[choiyounggi]

Knowledge flush — 1 insight(s)

Headless autoflush run, 2026-11-06 (claims by run 20261106-134854-81372). Claimed 3 candidates: 1 ingested as a new page, 2 released back to pending as pending duplicates of open PRs.

Candidate	Domain	Disposition
2c78c188c0a60329 — Apple Silicon: measure CPU fp32 / MPS fp32 / MPS fp16 p95 for batch-1 embedding inference; re-check accuracy after fp16	platforms hint → backend/ml	ingested — new page backend-common-ml-inference-device-choice-on-apple-silicon (verified)
b5863e2d758c98fb — Claude Code behind a gateway still sends bootstrap/penguin_mode requests while a claude.ai login remains	platforms	released — pending duplicate of open #328 / #298
fc76dc4dfa08ecde — Docker-gated schema validator: skip vs fail, known-bad document	testing	released — pending duplicate of open #300 / #317

Verified best-practice

2c78c188c0a60329 — confidence: verified. Official docs and PyTorch issue measurements back the method. The speed numbers are one machine's field measurement, and the page labels them that way.

• Claim: on Apple Silicon, for one-sentence-per-call transformer inference, do not choose the device from the "MPS loses at small batch" rule. Measure CPU fp32, MPS fp32 and MPS fp16 p95 on the target machine, and re-check the task metric after switching to fp16.
• Sources checked (WebSearch; this run has no WebFetch, so pages were confirmed through search summaries, not fetched):
  • https://docs.pytorch.org/docs/2.14/notes/mps.html — the MPS backend, is_available() / is_built(), and the macOS/device requirement.
  • MPS vs Metal vs CPU performance comparison pytorch/pytorch#148219 — measured per-op dispatch overhead: 1-element mul 24.5 µs on MPS (24.2 µs dispatch) vs 0.9 µs on CPU. This is why the rule holds for tiny ops.
  • MPS device appears much slower than CPU on M1 Mac Pro pytorch/pytorch#77799 — BERT inference about 2× slower on MPS than CPU in the report. The rule of thumb is real for some shapes, so the page asks you to measure.
  • https://sbert.net/docs/sentence_transformer/usage/efficiency.html — fp16 "can speed up inference on GPUs at a minimal loss of model accuracy". That is "minimal", not "zero", which supports the re-check step.
  • https://sbert.net/examples/sentence_transformer/applications/computing-embeddings/README.html — the model is placed "on the most performant available device, e.g. cuda or mps" when device is unset. This supports the "pin the device" step.
  • Warn when models implicitly load in low precision on CPU huggingface/sentence-transformers#4169 — Transformers v5 loads the checkpoint dtype, so half precision can land on CPU and be "significantly slower than float32". This became an edge row.
  • fix(mcp): sentence-transformers on cpu on Apple silicon; release accelerator memory on cooldown kkollsga/kglite#208 — a third-party server forced CPU because MPS placement grew memory by several GB (bge-m3 ~30 ms MPS vs ~65 ms CPU). This supports the memory row and gives an independent case of MPS winning at batch 1.
  • Field evidence from the source session: M4 Pro, torch 2.15. arctic-embed-l-v2.0-ko p95 was 46.0 / 20.6 / 13.5 ms (CPU fp32 / MPS fp32 / MPS fp16). Qwen3-Embedding-0.6B was 62.6 / 30.7 / 18.7 ms. The fp16 model kept F1 at 0.931.
• Changed from the candidate:
  • Added "pin device and dtype explicitly", because the library auto-selects MPS and Transformers v5 loads the checkpoint dtype.
  • Added warm-up exclusion and version recording.
  • Turned "re-check accuracy" into "re-run the task metric on the eval set".
  • Added a decision table, a small-model edge row, a memory row and a half-precision-on-CPU row.
  • The "600M급" size claim became an edge row stated as a field measurement, not a general size cutoff.
• Not verified externally: the exact model size at which MPS starts to beat CPU at batch 1. The page gives no cutoff and tells the reader to measure.

b5863e2d758c98fb, fc76dc4dfa08ecde: released before verification (see Open-PR check).

Existing-layer check

Pages read: backend-common-llm-self-hosted-model-load-latency, debugging-performance-profile-before-optimizing, backend-common-ml-mape-aligned-point-prediction

• Indexes read: INDEX.md, wiki/backend/index.md (full), wiki/platforms/index.md (full).

• wiki_search hits considered (3 queries: the device-choice/p95 phrasing, "Apple Silicon MPS GPU embedding model inference", and the fp16 accuracy phrasing). Top hits:

  • backend-common-reliability-timeouts-and-retries
  • debugging-concurrency-intermittent-failures
  • databases-selection-vector-search-engine-selection
  • backend-common-ml-mape-aligned-point-prediction
  • debugging-performance-profile-before-optimizing

  No hit covers choosing an inference device or dtype.

