Skip to content

Record PR42 page-base-flight research evidence - #149

Closed
FujitsuPolycom wants to merge 9 commits into
codex/recurrent-publication-contractfrom
codex/glm53-pr42-page-base-flight
Closed

Record PR42 page-base-flight research evidence#149
FujitsuPolycom wants to merge 9 commits into
codex/recurrent-publication-contractfrom
codex/glm53-pr42-page-base-flight

Conversation

@FujitsuPolycom

@FujitsuPolycom FujitsuPolycom commented Aug 30, 2026

Copy link
Copy Markdown
Owner

Resulting behavior

This draft records SparkCache PR42 opaque-page base-read-flight behavior as research-only evidence. It provides an isolated GLM-5.3 DFlash7 runtime profile, exact ARM64 image pins, stable codeword oracles, manifest inspection, bounded log parsing, and a verdict format that cannot emit a qualification status.

SparkCache's persistent base-segment host-I/O mechanism is implemented and GPU-free tested. Reconstructed page-delta admission is unsupported for GLM-5.3 by the retained semantic evidence. Cache identity, digest salts, 256-token logical geometry, and persisted schemas are unchanged.

Exact source and artifact contract

  • SparkRing base: 74a6a472e7f11a56029d52b879c9b3665aa60e01
  • SparkCache commit: a1511d26a1fe2b17b24561bc52e376bf7f54b06a
  • SparkCache tree: 4d5b8eb8c5c13793ee7a1e67b2b34bd38fcf4ddb
  • SparkCache source SHA-256: 6651f2823c816fac93779cbca54a8f19c0ed262830953149f3a87d189d1f833b
  • Page-base-flight feature commit: 6f3edede4b9f707ef9e8879359b05b0775ed5fee
  • Page-header accounting fix: 229d7d6158261e9510ab99d7e82d532abb9ade01
  • Singleton-later-cohort fix: a1511d26a1fe2b17b24561bc52e376bf7f54b06a
  • Image tag: sparkring-glm53-sparkcache:dflash7-pr42-page-base-flight-singletonfix-arm64
  • Image ID: sha256:35b58a7bf414059c65b8f74e4e4b17ee6a81b7008e1bffbc9bd298b5e08c739e
  • Build receipt SHA-256: ec51c5b99227fe14709977df026e25e3e60f220b81ae252155d048556e8ea90a
  • Archive SHA-256: d6b29b2a3f8595a70890e1795a50c1d65951280f7a302a3b0e1edc0417e6ab55

The profile retains vLLM compiled commit da4d7be6c97434f6942292ed8abbf4b32dc44355, Python overlay commit 0b67266a0f37d6146a8403fb8482403c62f412d5, B12X commit b1d541f9e71a35f030d45fae437630fff7507c2a, lease-contract SHA-256 8adbdfa3fd4b06b213c3aab45255a0b039f1c9940a4b1fad0efd004d263227c9, and CUDA placement-library SHA-256 d57509052b73853bcc8e3c3f47bb81748d87b9cbd8d908fc20d4c79a09aa400c.

Retained live observations

Condition Mechanical result Semantic result Conclusion
Sixteen concurrent 131,072-token page-delta requests with a shared 98,304-token base Every rank recorded 16 participants, one physical base read, and 15 avoided reads Codeword responses failed The host-I/O mechanism worked, but the workload exceeded the 20 GiB GLM hybrid-cache residency and cannot establish correctness
Two concurrent 131,072-token page-delta requests within resident capacity One admitted restore returned the wrong codeword; recomputation returned the correct codeword Python and SparkCache CUDA placement both exhibited the failure The defect is common to reconstructed page-delta admission and is not isolated to CUDA placement
One 131,072-token flat snapshot stored as 13 macro objects Restore completed in 1.55–1.70 seconds Exact red codeword Full-snapshot restore is the verified operational fallback for this evidence scope

Harness contract

The harness preserves stable single-token codeword oracles and the all-rank held-digest readiness check. Publication records a scheduler-log offset before the base request and submits no private-tail request until four ranks report the held digest. Replay writes all result receipts before returning, including rejected receipts.

A mechanically complete verifier result has kind=research-verdict and status=research-only. Semantic mismatches, missing manifests, missing bounded logs, wrong flight accounting, or incomplete restore timing remain rejected evidence. The harness never starts, stops, or restarts a service.

Validation

  • Focused runtime and harness tests: 11 passed
  • Ruff: passed
  • JSON parsing: passed
  • git diff --check: passed

This draft does not qualify reconstructed page-delta restore, the image, or the GLM-5.3 runtime.

Construct exact-boundary prompts with lane-specific natural-language codeword instructions, qualify normalized oracle values instead of ambiguous raw next-token hashes, and retain raw hashes as diagnostics. Retry scheduler readiness through at most eight recorded sub-threshold requests before publishing private tails. Receipt schema v2 rejects older synthetic evidence. Cache identity and on-disk namespaces are unchanged. Validation: 13 focused tests and Ruff.
Repeat each lane's required codeword immediately before the completion cue so the oracle remains stable after a 32,768-token private tail. Five previously unstable lanes returned their exact codewords at the full 131,072-token geometry. Cache identity and storage formats are unchanged. Validation: 13 focused tests and Ruff.
Use the original stable base instruction while retaining the strengthened lane instructions that passed the five full-length private-tail controls. This keeps the 98,304-token shared base identical to the successful control geometry. Cache identity and storage formats are unchanged. Validation: 13 focused tests and Ruff.
@FujitsuPolycom

Copy link
Copy Markdown
Owner Author

This was generated by AI during triage.

This draft records research on reconstructed page-delta admission for image sha256:35b58a7bf414059c65b8f74e4e4b17ee6a81b7008e1bffbc9bd298b5e08c739e. Its retained semantic evidence rejects that restore path, so the draft is not a candidate for main.

Closing the pull request to remove it from the active review queue. Branch codex/glm53-pr42-page-base-flight is retained with the harness and evidence.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant