Skip to content

Bind serving receipts to TraceLens image lineage - #76

Open
irvineoy wants to merge 11 commits into
mainfrom
codex/qwen-source-resolution-hardening
Open

Bind serving receipts to TraceLens image lineage#76
irvineoy wants to merge 11 commits into
mainfrom
codex/qwen-source-resolution-hardening

Conversation

@irvineoy

@irvineoy irvineoy commented Aug 8, 2026

Copy link
Copy Markdown
Collaborator

Summary

  • upgrade serving-runtime receipts to schema v2
  • preserve the frozen input image separately from the derived runtime image
  • bind TraceLens base-to-derived lineage, patch identity, source revision, immutable image IDs, benchmark commands, and raw result receipts
  • normalize local image references to immutable raw image IDs
  • protect shared-host benchmark containers and preserve explicit stop ownership

Exact Qwen live qualification

The complete Apex E2E flow was exercised with the unchanged config:

/home/viouyang/Magpie/examples/benchmarks/benchmark_vllm_qwen3_next_80b_fp8.yaml

  • Apex run: e2e-f6021fce3cc048a6ad7fc4de5cfd8faa
  • immutable result root: /data/viouyang/apex/results/e2e_qwen3_next_80b_fp8_acceptance_v19
  • Magpie revision exercised: 210513b31b2f3607920be4000d37fc51f14c5711
  • GPU: MI355X/gfx950, physical UUID GPU-5b46125e1dcf53d3
  • terminal status: no_gain / insufficient_throughput_gain
  • all three isolated candidate episodes were rejected and reverted
  • accepted patches: none; delivery was not attempted

The final replay returned to the exact baseline source/image identity. Initial and final same-source measurements were:

Measurement Throughput TTFT p99 TPOT p99 GSM8K
Initial baseline 2458.173883 445.489428 ms 13.055343 ms 0.954511
Final unchanged-source replay 2531.164661 442.661081 ms 12.729220 ms 0.957544

The observed +2.969% throughput delta is treated as repeat-run variance, not an optimization gain. This establishes a completed no-regression control flow; it does not claim a winner or formal source-rebuild delivery.

Tracing completed with reward disabled. The run retained a 159,520,658-byte raw trace, two 100,000-row targeted shards, and a typed partial TraceLens comparison. Partial semantic coverage was used only as agent context, never as grading evidence.

Integrity and tests

  • 411 contiguous Apex events across 408 transactions
  • 226/226 CAS objects rehashed successfully (1,198,846,855 bytes)
  • 16/16 supervised action receipts exact-bound and successful
  • regenerated reports and replication guides were byte-identical
  • focused Magpie suite: 109 passed
  • remaining available Magpie suite: 209 passed; the excluded collection requires the optional MCP environment
  • Apex-to-Magpie runtime receipt contract probe: passed
  • compileall and git diff --check: passed

The detailed qualification report lives in Apex at docs/validation_qwen3_next_80b_fp8.md.

@irvineoy
irvineoy marked this pull request as ready for review August 8, 2026 15:29
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