Skip to content

vLLM 0.30.0 host receipts on the workstation: install agree, use pending the winner pin - #267

Merged
seathatflowsinourveins merged 2 commits into
mainfrom
claude/vllm-030-receipts-20260925
Sep 25, 2026
Merged

seathatflowsinourveins merged 2 commits into
mainfrom
claude/vllm-030-receipts-20260925

Conversation

@seathatflowsinourveins

@seathatflowsinourveins seathatflowsinourveins commented Sep 25, 2026 •

Copy link
Copy Markdown
Collaborator

Summary

These are host receipts for vLLM 0.30.0, the production embedding runtime on nativestack-5975wx-20260925 since 15:52:14Z (#260). I took over this recording from the session that switched production, whose round-2 receipts were deferred by a model-usage limit.

Receipt What it shows Independent session
…--vllm--install--20260925 tools/vllm-0.30.0:
• imports vllm 0.30.0 with torch 2.13.0+cu130 on the RTX 4090
• all 5,100 files hashed in RECORD match
• uv pip check is clean
• PyPI's manylinux_2_28 x86_64 wheel sha256 ef52ee58… equals the switch receipt's
• the production unit runs this prefix
agree
…--vllm--use--20260925-3 The ten commands of the 0.25.0 receipt (-2), run against the live 0.30.0 server needs_changes: the claim quoted the dry run's cosines and delta instead of its own output, and called the changed command the seventh rather than the eighth
…--vllm--use--20260925-4 Re-record, superseding -3:
• the claim states only bars (4/4 within 0.005; exactly 0/4 for the rotated control), and exact values stay in the retained outputs
• the changed command is correctly called the eighth
• an added unit query backs the switch time and drop-in
needs_changes on step-8 point 4 only (see Residual)

Both reviews ran in separate headless claude -p sessions with Agent and Workflow disallowed. They re-ran the commands read-only; the rotated control gave exactly 0/4 in every re-run.

Residual: the winner pin (step 8, point 4)

  • The foundation semantic-rag winner pin in catalogs/landscape/foundation.json is still 0.25.0, while manifests/stack.json pins 0.30.0 (vLLM 0.30.0 for the production embeddings role on WSL (pin, receipt, docs) #260).
  • The receipts were recorded with --allow-unbound-version and say so. The reviewer held that an agree on a use receipt needs all four points and that the flag "never counts"; an agree now would let -4 qualify automatically once the pin moves, with no review against that pin.
  • Moving the winner pin needs a new sealed semantic-rag verdict. A candidates-only change is not enough: the sealed packet evidence/artifacts/layer-verdicts-20260922/packets/foundation__semantic-rag.json pins vLLM (candidate 7) at 0.25.0, and build_winners/check_winner_fields take the packet pin first. The vLLM qualification owner plans a new wave packet with vLLM at 0.30.0 (receipt vllm-030-switch-20260925, vLLM 0.30.0 for the production embeddings role on WSL (pin, receipt, docs) #260) and a two-family re-verdict after the Codex usage reset (2026-09-30). A fresh use receipt against the new pin follows; -4's commands are read-only and take about 4 seconds.
  • -4 stays in this PR as a preserved attempt: it fixed -3's issues, and its only open point is the pin binding.
  • Effect now: none on status. vllm on WSL2 stays host_verified through -2 at 0.25.0; the dissent count in the matrix rises.

Evidence discipline

  • All receipts are recorded at a published origin/main: 6752ede (install, -3) and 3ff5211 (-4).
  • The -2 receipt at 0.25.0 is unchanged and stays correct for 0.25.0.
  • No service was started, stopped or reconfigured. The commands send read-only HTTP to 127.0.0.1:18231 and read local files; hf cache verify reads the public Hub file list.
  • Lessons recorded: claims quote only the recording's own output, and run-to-run numbers are stated as bars.

Verification (uv Python 3.13 with jsonschema/PyYAML/numpy, like CI)

All of these pass on the head: host_receipts.py validate (76 receipts), component_matrix.py --write (0 flip-rule violations), new_host_grand_list.py --write, evidence_manifest.py --check (5,514 files) and validate.py.

Label: lane:foundation; the paths are evidence/hosts/** plus generated foundation reports. Codex cross-family review was skipped because of the Codex usage limit until 2026-09-30.

🤖 Generated with Claude Code

@seathatflowsinourveins seathatflowsinourveins added the lane:foundation Foundation lane: Claude/Codex setup, hosts, memory, RAG, research, workers label Sep 25, 2026
@chatgpt-codex-connector

Copy link
Copy Markdown

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.

Scout and others added 2 commits September 25, 2026 14:29
…60925, with their independent reviews

install: independent_session agree. use -3: needs_changes (claim quoted the
dry run's numbers; the eighth command). use -4 (supersedes -3): the claim
states bars and matches its own outputs; needs_changes on step-8 point 4 only,
because the semantic-rag winner pin is 0.25.0 until a new two-family
semantic-rag verdict moves it (after the 2026-09-30 Codex reset).

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

lane:foundation Foundation lane: Claude/Codex setup, hosts, memory, RAG, research, workers

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant