorch 2.9, Inductor currently has codegen support for fast_tanhf. However, there were some NaN issues in the original Triton implementation of fast_tanhf . Upstream Triton has an improved fast_tanhf where the NaN issues are now fixed. This upstream commit has been backported to ROCm fork of Triton (see code comments). A bump in the Triton commit is also needed. Other notes: - In support of [SWDEV-560271](https://ontrack-internal.amd.com/browse/SWDEV-560271) - Triton 3.5 backport of upstream Triton commit ROCm/triton#901 - Similar to #2802, #2803 - Related to pytorch#162052
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.
In the ROCm fork of PyTorch 2.7, Inductor currently has codegen support for fast_tanhf. However, it is currently guarded by
TORCHINDUCTOR_USE_FAST_MATHenvironment variable due to some NaN issues in the original Triton implementation of fast_tanhf.Upstream Triton has an improved fast_tanhf where the NaN issues are now fixed. This upstream commit has been backported to ROCm fork of Triton (see code comments).
Thus, I have removed the conditionalization on Triton versions as well. A bump in the Triton commit is also needed.
Other notes: