Skip to content

[Allocator 3/5 — after #71] Qualify physical engine lifecycles with real inference - #74

Draft
deehw wants to merge 8 commits into
codex/dynamic-resource-allocator-v2from
codex/allocator-runtime-qualification
Draft

[Allocator 3/5 — after #71] Qualify physical engine lifecycles with real inference#74
deehw wants to merge 8 commits into
codex/dynamic-resource-allocator-v2from
codex/allocator-runtime-qualification

Conversation

@deehw

@deehw deehw commented Sep 3, 2026

Copy link
Copy Markdown
Contributor

Stacked follow-up to #71. Adds an operator command and durable owner-only report for physical Ollama, ComfyUI, and vLLM lifecycle qualification: immutable inventory/fetch, warm, ownership, native readiness, real inference/workflow output, activity probe, drain/stop, and scoped cleanup. Tightens Ollama overwrite safety and native errors, and fixes vLLM child reaping. Machine A physically passed Ollama with a disposable SmolLM2 135M canary; the canary was removed. The same harness correctly caught the existing gpt-oss:20b tensor size-overflow failure without deleting it. 318 focused tests pass. Base will be retargeted to main after #71 merges.

@deehw
deehw marked this pull request as draft September 3, 2026 00:57
@deehw deehw changed the title Allocator: qualify physical engine lifecycles with real inference [Allocator 3/5 — after #71] Qualify physical engine lifecycles with real inference Sep 3, 2026
@deehw

deehw commented Sep 3, 2026

Copy link
Copy Markdown
Contributor Author

Qualification update: physical vLLM lifecycle passed on D with immutable Qwen2.5-Coder-0.5B-Instruct: fetch, bounded launch, ownership, readiness, real OpenAI-compatible inference returning GRID, activity probe, stop, and artifact cleanup. Physical testing found and fixed GPU-memory bounding, context/eager controls, venv/CUDA toolchain activation, and an explicit FlashInfer fallback. Ollama was also physically qualified on A. Remaining before Ready: physical ComfyUI qualification; no ComfyUI service is currently available on A/C/D.

@deehw

deehw commented Sep 3, 2026

Copy link
Copy Markdown
Contributor Author

Physical ComfyUI qualification completed on Forge machine A (MacBook Pro, MPS) with the installed Z-Image bundle.

Result: PASS. The CLI proved native inventory and artifact identity, warm/load, ownership, readiness, a real 256x256 one-step image workflow, exact output collection, an idle activity probe, and native /free memory release. Real inference took 63.89s and returned 156,361 encoded output units. Report: ~/.grid/allocator-qualification/reports/20260903-040255-3f82faed8470d770.json.

The first physical run found two real integration defects (root WebSocket URL and selection of an unrelated empty legacy output directory). Both are fixed in b60db20, with focused regressions passing (2/2) and qualification/orchestrator suites passing (13/13). The second full physical run passed end to end.

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