Publish the integrated native-MTP3 cache and checkpoint profile - #236
Merged
Conversation
Package the tested CUDA 13.3 and B12X compute composition with exact source verification, retain the BF16 verifier head, and isolate its persistent-cache namespace. Publish the immutable image reference with quickstart and repeated benchmark evidence. Validation: 318 CPU tests passed, 3 skipped; ARM64 image verification, exact installed-package parity, and anonymous registry access passed.
Preserve per-capture stream admission, sequence shared staging and output copies, and reject images without the pinned compute configuration. Record per-run benchmark variability and bundled vLLM provenance. Validation: 372 CPU tests passed, 21 skipped; ARM64 image content checks and anonymous publication passed. GPU stream/fault validation remains separate from these CPU checks.
The derived mesh image checks the pinned selector preimage, adds the block-entry barrier, and increments the kernel compile revision. Compute receipts and package maps identify the transformed source while retaining transport and native-library identities. Local patch tests and parent/result image-content verification pass; GPU validation is separate.
Resolve capture modes, report payload allocation budgets, and keep managed mesh supervision responsive during Docker status delays. Package source-attested MTP3 checkpoint and lease transforms with offline composition checks and request-level reuse analysis. Preserve published image pins, source verification boundaries, and merged startup memory gates. Record bounded four-GB10 text-only cache-pressure evidence without claiming 40 GiB or multimodal qualification. Validation: 2,748 offline tests passed, 42 skipped; numeric evidence verification and Ruff passed.
Exercise transport-error and receipt redaction with an unmistakable test value that does not match the release credential scanner. Production authentication behavior is unchanged. All nine conversation harness tests pass.
Keep service ownership checks intact while simulating root-owned fixture metadata on unprivileged Linux runners. Validation: all 37 managed-service tests pass under WSL, including both formerly failing monitor cases. No runtime or cache identity changes.
Reject absolute, traversing, non-vLLM, and symlinked override paths before reading package bytes. Read JSON as UTF-8 and write receipts as UTF-8 with LF endings. Cache and compute source-lock identities are unchanged; published images are not rebuilt. Validation: five failing path regressions now pass; 16 compute tests and 353 mesh tests pass with one optional native-bundle skip.
Document incompatible indexer postimage hashes and the need for one attested combined image. Source branch integration does not repin images or cache namespaces. Documentation-only change; no stack activity.
Recommend an optional pre-installation reboot after stopping workloads on hosts that have run large GPU jobs. Explain contiguous-memory requirements and retain the managed startup gate as the readiness check. Documentation-only change; relative links and whitespace verified.
Combine source-attested compute overrides and stream-ordered transport with the repository's supported profiles and host memory gates. Preserve the node-count README layout. Existing published qualification remains scoped to its recorded image; additional profile integration is pending validation.
Retain offline container rendering and payload-memory planning as independent interfaces. Include verified checkpoint reuse transforms, responsive Docker supervision, and source-bound evidence without changing the running cluster.
Retain the separately qualified histogram-publication image and patch tooling alongside the compute and cache-reuse composition. Artifact-specific qualification remains explicit.
Compose merged SparkCache source with attested recurrent-checkpoint payloads, graph-only mesh routing, native transport fixes, and the GLM reasoning contract. Add the published image receipt, source preparation tools, strict runtime verification, recipe, and operator quickstart. Select the matching image, placement digest, bundle and namespace together. Preserve the compute-parent profile and artifact-specific qualification; periodic full capture remains disabled by default. Validation: 2,964 offline tests passed, 62 skipped; lint and 550 documentation links passed. The rebuilt image verifies 5,472 files and matches 5,308 deployed runtime files; registry pull and independent file verification passed. No serving replacement or rebuilt-image soak was performed.
Incorporate main's profile attribution and CI corrections while retaining the cache/checkpoint guide. Validate navigation links without coupling the test to a README table layout. Profile documentation tests pass.
Refresh the reviewed MTP3 compute and public-image source identities after integrating the compute profile. Preserve the descriptor's serving limits and source-verification checks. Offline bridge tests: 49 passed, 23 platform or optional companion skips; Linux companion execution remains a CI gate.
Assert the complete immutable registry reference selected by the descriptor instead of a historical digest prefix. Preserve native-MTP and target-identity assertions. Linux export tests: eight passed and sixteen optional companion skips; lint passed.
FujitsuPolycom
marked this pull request as ready for review
September 6, 2026 21:40
This was referenced Sep 6, 2026
This was referenced Sep 6, 2026
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
1 participant
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.
The public MTP3 profile must select a compatible compute runtime, recurrent checkpoint implementation, transport bundle, and SparkCache source together. This integration publishes that composition and adds a root-linked quickstart and recipe, rather than requiring operators to reconstruct private deployment layers.
Status: implemented; research-only performance profile. Includes the integration histories from #219, #226, and #227 without closing the underlying performance issues automatically.
Published composition
Config image:
sha256:6921a6c163ea40b603e19a0332330efe3dbccbf4dce9f6cbbf6b756c9231835a.48bbd2be4a7b972e56632a2d7b934bac5460f272, including merged Let a launch profile and a parent image be resolved before the build #62/Publish the GB10 runtime base and make DeepSeek runnable from it #63 functionality.The exact checkpoint source payloads and native source archive preserve licenses and file hashes. Git attributes prevent line-ending conversion from invalidating attestations. Build tools reject unknown preimages and native inputs; they do not derive acceptance hashes from arbitrary installed code.
Operator interface
docs/GLM53_MTP3_CACHE_CHECKPOINTS_QUICKSTART.mdlinks the published image, recipe, artifact extraction, and managed rendering/installation workflow. The root README selects this guide for the cache/checkpoint mesh. The compute-parent profile remains separately documented with its own evidence.The profile uses 40 GiB/32 GiB persistent-cache capacity defaults and leaves periodic full captures disabled. The earlier stress evidence used 2 GiB per rank and opt-in periodic captures; it does not qualify the 40 GiB multimodal workload.
Validation
AGENTS.md: 2,964 passed, 62 skipped on Windows; platform, optional native-artifact, and companion-integration skips remain explicit in the test output.No serving replacement or soak of this exact rebuilt image was performed. File equivalence does not relabel an earlier image's hardware tests as qualification of this image. The running cluster was left unchanged.
Related: SparkCache #60, #61. This PR does not claim those original scenarios are universally resolved.