Skip to content

fixes: minor fixes#134

Open
droot wants to merge 6 commits into
gke-labs:fftfrom
droot:pr-132
Open

fixes: minor fixes#134
droot wants to merge 6 commits into
gke-labs:fftfrom
droot:pr-132

Conversation

@droot

@droot droot commented Jun 24, 2026

Copy link
Copy Markdown
Collaborator
  • Add engine.sleep(level=2) in vllm_sampler.py during initialization so vLLM cleanly flushes CUDA graphs and yields GPU VRAM before checkpointing
  • Switch worker templates in k8s/deploy/distributed-fft-timeslice/ to imagePullPolicy: IfNotPresent for instant (< 1s) pod startup during dev/test
  • Update Makefile to resolve image tags via git rev-parse --short HEAD (with fallback to VERSION file stamped in push-vm)
  • Enhance make push-images to automatically run kubectl set image and update Gateway OPEN_RL_WORKER_IMAGE env var
  • Add detailed llmd-snapshot-agent installation and Helm deployment instructions to docs/setup/gke-fft-timeslice.md
  • Add automated worker pod cleanup (kubectl delete pods -l snapshot-agent=true --force) to make test e2e
  • Exclude VERSION file in .gitignore

…nd vLLM GPU yielding for RL time slicing

- Add engine.sleep(level=2) in vllm_sampler.py during initialization so vLLM cleanly flushes CUDA graphs and yields GPU VRAM before checkpointing
- Switch worker templates in k8s/deploy/distributed-fft-timeslice/ to imagePullPolicy: IfNotPresent for instant (< 1s) pod startup during dev/test
- Update Makefile to resolve image tags via git rev-parse --short HEAD (with fallback to VERSION file stamped in push-vm)
- Enhance make push-images to automatically run kubectl set image and update Gateway OPEN_RL_WORKER_IMAGE env var
- Add detailed llmd-snapshot-agent installation and Helm deployment instructions to docs/setup/gke-fft-timeslice.md
- Add automated worker pod cleanup (kubectl delete pods -l snapshot-agent=true --force) to make test e2e
- Exclude VERSION file in .gitignore
@droot droot marked this pull request as ready for review June 24, 2026 18:59
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.

1 participant