Skip to content

[release/2.12] Backport Upstream PR: Bump AOTriton to 0.12b (#184288)#3290

Open
xinyazhang wants to merge 1 commit into
release/2.12from
xinyazhang/backport-aotriton-0.12b-2.12
Open

[release/2.12] Backport Upstream PR: Bump AOTriton to 0.12b (#184288)#3290
xinyazhang wants to merge 1 commit into
release/2.12from
xinyazhang/backport-aotriton-0.12b-2.12

Conversation

@xinyazhang

Copy link
Copy Markdown

Notable new features:

  • BREAKING Varlen LSE tensor shape changes to (H, Total_seqlen)
  • Support head_dim != head_dim_v
  • Support use_deterministic_algorithims
  • Support seqused_k in test/test_varlen_attention.py
  • gfx1100 and gfx1151 promoted out of experimental
  • Partial FAv3 support on gfx950

Bug Fixes:

  • GQA kernel failed to read bias tensor with the right offset.

Known Issues

  • gfx950's Triton kernel has problem handling hdim=16's fwd, in addition to hdim=48/80's bwd.
  • Disables gfx90a's CK SDPA support due to GPU Segfault.

Pull Request resolved: pytorch#184288
Approved by: https://github.com/jeffdaily

@xinyazhang

Copy link
Copy Markdown
Author

Required to fix ROCM-20626 and ROCM-20683

@rocm-repo-management-api

rocm-repo-management-api Bot commented Jun 9, 2026

Copy link
Copy Markdown

Jenkins build for d977c5e70abe135be09abca2abee48050c0e619e commit finished as FAILURE
Links: Pipeline Overview / Build artifacts / Test Results

@xinyazhang xinyazhang changed the title [release/2.12] Backup Upstream PR: Bump AOTriton to 0.12b (#184288) [release/2.12] Backport Upstream PR: Bump AOTriton to 0.12b (#184288) Jun 9, 2026
@xinyazhang xinyazhang requested a review from jeffdaily June 10, 2026 18:58
Notable new features:

* **BREAKING** Varlen LSE tensor shape changes to (H, Total_seqlen)
* Support head_dim != head_dim_v
* Support `use_deterministic_algorithims`
* Support seqused_k in test/test_varlen_attention.py
* gfx1100 and gfx1151 promoted out of experimental
* Partial FAv3 support on gfx950

Bug Fixes:

* GQA kernel failed to read bias tensor with the right offset.

Known Issues

* gfx950's Triton kernel has problem handling hdim=16's fwd, in addition to hdim=48/80's bwd.
* Disables gfx90a's CK SDPA support due to GPU Segfault.

Pull Request resolved: pytorch#184288
Approved by: https://github.com/jeffdaily
@xinyazhang xinyazhang force-pushed the xinyazhang/backport-aotriton-0.12b-2.12 branch from d977c5e to a1ff90b Compare June 10, 2026 19:00
@rocm-repo-management-api

rocm-repo-management-api Bot commented Jun 10, 2026

Copy link
Copy Markdown

Jenkins build for a1ff90b34fdb5588cdee78d1df82065745f39a4c commit finished as FAILURE
Links: Pipeline Overview / Build artifacts / Test Results

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant