Repository navigation
Support for NVCRE init containers - #6375
sjayaram-nv wants to merge 6 commits into
Conversation
|
I think we should invert the call logic. Right now this is a passthrough essentially of nvcre nemo-run flags. When jobs are launched via setup_experiment.py it already knows which commit is and repo location. I'd argue we should have a flag like Makes the whole flow simpler. |
|
/ok to test e52380c |
|
Automatic Claude reviews have been retired. To request a pull-request review, post a comment containing: Add |
sudostock
left a comment
There was a problem hiding this comment.
hf_token in arg parse from previous NVCRE PR has offline mode issue.
Signed-off-by: Smitha Jayaram <sjayaram@nvidia.com> Signed-off-by: Alex Filby <afilby@nvidia.com>
Signed-off-by: Smitha Jayaram <sjayaram@nvidia.com>
for NVCRE init_container in Run repo Signed-off-by: Smitha Jayaram <sjayaram@nvidia.com>
Signed-off-by: Smitha Jayaram <sjayaram@nvidia.com>
Signed-off-by: Smitha Jayaram <sjayaram@nvidia.com>
Default the NVCRE per-job timeout to 12h Signed-off-by: Smitha Jayaram <sjayaram@nvidia.com>
Signed-off-by: Alex Filby <alexfilby@gmail.com>
063422a to
bb0d261
Compare
What does this PR do ?
Support NVCRE init containers for llmb K8s workloads
Changelog
Added --nvcre_init_containers_json, forwarded to NvcreExecutor.init_containers, and
--nvcre_bridge_root (default /opt/Megatron-Bridge) so the generated command and
PYTHONPATH can use a Megatron-Bridge checkout that an init container placed in the
pod instead of the image's copy.