feat: add agent-aware admission and artifact observations - #154
Merged
Conversation
Add a generic dense causal-convolution reference with explicit raw-input history, checked shape admission, finite arithmetic, and independent CPU oracle coverage.
MetaValue::Array now carries MetaArray instead of Vec<MetaValue>; this intentional source-level API correction preserves declared element types for empty arrays without parallel authority.
This was referenced Sep 6, 2026
forkwright
added a commit
that referenced
this pull request
Sep 6, 2026
Continues #150 after merged #154. - Add inspection-only IQ4_XS (256 elements/136 bytes) and IQ4_NL (32/18) layouts verified against pinned primary GGML sources; preserve explicit decode refusal and reject malformed block rows, offsets and unknown types. - Add a typed unknown-storage diagnostic with numeric-only stderr while preserving byte-exact v1 error stdout. - Add original VGPR-only v_mul_f32_e32 emulation in the existing finite-normal input/result domain, with pinned compiler-byte round-trip and independent numeric oracle. No new FP mode, memory/control state or full-kernel claim. - Reconcile authored scheduler/architecture descriptions with the merged process-local controller. Validation: denied locked workspace tests, strict all-target Clippy, release loader/emulation tests, three pinned AOMP fixture witnesses, focused final loader/bin checks, format and repository guards. Independent format and ISA reviews cleared the changes. Minimal optimized inspector has no HIP/device-runtime dependency. Exact file observations through the reviewed one-file read-only boundary now bind the text Qwen3.8-27B GGUF and Qwen3 embedding Q8 artifact to their immutable publisher LFS hashes. The text file declares qwen35 and seven actual storage types. Reranker bytes have an observed digest, but its conversion/publisher chain remains unverified. These observations do not establish payload decoding, model quality, derived VRAM requirements, native serving or GPU qualification. No model execution or live fleet changes occurred.
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
1 participant
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.
Continues #150 without closing the full program.
Implements one coherent CPU-only increment:
Existing plan and no-flag inspect v1 JSON remain unchanged. MetaValue::Array now retains its declared element type; this intentionally corrects the unpublished loader Rust API and updates every in-repo consumer.
Validation: combined GPU-denied locked workspace tests, strict all-target Clippy and formatting; release placement/sched/kernels tests; repository scope/document guards; synthetic input/device/descriptor/network boundary witnesses. Independent authority and numerical reviews cleared the integrated code. Required hosted checks remain authoritative for landing.
Limits: no native model executor, physical reservation, model-quality claim, GPU execution, service/mode/routing changes, or pooled VRAM. One service owner must own each resource grant and drain/reconcile on restart. Exact local model inspection currently refuses unsupported GGML storage ID 23; format support and artifact-derived profiles are follow-up work, not claimed by this PR.