[build] chore: bump transformer-engine to release_v2.15

[ko3n1g]

To be merged on 05/11

Claude summary

What

• Bump transformer-engine to track the release_v2.15 branch (currently 3378ef15), replacing the prior SHA pin (f031cf87).
• Regenerate uv.lock against megatron-bridge:latest.

Why

• TE release_v2.15 is the latest release branch and pulls in upstream fixes/perf since the previously pinned SHA.

Lockfile delta

Updated transformer-engine v2.14.0+f031cf87 -> v2.15.0+3378ef15

Usage (unchanged)

uv sync --extra te

Test plan

• L0 CI green
• L1 CI green (label needs-more-tests applied to broaden coverage)

Quarantined tests (this bump)

None — full L0 + L1 matrix passed first try on H100 and GB200.

[ko3n1g]

/ok to test 3792466

[claude]

Light Code Review - Clean dependency PR. Two observations: (1) No version constraint on flash-attn-4 - the fa4 extra specifies flash-attn-4 cu13 with no version pin. Since this is a pre-1.0 beta 4.0.0b11, future uv lock runs could pull in a version with breaking API changes. Consider adding a lower bound. (2) TE override now tracks a branch instead of a SHA - release_v2.15 replaces the previous f031cf87 pin. The lockfile still pins the exact commit 3378ef15, so uv sync --locked is deterministic. On lockfile regen, the resolved commit will float to the branch tip. Neither is a blocker. Suggested test cases: No perf tests impacted.

[ko3n1g]

/ok to test 24fa152

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[claude]

test

[claude]

Clean dependency bump. Only pyproject.toml and uv.lock changed, lockfile delta is limited to the TE package as expected. Two observations: (1) Missing full-test-suite label - the bump-dependency skill says TE bumps are high-blast-radius and should carry full-test-suite to pull L2 into the PR run. L0+L1 green is good, but L2 coverage would be safer for a TE major-minor bump. Consider adding the label and running L2 before merge, or explicitly noting the skip rationale. (2) Branch pin vs SHA pin - the previous TE override used a full SHA (f031cf87); the new one uses a branch name (release_v2.15). The lockfile still pins the resolved commit (3378ef15), so builds from the lockfile are reproducible. However, a future uv lock will silently advance to whatever release_v2.15 points to. This is fine if intentional, just noting the style change. No code changes, no bugs to flag. Suggested test cases: No perf tests impacted.

[claude]

(Please ignore the two comments above — sandbox formatting issues. This is the final review.)

[claude]

Review: Clean dependency bump. Only pyproject.toml and uv.lock changed; lockfile delta is limited to the TE package as expected. Two observations: (1) Missing full-test-suite label — the bump-dependency skill says TE bumps are high-blast-radius and should carry full-test-suite to pull L2 (VL models, checkpoint conversion, heavy quantization) into the PR run. L0+L1 green is good, but L2 coverage would be safer for a TE major-minor bump. Consider adding the label and running L2 before merge, or explicitly noting the skip rationale. (2) Branch pin vs SHA pin — the previous TE override used a full SHA (f031cf87); the new one uses a branch name (release_v2.15). The lockfile still pins the resolved commit (3378ef15), so builds from the lockfile are reproducible, but a future uv lock will silently advance to whatever release_v2.15 points to at that time. Fine if intentional, just noting the style change. No code changes, no bugs to flag. Suggested test cases: No perf tests impacted.