Skip to content

ledger(H7x): amendment V1, arms B and C on EAT release-1.1.0 (vLLM CPU v0.28.0) - #235

Closed
88plug wants to merge 1 commit into
claude/xhost-h7xfrom
claude/latest-bench-arms-1029-h7x
Closed

88plug wants to merge 1 commit into
claude/xhost-h7xfrom
claude/latest-bench-arms-1029-h7x

Conversation

@88plug

@88plug 88plug commented Oct 29, 2026

Copy link
Copy Markdown
Owner

This branch is stacked on #196 and targets its branch. Merge it into claude/xhost-h7x before #196 lands, so the amendment is in the ledger before H7x's first hold.

H7x has not run. Its six setup smokes ran on vLLM v0.24.0 and stay as records. B is "EAT's vLLM CPU engine with EAT's chart defaults", which makes it EAT stock, and EAT-stock arms track EAT's latest release. That release is release-1.1.0 (0c726f57, 2026-10-22), still the latest on 2026-10-29. Its xeon-values.yaml ships vllm/vllm-openai-cpu:v0.28.0. Amendment V1 moves B and C there, as crescent amendment E1 did for H50 and H51.

Runner changes in bench/apb/h7x.sh:

  • B's tag is pinned in the script itself: v0.28.0@sha256:197354e6…, the tag's index digest, re-resolved on 2026-10-29. It is passed as --set image.tag, so the shared overlay integrations/eat/vllm-values.yaml (still v0.24.0, read by the done row H7) no longer decides H7x's image.
  • The script refuses an EAT chart whose xeon-values.yaml, values.yaml or templates/deployment.yaml does not match release-1.1.0 by sha256. The run directory keeps chart.sha256 and chart-check.txt.

What changes in the arms. B and C were rendered through h7x.sh itself with both releases (H7X_RENDER_ONLY=1, H7X_VLLM_MEM=fit). The objects differ in two places: the image, and the chart's fsGroup: 1001. The 0d3438d render reproduces smoke 6's vllm.yaml and vllm-c.yaml byte for byte. The amendment also records an engine-level change: vLLM v0.28.0 no longer reads VLLM_CPU_SGL_KERNEL (vllm-project/vllm#50801). Its dispatch sends bf16 weights over 1 MiB to oneDNN, so every dense projection of Qwen3-4B moves from the SGL kernel to oneDNN.

The amendment requires a seventh setup smoke (B and C, the run's settings) before the hold. If B is not Ready, the hold waits for a further amendment. No prediction, rule, sizing or cap changes.

Tests run on this branch:

  • tests/bench_test.sh: 64 passed. Two checks are new. One asserts that h7x.sh refuses a non-release chart. The other asserts the digest-pinned v0.28.0 tag. Both were confirmed failing on the pre-change script.
  • bench/apb/test_h7x.py.
  • tests/claims_check.sh and its test.
  • tests/ledger_drift_check.py and its test.
  • shellcheck at warning severity.

The amendment links to crescent amendment E1's heading, which is on main. The link resolves once #196 merges main.

…U v0.28.0)

H7x has not run; only its six setup smokes have, all on v0.24.0, and
they stay as records. B and C are EAT stock, and EAT's latest release
is now release-1.1.0 (0c726f57), whose xeon-values.yaml ships
vllm/vllm-openai-cpu v0.28.0. The amendment moves both arms there, as
crescent amendment E1 did for H50 and H51, and asks for one more setup
smoke on the new image before the hold.

h7x.sh now pins B's tag itself (v0.28.0 by index digest) instead of
reading it from integrations/eat/vllm-values.yaml, and refuses an EAT
chart whose xeon-values.yaml, values.yaml or deployment template does
not match release-1.1.0 by sha256. Rendered through h7x.sh, the two
releases differ only in the image and the chart's fsGroup 1001; the old
render reproduces smoke 6's manifests byte for byte. The amendment also
records that vLLM v0.28.0 no longer reads VLLM_CPU_SGL_KERNEL, so B's
dense bf16 projections run on oneDNN instead of the SGL kernel. No
prediction, rule or cap changes.

Tests: bench_test.sh checks that h7x.sh refuses a non-release chart and
pins v0.28.0 by digest; the early-exit test skips the chart check.
@88plug

88plug commented Oct 29, 2026

Copy link
Copy Markdown
Owner Author

Landed in #241 (merge commit 76bbfbd).

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