Repository navigation
Build the agent-aware gfx1100 compute runtime and qualification ladder #150
Description
Direction
An agent-aware operating environment for local AI compute. Target gfx1100 rather than a W7900 SKU. W7900-only operation remains a normal configuration; a planned RX 7900 XTX is optional until installed and independently qualified. No unified 72-GB allocation pool, implicit sharding or automatic fleet cutover.
Ownership
Aletheia owns workload/session intent and privacy. Logismos owns inference admission, device-local placement, residency and execution. Arche/Tropos and systemd retain host modes, display/gaming and external process lifecycle fences. Kanon retains development dispatch. The pure planner must not initialize devices, reserve real memory or start/stop services.
Program acceptance
- Stable device identities separate from ordinals; one typed architecture contract; checked per-device budgets and deterministic transactional refusal.
- Fail-closed GPU-denied agent iteration; explicit CPU-only versus required HIP compilation; real compiler/code-object evidence (The five .hip kernel sources are compiled by nothing in the fleet, so a change to them is not merely untested — it is never syntax-checked #95).
- Original bounded gfx1100 emulator with exact supported-domain coverage and explicit unsupported-state errors. CPU references and original independent oracles; no timing or performance claims.
- Exact artifact inspection/digests, mixed tensor formats, templates/tokenizers and architecture-derived execution profiles. No filename-based compatibility or parameter-count memory formula.
- Native Qwen3.8 hybrid GDN/attention text plus Qwen3 embedding/reranking continuity. Local quality/cost evaluation selects catalogue challengers; embedding-space changes require reindex and rollback.
- Bounded scheduler/admission, cancellation, residency lifecycle and ordinary/agent-aware APIs; stale grants cannot authorize execution.
- Original HSA/ROCr provider experiment retaining amdgpu, followed by separately approved qualification. No PCI takeover, firmware/reset or power-policy mutation.
- W7900 service qualification first, then independent second-device services after installation. Exact artifacts/toolchain/device/topology and numerical/quality/latency/memory evidence; operator-reserved windows before any incumbent displacement.
- Reconcile authored and derived consumer records (Reconcile repo-facing architecture and status docs with the live runtime #144); upstream repositories are pinned Read-only reference corpora, never vendored implementations or near-ports.
The foundation PR is an increment toward this program, not completion of native serving, fleet integration, general instruction emulation or the below-HIP provider. Existing #46 owns the GDN kernel work; #47 owns deferred TurboQuant work. Training remains outside the runtime scope (#131).
Activity
feat: establish agent-aware gfx1100 compute foundations
forkwright commentedon Sep 5, 2026
The foundation is merged in #151 (a0f4250), with the bounded CI permission/doc follow-up in #153 (041dee5). It establishes gfx1100/stable device identity and ownership checks; strict supplied-budget placement and an action-free CLI; coherent bounded GGUF inspection; original one-head CPU GDN and bounded wave32/WMMA witnesses; and GPU-denied CPU/real-HIP tooling. Hosted ROCm 7.1.1 proved five gfx1100 objects and five source-local syntax rejections. #95 and #144 were closed only after clause-by-clause verification against merged code; this umbrella stays open.
The current work-package authority is the Kanon update https://github.com/forkwright/kanon/pull/3883. Its implementation plan and STATE distinguish the foundation from artifact-derived requirements, deterministic admission/revocation/residency, native hybrid-GDN/retrieval execution, agent integration, the experimental original HSA/ROCr provider, and operator-reserved hardware qualification. Exact artifact provenance/metadata and local evals still determine model choice; the Qwen continuity baselines are not a globally-best claim. W7900-alone remains the baseline; XTX capacity, P2P and deployment are not assumed. No inference workload or incumbent mode/service routing was changed.
The Ardent source update ardent-tools/ardent-tools-site#181 and profile README forkwright/forkwright#9 are merged. Site publishing remains separately blocked by credential authorization (ardent-tools/ardent-tools-site#183); its unpatched browser-installer dependency is #182 there. Release-generator lockfile drift remains #135; the shared check-rollup propagation defect is forkwright/.github#71. Those findings do not change the inference/hardware authority boundary.
forkwright commentedon Sep 5, 2026
Cross-repository landing update: Logismos foundations remain merged in #151 and #153; Ardent site source PR ardent-tools/ardent-tools-site#181 and Forkwright profile PR forkwright/forkwright#9 are merged.
The canonical Kanon planning update forkwright/kanon#3883 is still OPEN, not landed: its required standards job hit the 45-minute cap twice on the same e434233f tree without a lint verdict. All other hosted checks and the local full gate passed. The exact blocker and needed checker/CI investigation are in forkwright/kanon#3884; no gate bypass or timeout relaxation was used.
Ardent production publication independently remains blocked on the existing Cloudflare authentication failure in ardent-tools/ardent-tools-site#183. Native serving, artifact-derived profiles and local model evaluations, runtime admission/residency, the HSA provider and hardware qualification remain program work. This umbrella stays open; no live modes, provider routing, incumbent services or GPU workloads were changed by this delivery increment.
forkwright commentedon Sep 6, 2026
Continuation scope after the foundation landing: implement an opaque loader-owned artifact observation (JSON remains a report, not admission authority); a one-file read-only input lane under the GPU-denied boundary; original grouped multi-head CPU GDN; and a deterministic sched controller backed by placement-owned resource accounting. Runtime tokens must be instance/revision/grant-bound, caller requests cannot inflate granted capacity, revocation immediately prevents new work, and in-flight/failed-eviction allocations remain accounted until explicit reclamation. Preserve v1 planning wire behavior, use independent CPU/property witnesses, and keep the CLI's minimal dependency graph HIP-free. These are implementation increments, not a native model-serving or hardware-qualification claim. Live modes/services/GPU workloads and the separate Kanon standards-timeout issue remain untouched.
forkwright commentedon Sep 6, 2026
The next coherent implementation is now in PR #154: #154. It adds placement-owned process-local reservation accounting, bounded admission/residency/revocation transitions, opaque retained-file observation and typed metadata/tensor diagnostics, the reviewed one-file read-only test input boundary, and original grouped GDN/causal-convolution CPU references. Combined local denied CPU checks and independent authority/numerical review passed; hosted CI is running. This does not close #150 or establish native serving, physical reservations, artifact-derived memory requirements, or GPU qualification. Exact current-model inspection failed closed at unsupported GGML type 23; the next bounded change will validate that format layout before attempting observation again.
feat: add agent-aware admission and artifact observations (#154)
feat: inspect exact IQ4 artifacts and extend bounded emulation (#155)
forkwright commentedon Sep 6, 2026
The next bounded native-data increment is merged in PR #156, with every hosted check passing at 52ee19b. It adds original CPU Q8_0 block decoding with one shared layout authority, and an opaque-observation-bound Qwen35 structural profile whose checked counts and exact role/shape map derive from the same templates. The final optimized HIP-free witness re-observed the exact text artifact under the one-RO-input denied runner: pinned SHA-256 match, 866 roles/shapes, 64 main + 1 NextN block. This is not payload/model execution, provenance admission, VRAM accounting or physical qualification. Native tensor access must next bind to the artifact actually read with an explicit mutation policy; mixed-text formats, operators, tokenizer/template behavior, retrieval conversion provenance, execution requirements and serving remain open. #150 and #46 stay open. Kanon PR #3883 now carries the reconciled state and research/15 exact-artifact evidence.
forkwright commentedon Sep 6, 2026
Completion audit correction: this umbrella was closed at the #156 merge (2026-09-06 02:45:15 UTC), although its program acceptance is not complete. On main f6a21d7, crates/decoders/src/qwen35.rs:117 constructs structural preflight only; crates/loader/src/gguf.rs:461 documents non-atomic observation rather than execution/admission authority; crates/sched/src/lib.rs:1 explicitly limits the controller to process-local state, not hardware reservations. Native hybrid text/retrieval execution, artifact-derived execution admission, model-quality evaluation, the HSA experiment, and physical qualification remain outstanding. Reopening and retaining the original acceptance list. The negated closing-keyword phrase in #156 was replaced with unambiguous outstanding-work wording; the merged implementation is unchanged.
