Skip to content

docker: fix main image build — install TransformerEngine from PyPI, drop stale wheel path - #1789

Merged
yueming-yuan merged 1 commit into
mainfrom
zhichen/te-wheel-version-gate
Jul 24, 2026
Merged

docker: fix main image build — install TransformerEngine from PyPI, drop stale wheel path#1789
yueming-yuan merged 1 commit into
mainfrom
zhichen/te-wheel-version-gate

Conversation

@Zhichenzzz

@Zhichenzzz Zhichenzzz commented Jul 24, 2026

Copy link
Copy Markdown
Contributor

Problem

build-and-push on main has been failing since the TE 2.17.0 bump (#1781):

AssertionError: Transformer Engine package version mismatch. Found transformer-engine v2.17.0 and transformer-engine-cu13 v2.12.0.

Root cause: the wheels release (miles-wheels / cu130-x86_64) still ships transformer_engine_torch-2.12.0, and the TE install block prefers any local wheel it finds. So the build first installs transformer_engine==2.17.0 + transformer_engine_cu13==2.17.0 from PyPI, then installs the stale 2.12.0 torch wheel — whose dependency pin silently downgrades transformer-engine-cu13 back to 2.12.0. The TE patch step's import transformer_engine then trips the package sanity check and the image build dies.

Change

TE is meant to come from PyPI now (per discussion with @yueming-yuan), so:

  • Drop the prebuilt TE-torch wheel branch: always pip install --no-build-isolation transformer_engine_torch==2.17.0 (the previous fallback path becomes the only path).
  • rm -f /tmp/wheels/transformer_engine*.whl first, so the stale wheels can't be picked up by anything downstream.
  • The wheels release no longer needs to track TE version bumps.

Verification

  • build-and-push on this PR exercises the changed path directly (the source build of transformer_engine_torch adds some build time; that cost only exists until a wheel for the pinned version exists upstream).

@gemini-code-assist

Copy link
Copy Markdown
Contributor

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

The wheels release still ships transformer_engine_torch 2.12.0; installing
it after the 2.17.0 bump (#1781) downgrades transformer-engine-cu13 via its
dependency pin and the TE import sanity check fails the build. TE now comes
from PyPI, so remove the wheel branch (and delete the stale wheel from
/tmp/wheels so nothing else can pick it up).
@Zhichenzzz
Zhichenzzz force-pushed the zhichen/te-wheel-version-gate branch from fed9995 to 98d15d6 Compare July 24, 2026 19:16
@Zhichenzzz Zhichenzzz changed the title docker: version-gate the prebuilt TE-torch wheel (fixes main build) docker: install TE from PyPI, drop the stale prebuilt TE-torch wheel path (fixes main build) Jul 24, 2026
@Zhichenzzz Zhichenzzz changed the title docker: install TE from PyPI, drop the stale prebuilt TE-torch wheel path (fixes main build) docker: fix main image build — install TransformerEngine from PyPI, drop stale wheel path Jul 24, 2026
@yueming-yuan
yueming-yuan merged commit 667ffbc into main Jul 24, 2026
37 checks passed
@yueming-yuan
yueming-yuan deleted the zhichen/te-wheel-version-gate branch July 24, 2026 19:34
Zhichenzzz added a commit that referenced this pull request Jul 24, 2026
Dockerfile changes were only build-tested by the post-merge push
workflow, so breakage landed on main first (#1781 -> #1789 -> this PR).
Single-arch cu13-x86 build, no registry push, cancels superseded runs.
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.

2 participants