b70-optimization-lab

Laguna reproducibility and native-provenance audit

Date: 2026-07-26 America/Toronto

Status: documentation/repro hardening complete; benchmark result unchanged

This audit made no performance claim and ran no score-bearing GPU work. The sealed result remains:

• 102.97143559613157 tok/s under the submitted legacy event convention;
• 101.94172124017027 tok/s under conventional 99-interval accounting;
• 13/13 canonical-q1 token IDs and output-text hashes exact;
• LocalMaxxing cms2ccv2d00lps201rej94pjy, APPROVED.

Gap found

The first promoted repro hashed the sibling vllm_xpu_kernels/libattn_kernels_xe_2.so, but the measured _vllm_fa2_C.abi3.so contains an absolute RUNPATH into:

/home/steve/src/deepseek-v4-xpu-kernels-mwidth-mhc/build/temp

With the historical launch environment, /proc/self/maps proved that the loader selected the library from that external build tree. It happened to be byte-identical to the sibling file, but the preflight did not prove that. External drift or deletion could therefore change or break the runtime while the old preflight still passed.

Importing _xpu_C also loaded four unpinned helper DSOs:

• libgrouped_gemm_xe_default.so;
• libgdn_attn_kernels_xe_2.so;
• libmqa_logits_kernels_xe_2.so;
• libmhc_kernels_xe_2.so.

Finally, editable package metadata supplied xpumem_allocator from a separate DeepSeek kernel tree. Its source and binary were compatible, but neither its origin nor hash was part of the initial promoted contract.

Reconstructed native provenance

Runtime object family	Observed source commit
final router/workspace source and _moe_C	6f9dd3c3a7b1b677a992ca4f431a968408f9c816
_vllm_fa2_C and libattn_kernels_xe_2.so	906190641d708b8028018c5dde653e265c835348
_C, _xpu_C, grouped GEMM, and transitive helper DSOs	4772f727590c51b72add79350b913d098cf67872
xpumem_allocator	18a44f440ca3ac2006d5ba19cd12ccca0a0c9982

The supplemental 906190641 bundle and combined patch are now tracked beside the final vLLM and kernel snapshots. The earlier 4772f727 and 18a44f440 commits are present in the main kernel bundle’s history.

Hardening applied

The standalone repro now:

1. puts the pinned kernel package first in LD_LIBRARY_PATH, overriding the extension’s absolute RUNPATH selection;
2. hashes four Python extension modules, six helper DSOs, and xpumem_allocator;
3. imports every module without allocating an XPU and checks its real origin;
4. inspects /proc/self/maps and requires each helper DSO to map only from the pinned kernel package;
5. records the runtime-verification JSON in every new run and binds its hash into identity.txt;
6. tracks the full observed runtime/package/host identity;
7. tracks a release-only 32-file model manifest and download-at-revision helper;
8. embeds the sealed benchmark, exactness report, server log, identity, service environment, metrics, cleanup, and every idle snapshot;
9. makes sealed-record verification fail if that raw evidence is absent;
10. validates a new run’s exact identity, selectors, environment, runtime maps, cleanup keys, idle phases, token/text equality, topology, treatment, metric math, and a conventional performance floor.

Reproducibility claims

Three evidence classes are intentionally separate:

• portable sealed-evidence audit: recomputes and verifies the historical claim without a GPU;
• artifact-exact lab replay: requires the originating host’s byte-identical venv and native files;
• source-equivalent rebuild: restores all sources and models but produces a new binary environment that must pass the full gate.

The original single monolithic native build invocation was not sealed. Consequently, a clean rebuild is not called bit-identical merely because it uses the same Git commits. The tracked BUILD.md provides the observed toolchain and standard source-equivalent build entry point while preserving that limitation.

General rules for future model work

1. Hashing a sibling DSO is not loader provenance. Record DT_NEEDED, RUNPATH/RPATH, LD_LIBRARY_PATH, module origins, and actual mapped paths.
2. Pin transitive native dependencies, not only the Python extension directly imported by the model.
3. Treat editable package metadata as advisory. Git HEAD, PYTHONPATH, module origin, file hash, and mapped object identity are the runtime facts.
4. Keep sealed-evidence verification portable and fail closed when evidence is missing. “Not mounted” is not PASS.
5. Separate exact artifact replay from source-equivalent rebuild. Compiler identity alone does not promise bit-identical native output.
6. Model manifests should contain immutable release payloads, not cache locks, metadata, or incomplete downloads.
7. A reproduction validator must test both semantics and performance. Correct-but-slow is useful evidence, but it is not reproduction of a speed headline.
8. Record exact cleanup keys and pre/post idle evidence; empty or partial status files must fail.

Pointers

• standalone repro
• build and reproducibility tiers
• source snapshot index
• qualified result packet
• campaign transfer ledger

This site is open source. Improve this page.