• debugging-performance-profile-before-optimizing owns the general method: measure distributions (median and p95), and separate cold runs from warm runs. The new page applies it to one decision (device and dtype) and links to it instead of restating it. Not a merge target: its trigger is "something is slow", not "choosing a device".

• backend-common-llm-self-hosted-model-load-latency covers warm-up for a model server (Ollama), not in-process device choice. Linked, not merged.

• platforms has no ML or accelerator category, and its scope is OS differences that break code. Device choice is a model-serving decision, so it does not fit there.

• Created: new page wiki/backend/common/ml/inference-device-choice-on-apple-silicon.md.

• Related links: the new page links to both pages above. Back-links were added to both pages' related:.

• Conflicts: none.

• wiki/backend/index.md: one row added at the top of the ml table. log.md: one ingest entry appended.

• Not edited: the root INDEX.md backend line and the common summary row in wiki/backend/index.md. Open PR knowledge: gateway alias resolution, Spring env-var names, CDN traffic path, stderr in captures #298 edits INDEX.md, and knowledge: re-runnable multi-artifact release, server decision in status responses, REA env -i run token #329/knowledge: Jackson 3 asString() throws on container nodes; node:test files misread by a Jest-only test floor #315/knowledge: gateway alias resolution, Spring env-var names, CDN traffic path, stderr in captures #298/knowledge: group near-duplicate templated examples before cross-validating a text classifier #334 edit wiki/backend/index.md. The ml table row is enough to route there.

Open-PR check

23 open knowledge/* PRs: #334, #332, #331, #330, #329, #328, #324, #319, #318, #317, #315, #310, #309, #308, #307, #306, #305, #304, #301, #300, #299, #298, #293. This headless run cannot fetch or diff branch content. Overlap was judged from each PR's title and changed-file list (gh pr list --repo choiyounggi/dev-loop --state open --search "head:knowledge/" --json number,title,files), plus #334's report body.

Candidate	Overlapping open PR	Verdict
2c78c188c0a60329 inference device choice on Apple Silicon	None on trigger. #334 adds backend/common/ml/grouped-cross-validation-for-templated-examples.md, which is about model-evaluation splits, a different trigger. It also adds one row to the same ml index table; its position could not be diffed from here. This PR inserts its row at the top of that table. If the two rows land next to each other, the merge conflict is two independent rows, and keeping both resolves it. No open PR touches debugging/performance/ or backend/common/llm/self-hosted-model-load-latency.md. The platforms PRs (#328, #324, #319, #309, #299, #298) cover agent CLIs, shells, tmux and tools, not ML inference	new
b5863e2d758c98fb gateway + leftover claude.ai login	#328 edits platforms/processes/spawning-an-agent-cli-from-a-server.md with ANTHROPIC_BASE_URL/ANTHROPIC_AUTH_TOKEN gateway rows. #298 adds backend/common/llm/gateway-model-alias-resolution.md. Same trigger family, and both PRs are still open	pending duplicate → claim released
fc76dc4dfa08ecde Docker-gated schema validator: skip vs fail	#300 and #317 edit testing/quality/checks-that-cannot-pass.md (and #300 also tests-that-cannot-fail.md). Same theme (a validator that skips or always passes), and both PRs are still open. This run could not locate the row's queue file. Its summary is taken from #334's report, which read it	pending duplicate → claim released

Routing decision

• 2c78c188c0a60329 → backend / common/ml → new page wiki/backend/common/ml/inference-device-choice-on-apple-silicon.md.
  • The candidate's domain hint was platforms. That domain covers OS-level differences that break code across macOS/Linux/Windows, and this lesson is about which device a model runs on, which backend/common/ml owns.
  • The directive is general: any PyTorch transformer served one request at a time on Apple Silicon. Project names and numbers appear only in the field-evidence Sources line.
  • No new category.

Local-layer candidates

none

Queue bookkeeping

• Retired in place (status: retired, retiredBy, retiredAt, outcome), the same way earlier headless runs retired rows: 2c78c188c0a60329.
• Released back to pending with queue-claim.js release: b5863e2d758c98fb, fc76dc4dfa08ecde.
• Reviewer note: b5863e2d758c98fb and fc76dc4dfa08ecde have now been released by seven or more consecutive runs while knowledge: gateway alias resolution, Spring env-var names, CDN traffic path, stderr in captures #298/knowledge: Gradle include filters, HMAC key length, Slack webhook probe, approved-command secrets, .env targets, worktree pre-push #300/knowledge: grep -q SIGPIPE in verify pipes, self-matching gate CHECKs, id-only failure allowlists, done-gates counting their own appends #317/knowledge: child claude hooks/env (Stop per turn, --settings env, gateway proxy), Explore ignores CLAUDE_CODE_SUBAGENT_MODEL #328 stay open. Merging or closing those PRs, or one manual /dev-loop:knowledge-flush (which can diff them), would clear them.

[github-actions]

Wiki agent gate: error — the agent run produced no verdict (failed closed). See the workflow log.