docs: point planning guidance at private corpus (#157)
forkwright commentedon Sep 6, 2026
Deep-pause reconciliation is durable: the owned planning changes now live on the canonical private planning repository's main (
d40d7f9), including current bounded verdicts for all nine pinned reference corpora, a lean charter/index/acceptance contract, exact artifact evidence limits, and preserved Aletheia/Arche ownership boundaries. Public documentation guidance landed via PR #157 with all checks green; its merge tree matches the reviewed candidate.The completion audit confirmed the earlier bounded increments, but this umbrella remains OPEN: structural GGUF/profile evidence and deterministic reservation bookkeeping do not establish native inference, physical residency, or model quality. Issue #46 now reflects the original CPU GDN implementation and the still-missing GPU/parity qualification. Issue #47 explicitly records the existing copied Turbo3 table and requires original derivation or reviewed API retirement before any extension/qualification; no API was silently removed.
Recommended next implementation remains checked artifact-bound payload access with an explicit mutation policy, then independent CPU model/operator composition and artifact-derived execution requirements connected to admission. Routine agent work remains GPU-denied; W7900-alone stays valid and the optional XTX is not assumed installed. No best-model claim, hardware run, service cutover, or host-mode change is implied.
forkwright commentedon Sep 6, 2026
Native execution increment selected after the deep pause: implement an expected-SHA-256-verified, privately owned GGUF byte backing; borrow checked tensor payloads from that same owner; connect it to an original Q8 CPU row operation and a Qwen-structure-bound named tensor projection. The existing cheap reporting/structural observation path remains, but cannot authorize payload execution.
Decision rationale: retained descriptors plus file-size checks cannot prevent same-length rewrites. One bounded owned backing makes byte lifetime/mutation guarantees enforceable in the types. Its explicit limit covers serialized backing bytes only, not all metadata or runtime allocations. Digest equality binds contents to the supplied expectation, not publisher authenticity. File-backed paging/immutable storage needs its own qualified contract; it is not silently substituted underneath this proof.
This increment is deliberately a real weighted-tensor execution path, not another full-model support claim. Full mixed-format model execution, derived execution requirements/admission, quality evaluation, expanded emulation and hardware qualification remain outstanding. Turbo3 disposition remains #47; the Kanon migration/standards lane remains separate. No physical GPU, real model-forward run, incumbent displacement, or provider/mode change is authorized by these CPU tests.
32 remaining items
forkwright commentedon Sep 9, 2026
Progress: #184 merged as
52dde571de649fdbf0d2baf55b3ebdff23df104a; its tree exactly matches the tested candidate.The real hybrid CPU decoder now consumes the checked Q=1 paged-attention operation directly through borrowed KV rows, and uses that same plan for attention workspace accounting. Whole-call cache/recurrent/position publication remains atomic. An original standalone wave32 HIP counterpart has explicit B8/B16/B32 dense-page/ABI admission and independent native-order f32 versus logical f64 witnesses, including meaningful query-head/GQA/page/prefix/rescale/stride falsifiers and normal-domain refusals.
Exact candidate checks passed: workspace/release-kernel gate (618 workspace tests / 108 release-kernel tests), denied boundary and HIP/code-object evidence, and security. The PR records skips and scope limits. Post-merge checks are running on the same tree.
This is not native whole-model execution, general FA-2, shared-prefix policy, physical residency, serving or device qualification. No local compiler/model/GPU/service/mode action was performed. The next consumer slice is complete f32 native model-layer composition with device weight/state/workspace ownership, preserving the independent CPU-only graph. #150 and the broader #173/#46 acceptances remain open.
feat: compose native gfx1100 full-attention layers (#185)
feat: prepare native recurrent hybrid composition (#186)
feat: execute native gfx1100 hybrid main models (#187)
feat: refuse invalid native arithmetic before publication (#188)
forkwright commentedon Sep 9, 2026
Native numerical publication is now checked in merged #188 (
981ce290cf0d8d48363f336b200a80ed90139e94). One sticky typed status spans the actual main-model path; successful synchronization and a clear status read precede state/logit publication. Raw/checked launch ownership is shared and dead reduction lanes cannot cause false status failures. The merged tree matches the independently reviewed candidate with passing hosted gate, security and denied boundary/HIP compilation. Exact proof and limits. This does not qualify hardware, real-artifact quality, emitted math behavior or safe serving; no GPU/service/mode operation occurred.forkwright commentedon Sep 9, 2026
CI-only follow-through and op-pause closeout.
Shared package-source owner PR74 and opt-in PR75 are merged. Logismos PR193 is merged at
51b965dde7e06170cf02c142540a31867a560a7c, with a tree identical to the fully tested candidate. Canonical proof includes passing branch AND main runs, 824 workspace + 157 release-kernel tests (27/13 skips), CPU-denied checks, 13 gfx1100 compile-only objects/rejections and security/CodeQL. Repair forkwright/.github#73 is closed after independent actual-main acceptance audit, with source-only rollback recorded.PR178 is closed as superseded, not as proof that all issue acceptance is complete. #169/#170/#171/#177 remain open; major release PR152 is separate and untouched. No new S4 implementation, GPU/model payload, local Cargo/compiler, host-service/driver/power or deployment action occurred.
Standard pause shape remains assumed rather than separately reconfirmed. Ledger:
No unresolved source/package CI blocker remains. Remaining acceptance is explicit: taxis/cache release-profile proof, real-checkpoint Stella golden parity, later safe serving/host grants and physical qualification. This closeout is not completion of the full Logismos programme.
fix(runtime)!: enforce checked tensor and numerical boundaries (#193)
fix(kernels): tighten packed prefill proof fixtures
test(kernels): compare packed state bits
feat(kernels): compose checked packed CPU prefill (#194)
feat(decoders): stage native single-sequence recurrent chunks (#195)
forkwright commentedon Sep 10, 2026
Native whole-model B=1 chunk prefill landed in #196 (main4e2848185e51c2b72b9a4c337d76bf8b5e81b7cd), with the same tree as tested16ef024. One exact capacity plan drives transient workspaces, causal multiquery attention, whole-chunk KV/recurrent publication and terminal logits; legacy constructors remain capacity1. Q1 and B1 share one private wave32 arithmetic body. Gate34502411614 passed855 workspace tests and178 release kernel tests; denied boundary/HIP witness34502410513 and security34502411038 passed. 14gfx1100 code objects/14malformed-source rejections and the external launcher symbol were verified. See PR196 for full receipts and skip counts. Native device witnesses remain ignored/unrun; no hardware, artifact, performance, host-grant or serving qualification. Full mixed-sequence model ownership/S4 remains open. Next bounded consumer is Hermeneus prompt chunking via the exact per-use plan, without a new service or scheduler authority.
feat(decoders): publish CPU sequence batches atomically (#198)
forkwright commentedon Sep 10, 2026
Two further source increments are landed: #197 consumes exact per-use native prompt chunks in Hermeneus, and #198 adds the whole-model CPU mixed-sequence transaction/reference owner. The latter preflights every owner, stages all recurrent/KV/logit results, and prepares all cache appends before infallible group publication. It retains distinct sequence contexts and truthful checked live-set accounting; its private scalar arithmetic is not packed/vectorized native execution.
At main
2d324f92330303ba9eeaa83c4b768550b677bbea, the tested and merged trees match; 868 workspace and 178 release-kernel tests passed, and all post-merge gate, denied-boundary/HIP, security and CodeQL checks passed. Exact receipts and skip/proof limits are in #198. Native B>1 arithmetic, physical/device/model-quality qualification and host-authorized serving remain open. No GPU execution, model payload runs, host modes/services, dependency changes or deployment occurred.