Integrate the build-ready DFlash7 image contract - #140
Closed
FujitsuPolycom wants to merge 3 commits into
Closed
Conversation
Attest deep_ep module ownership by the exact deep_ep 2.0.0+local distribution, uninstall only that distribution in the derived image, and verify the module and distribution metadata are absent. Carry the content-addressed removal receipt through prepared-context hashes, OCI labels, profile attestation, and composed-image verification. DFlash profiles leave Torch thread selection unset, select language-model-only serving, disable unsupported symmetric-memory and FlashInfer all-reduce candidates, and disable all-reduce RMS fusion. PYNCCL, the pinned NCCL library, and B12X attention, MoE, and linear bindings remain unchanged. ModelOpt and FP8 KV warnings remain visible. Compatibility: model, draft, loader, KV, topology, SparkCache identity, cache wire values, digest salts, and 256-token geometry are unchanged. Cache namespace impact: none. Validation: 34 focused GPU-free tests passed; repository-wide Ruff passed; Bash syntax, Python compilation, JSON parsing, and diff checks passed. The maintained suite passed 1,944 tests with 9 skipped and retained one unrelated PR #136 base README assertion failure. No image was built and no service was modified.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Resulting behavior
The GLM-5.3 DFlash7 image contract now binds SparkCache commit
19e2ec8b59c84ef359c2a3290f86962e3ff71d96, Git treed8b417bb4b6d734c4403c0a73e7e42b95abd8343, and deployable source SHA-256bc7cae86732c869ee8b2205d48ac5be6f580ee8b77a3e4ffd4c69dcd4f1bfae5.That SparkCache source accepts canonical CUDA configuration keys and replaces a partial terminal HMA page when an authenticated cache boundary falls inside that page. Null-block publication remains unsupported.
The image builder also removes the unused
deep_ep==2.0.0+localdistribution under an exact ownership and receipt contract. Both profiles disable unsupported symmetric-memory and FlashInfer all-reduce probes, disable all-reduce RMS fusion, select language-model-only serving, and retain ModelOpt and FP8 KV warnings that describe supported-runtime limitations. The public benchmark summary states the capacity-limited C4 and C8 scope required by its evidence contract.Compatibility and cache namespace
Cache namespace impact: none. Cache identities, digest salts, chunk geometry, page-delta wire schemas, model identities, vLLM patches, the lease contract, and the CUDA placement ABI are unchanged. Compatible
page-tail-cow-v1entries remain eligible.The image identity and source receipt must change because the SparkCache source and installed Python distributions change. The builder and runtime attestation fail closed unless the PR29 source receipt, canonical CUDA settings, DeepEP removal receipt, and module-absence postcondition all match.
Validation
python -m pytest runtime/glm53-flash-dflash7-python-overlay/test_dflash7_python_overlay.py scripts/test_prepare_glm53_dflash7_python_overlay_profile.py -q— 18 passedpython -m pytest -q— 1,979 passed, 9 skippedgit diff --check— passedNo image was built or published. No serving process or host configuration was changed.