Skip to content

[AMD] Pin the exact ROCm Triton wheel instead of deferring to AITER - #34403

Draft
bingxche wants to merge 1 commit into
mainfrom
amd/rocm720-pin-triton-wheel
Draft

bingxche wants to merge 1 commit into
mainfrom
amd/rocm720-pin-triton-wheel

Conversation

@bingxche

@bingxche bingxche commented Aug 11, 2026

Copy link
Copy Markdown
Collaborator

Motivation

#34364 handed the ROCm 7.2 image's Triton to AITER's install_triton.sh, so an unversioned helper script in another repo decides what the image ships. AITER does not claim that role — its requirements.txt says torch and triton are "provided by the base image ... their provenance is governed by the base image, not this file", and the only thing it enforces is triton>=3.6.0.

Modifications

Install the wheels directly from AMD's ROCm index, pinned in docker/rocm.Dockerfile as TRITON_INDEX_URL / TRITON_VERSION / TRITON_KERNELS_VERSION, so the version is visible and bumping it is a reviewed edit. The local version is part of the pin: triton==3.7.0 accepts any revision the index later publishes under that number, and when two exist pip picks by lexical order of the git hash rather than by date, so a republish could silently move the image forward or backward.

Same wheels as today, both triton-lang/triton@89002410 — this changes who decides the version, not what the image currently ships.


CI States

Latest PR Test (Base): ✅ Run #31481303981
Latest PR Test (Extra): ❌ Run #31481303711
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.

#34364 delegated the ROCm 7.2 image's Triton to AITER's install_triton.sh, which
means an unversioned helper script in another repo decides what the image ships.
AITER does not claim that role: its requirements.txt says torch and triton are
"provided by the base image ... their provenance is governed by the base image,
not this file", and all it enforces is triton>=3.6.0. So pin it here, where the
choice is visible and changing it is a reviewed edit.

The local version is part of the pin. `triton==3.7.0` accepts any revision the
AMD index later publishes under that number, and when two exist pip picks by
lexical order of the git hash, not by date - so a republish could silently move
the image either forward or backward.

Same wheels as before, both triton-lang/triton@89002410: no change to what the
image ships today, only to who decides it.
akao-amd added a commit to akao-amd/sglang that referenced this pull request Aug 13, 2026
…rocm7_15

rocm.Dockerfile:
- Torch metadata patch: replace wheel-file-based hack with in-place
  METADATA patching via importlib.metadata. Works for both rocm720
  (pre-baked wheel) and rocm7_15 (pip-installed torch). Handles bare
  continuation lines (no Requires-Dist: prefix) seen in ROCm wheels.
- Remove old triton rebuild block that faked version strings.
- Adopt PR sgl-project#34403: add TRITON_INDEX_URL/TRITON_VERSION/TRITON_KERNELS_VERSION
  ARGs for rocm720 pinned wheels from pypi.amd.com.
- Merge triton install: rocm7_15 builds from source at TRITON_COMMIT;
  rocm720 installs pinned wheels from TRITON_INDEX_URL.

amd_ci_install_dependency.sh:
- Keep ROCm version detection from HEAD for 7.15 pip install path.
- Add AITER triton pre-install (install_triton.sh) in the rocm7.2 path.

Co-Authored-By: Claude <noreply@anthropic.com>
chuyeh added a commit to chuyeh/sglang that referenced this pull request Aug 19, 2026
Drop the helper scripts so the image build does not call files outside
the Dockerfile, and install the same AMD Triton wheels as sgl-project#34403.
@Jiminator Jiminator closed this Sep 14, 2026
@Jiminator
Jiminator deleted the amd/rocm720-pin-triton-wheel branch September 14, 2026 04:41
@alexnails
alexnails restored the amd/rocm720-pin-triton-wheel branch September 14, 2026 05:33
@hnyls2002 hnyls2002 reopened this Sep 14, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants