Repository navigation
Bundle the DMTCP fork release and accelerate large-memory restores - #64
Merged
Merged
Conversation
kbespalov
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Linux release archives now carry the complete TensorRails DMTCP runtime and fdkeep plugin for amd64 and arm64. warmctl resolves the versioned runtime beside its executable, including custom installation prefixes. The installer retains existing component versions and mtimes, and refuses to overwrite a different build at the same version. macOS remains a registry client. DMTCP packages are native Linux/glibc binaries, not architecture-independent executables.
Both CLI packaging and the runtime Dockerfile fetch the independently released
v5.0.0-warmctl.1packages and verify SHA-256 values pinned in source. The duplicate fdkeep source is removed: the DMTCP fork owns and releases it. The release workflow builds archives, then runs native package/plugin and install/reinstall/preservation tests on amd64 and arm64 before publishing. A separate manual workflow builds the full vLLM runtime from a published CLI release, validates its programs, and publishes to registry.warmctl.io in bounded chunks; it does not promotelatestbefore GPU qualification.For large local snapshots,
upsupplies restart-only huge-page advice and a supported reader count bounded by Go's container-aware CPU budget (up to 16), preserving explicit overrides. Private mmap remains off. Four H200 cache-hot restores reached health in 20.21–23.43 s (median 20.83 s; streaming first token 20.92 s), versus 77.36/82.14 s on the same host and snapshot. Documentation records phases, cache/RAM conditions and the unresolved numerical-canary qualification. These timings exclude pod provisioning and snapshot transfer; ARM64 CPU tests do not qualify ARM64 GPU serving.Validation: Go vet and race tests, Linux lint and profile validation; focused tests for reader limits, environment propagation and bundled runtime resolution. CI checks both native Linux architectures, all four CLI build targets, adapters, installer and release archives. DMTCP's merged PR includes full-byte memory/descriptor checkpoint-restore tests and the 20 H200 runs / 640 short requests.
Dependency: https://github.com/TensorRails/dmtcp/releases/tag/v5.0.0-warmctl.1 (merged TensorRails/dmtcp#2; closed performance issue TensorRails/dmtcp#1).
Prepares warmctl v2.0.0-beta.17. Bake fresh snapshots with the released runtime; old snapshots must retain their original mapped libraries and paths. Publishing the CLI and publishing a full vLLM runtime image are separate steps.