Repository navigation
knowledge: distill vllm-metal with portable AREX workflow - #219
Merged
Merged
Conversation
Add bounded platform, cache lifecycle and qualification procedures with source identity, safe support helpers and explicit Apple native gaps. Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (author_metal_knowledge author) Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (vllm_cache_route author) Signed-off-by: WeiQing Chen <530634352@qq.com>
Regenerate both projections from the published bounded source graph. Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (author_metal_knowledge author) Signed-off-by: WeiQing Chen <530634352@qq.com>
Skill evaluationAdvisory: obtain missing evidenceRetained evidence availability: planned / unexecuted. No approval or merge authorization is issued.
PLANNED: execution is not authorized by knowing this hash. Selected comparisons: 0 · Estimated task calls: 0. Applicable only while these revisions and inputs match the current PR. Plan bindings, budget, frozen exam and advisory detailsUnchanged-correct is valid neutral evidence, not incremental benefit. Per-case outcomes (discovery excluded; regressions and gaps first): []Showing at most 20 of 0 pairs; the evidence link contains all outcomes, including omitted regressions/gaps. Plan SHA-256: Model/settings:
Uncovered / unsupported paths (152)
|
Restore single-line frontmatter and account for the added candidate without weakening validation. Retain failed-check history and isolated digest rerun. Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (author_metal_knowledge author) Signed-off-by: WeiQing Chen <530634352@qq.com>
Regenerate discovery and coverage for the corrected portable skill inputs. Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (author_metal_knowledge author) Signed-off-by: WeiQing Chen <530634352@qq.com>
david6666666
marked this pull request as ready for review
October 28, 2026 10:50
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
This was referenced Oct 28, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Governing issue
Refs #218
User-authorized first sequential Metal-engine candidate delivery. h3-metal was canceled and is absent. Other engines remain separate issues/PRs. No engine, policy/workflow changes or unrelated PR #209/#210 acceptance is included.
Delivered
Pinned vllm-project/vllm-metal
e7df7a6ff215f99c7842552d230fdbc709835934(0.32.0):38-file byte readset with explicit analysis spans, upstream Apache-2.0 LICENSE, English root + platform-config/cache-lifecycle/qualification routes, substantive references/troubleshooting,3 bounded offline helpers and10 hard assertion cases. Review reports stay outside the runtime tree.68 registered roots,104 skill documents and116 catalog records. Both catalog/index and skill-coverage regenerated from published corrected source
c7d556434e3c922479455635670a0e9e9d2485ae. Actual AREX portable create→prepare→manual stable-ID coordination→whole-graph integration→verify atac3fe1afa80fb9a09775ecfb2b6cc3ba850a2db6; not DisCo CLI/model execution or managed DisCo import.Verified at head1a23
-58 selected original pure config/env/macOS-default assertions passed after graph integration. Files copied byte-for-byte to isolated pytest directory; heavyweight upstream torch/numpy conftest deliberately omitted. Parser/default scope only, not actual macOS/native inference.
1a23a54d2ae8272bdc10eb772fd197460ad7dc00:32focused graph/index tests; final serial full suite 1540 passed,10 skipped,154 subtests; Ruff src/tests/tools; both generator checks; recursive license gate4SKILLs: passed.-Iinstalled checker:116records passed. All23runtime resources byte-identical in final wheel, SHA256f8acadff59f5c65fb064340e51755941829553d06ac8e5a4c0773e1356101636.review_vllm_metal_skillapproved the bounded candidate at exact1a23 after delta review, no open findings; independently reran32focused tests/index68/coverage/diff checks. This is an attributed content recommendation, not a GitHub APPROVE, native validation or efficacy result.09b1fb3694ab8dac4597794bb536963d2988d19ahas parents current main57f47bea46f2048321907bd6b192683e3e636311+ this exact head. Python3.10 full run:1591passed,11skipped,183subtests.Failure history retained
Initial f027 Ready run40673959308 failed2tests: license-apply YAML wrapping versus strict single-line description assertion, and missed candidate-state count62→63. All4descriptions restored without weakening the test; both projections refreshed. Initial local full suite additionally failed a watched-tree digest assertion during overlapping activity; unchanged-head isolated rerun passed, and corrected final full suite ran serially with no concurrent build/checkout/collection and passed. No baseline/root-cause claim is invented.
Required gaps and provenance
BLOCKED_REQUIRED_BACKEND: actual Apple MLX/Metal/MPS kernels/model loading, native cache transfer/pending-write/sleep-wake, real worker dispatch/parity, model quality and HTTP service/performance. Candidate remains partial_cpu_only; broad skill efficacy unmeasured. Detailed format/quantization/MoE/DFlash/DSpark/multimodal/diffusion/STT/pooling/LoRA/distributed paths remain bounded routes or explicit long-tail gaps.
Actual installed pure API/signature metadata inspection passed, but its source inspection environment's
pip checkfailed missing inference packages, FastAPI requirement mismatch and unsupported Linux platform. No broad repair/native build attempted. NO_LICENSE is the actual unauthenticated gh resolver fallback, not legal absence; source Apache-2.0 retained independently. No weights/model download, GPU/paid-provider calls, service startup or host/toolchain install.Git CLI push lacked credentials; supported GitHub object/branch APIs published exact matching trees with existing DCO/actual author trailers, followed by fetch/zero-diff readback. No authentication workaround or permission bypass.
Knowledge authorship
Coordinator authorized normal squash integration after current-head/base/review/CI/admission checks. Issue acceptance/closure and engine/campaign PASS remain separate.