chore: Bump Gym, Megatron-Bridge, vllm for superv3.5 - #3958
Conversation
Point the Megatron Bridge submodule at the tip of chcui/sage/nmvl to test the Sage NMVL integration on the latest super-v3.5-posttraining branch. Signed-off-by: Yi-Fu Wu <yifu.wu@gmail.com>
Signed-off-by: Yi-Fu Wu <yifu.wu@gmail.com>
Signed-off-by: Yi-Fu Wu <yifu.wu@gmail.com>
Signed-off-by: Yi-Fu Wu <yifu.wu@gmail.com>
Signed-off-by: Yi-Fu Wu <yifu.wu@gmail.com>
Cherry of 545ba70 (#3905) from main, which this branch's merge base predates. torch-memory-saver arrives transitively through megatron-core's dev extra as a git dependency, and its build backend refuses to produce metadata without TMS_CUDA_MAJOR: RuntimeError: TMS_CUDA_MAJOR env var must be set for CUDA builds (use `make build-wheel-multi-cuda` or scripts/build_multi_cuda.sh). That aborts `uv lock --check` inside the image build (nemo-ci pipeline 65727784 died there), because an env var exported in a CI step does not reach the Docker build. Declaring it in pyproject.toml makes it travel with the project, so CI, the Docker build, and local uv runs all get it -- which is exactly why #3905 moved it here and dropped the export from cicd-main.yml. 13 matches docker/Dockerfile's BASE_IMAGE (cuda13.2). Verified: `uv lock --check` now resolves 551 packages and exits clean on this branch. The lockfile itself was not stale, so no regeneration is needed. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Yi-Fu Wu <yifu.wu@gmail.com>
Signed-off-by: Yi-Fu Wu <yifu.wu@gmail.com>
✅ Submodule Fast-Forward Check ResultsCheck based on commit: ea67dfd (PR #3958 from ✅ Submodules that are properly updated:Gym: ✅ PR branch is ahead of super-v3.5-posttraining branch (fast-forward) All submodule changes look good! ✨ |
✅ Submodule Fast-Forward Check ResultsCheck based on commit: b10602f (PR #3958 from ✅ Submodules that are properly updated:Gym: ✅ PR branch is ahead of super-v3.5-posttraining branch (fast-forward) All submodule changes look good! ✨ |
yfw
changed the title
✅ Submodule Fast-Forward Check ResultsCheck based on commit: f4dce7c (PR #3958 from ✅ Submodules that are properly updated:Gym: ✅ PR branch is ahead of super-v3.5-posttraining branch (fast-forward) All submodule changes look good! ✨ |
1 participant
What does this PR do ?
Bump Gym, Megatron-Bridge, vllm and use dedicated super-v3.5-posttraining branches for Gym and Megatron-Bridge.
Issues
List issues that this PR closes (syntax):
Usage
# Add a code snippet demonstrating how to use thisBefore your PR is "Ready for review"
Pre checks:
Additional Information