Build GLM adaptive MTP from the public native runtime - #130
Closed
FujitsuPolycom wants to merge 7 commits into
Closed
Build GLM adaptive MTP from the public native runtime#130FujitsuPolycom wants to merge 7 commits into
FujitsuPolycom wants to merge 7 commits into
Conversation
Add a byte-attested ARM64 image builder that retains the qualified da4d7be vLLM native extensions, overlays the 31 Python files from 0b67266, installs B12X b1d541f, and composes SparkCache tail publication with its exact patch and lease contracts. The output records mixed vLLM provenance explicitly, preserves da4 wheel metadata, isolates compilation and cache namespaces, and supplies a tail-cow-v1 TP4 adaptive-MTP/fastsafetensors profile with an offline four-rank launcher plan. Compatibility: no existing profile or runtime builder changes. The composed MTP and SparkCache namespaces are distinct. The image remains unsupported for serving until four-rank live qualification. Validation: 14 focused tests passed; Ruff passed; Python bytecode compilation, JSON parsing, shell syntax, and git diff checks passed.
Create a nonexistent context path beneath the process-owned temporary workspace so the source preparer can publish its verified tree atomically. Cleanup removes the enclosing workspace. Validation: 15 focused tests passed; Ruff and shell syntax passed.
Print the measured SparkCache source, native library, and vLLM configuration digests before enforcing their exact build contracts. This makes a rejected composed image identify the first mismatched input without weakening verification. Validation: 15 focused tests passed and the diff check passed.
Measure the exact SparkCache source tree before copying the compiled placement library into its runtime directory. The native library retains an independent SHA-256 check. Validation: 15 focused tests passed; Ruff and diff checks passed.
Require the complete installed Torch version 2.13.0+cu130, including its CUDA-local version suffix. The composed-image verifier therefore matches the immutable public base without weakening version comparison. Validation: 15 focused tests passed and the diff check passed.
Write the verified deployable SparkCache source digest to a runtime receipt before the placement library enters the source directory. Image verification and the TP4 profile consume that receipt instead of hashing a tree that contains generated native artifacts. Compatibility: image labels, cache identities, patch order, and vLLM source contracts are unchanged. Construction and profile resolution fail closed when the clean-source receipt is missing or differs from the pinned digest. Validation: 16 focused tests passed; Ruff, JSON parsing, shell syntax, and git diff checks passed.
Pin both GLM adaptive-MTP construction paths and their executable profiles to SparkCache commit 5d571018, Git tree e864ed9a, and clean deployable-source digest f7c05655. The source provides tail copy-on-write publication and routes reconstructed opaque pages through the native placement library. The four vLLM patches and eleven-file lease contract remain byte-identical. The public-base overlay now labels and verifies the SparkCache Git tree in addition to its commit and clean source digest. Existing cache wire values, digest salts, and logical geometry are unchanged. Validation: 26 relevant GPU-free tests passed; Ruff, JSON parsing, Bash syntax, git diff checks, exact-source digest reproduction, patch/contract comparison, and changed-prose checks passed.
This was referenced Aug 30, 2026
Owner
Author
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.
Result
Adds an executable ARM64 image builder that retains the qualified vLLM da4d7be native extensions and wheel metadata, overlays the exact 31-file Python delta from 0b67266, installs the matching B12X b1d541f source, and composes SparkCache with its SM121 placement library and exact four-patch/eleven-file runtime contract.
The image labels identify retained native code and overlaid Python separately. It is not described as a source-built 0b67266 wheel.
Runtime profile
Adds a test-only TP4/DCP1 profile for adaptive embedded MTP5, fastsafetensors queue size one, native SparkCache restore, and opaque page-tail copy-on-write publication through spark_cache_publication_schema=tail-cow-v1. The profile uses distinct served-model, container, JIT cache, SparkCache root, composed MTP identity, and clear-token values.
The resolver requires the public base digest, mixed vLLM labels, B12X commit and tree, Python overlay manifest, retained native ELF and dispatch manifests, SparkCache source, lease contract, placement-library hash, and source receipt. Its generated four-rank launcher plan is offline.
Compatibility
Existing profiles and runtime builders are unchanged. The composed cache namespace is distinct. Construction fails closed on any source, dependency, patch, native artifact, metadata, or label mismatch.
Validation
Status
Construction and offline verification are implemented. Serving is unsupported until an immutable output digest passes four-rank TP4/DCP1 model loading, semantic generation, SparkCache store/restart/restore, failure-recovery, tail-publication, and concurrency checks.