Repository navigation
compute: explicit owner-bound BF16 native CPU candidate - #212
48 commits merged into
Conversation
…native-cpu-inference
…native-cpu-inference # Conflicts: # THIRD_PARTY_LICENSES.md
Next native coding slice from the original failureRead-only review of exact head The next implementation candidate is an explicitly negotiated, versioned output constraint, separate from unchanged The already-pinned llama.cpp Required closure before an actual trial:
This is a concrete next slice, not implemented or validated functionality. It does not prove correct reasoning or protected execution on other people's hardware. Current Storage/Android integrations continue; the stale native worktree and its unrelated uncommitted documentation have been left untouched. |
|
Candidate 4e5126a (+ hosted PATH correction b8bd822) continues the original OpenCode invalid-JSON blocker with explicit 171 targeted inert Python tests pass, including a real serialized selected dataset through worker/encoding/decoding with a synthetic native runtime. 13 new Rust tests await hosted execution. Independent source reviews cover Rust/Python/C++ and the Code adapter (76 Node contracts +10 inert Python session tests pass). New source-only normal and ASan/UBSan jobs use the pinned grammar engine without models; the known upstream all-type sanitizer finding remains unsuppressed. Initial source-only run38671269423 failed before compilation/log creation. Original preflight finds /usr/local/bin/cmake but scrubbed build omitted that path; both now use the same fixed hosted PATH. Original logs SHA256 f05e248ac6685079055b8a4db631d43b85ded5792be0d53a2374ce446c0e0177. No native test passed in that attempt. Quality is dispatched explicitly because this long-lived branch conflicts with main; this does not resolve merge conflicts or authorize a merge. No real OpenCode/model trial restarted before native/contract checks. |
|
Native output-constraint checkpoint: normal source-only build on b8bd822 passes all five targets; ABI, loader, BF16/F32 and output-constraint targets also pass under ASan/UBSan. The unchanged upstream all-type target still fails at its previously recorded Q1_0/Q8_0 unaligned load (quants.c:590). Overall native workflow38671551374 therefore remains failed; this is not release readiness or actual model proof. Original artifact SHA256 normal40ecaab5262d043a25b3d30d52a641f8406130103741e00d6857eca925201163, ASan3c948ae71d8c54870fe1a151b95d699556aa0907f9ccd17b5a9d3d18d71b5272, logs21b840e5ec8c39d8e7af2d9b23214b6eb1a5b04f324343fd57386a6ab8d21dde. Rust Quality38671767972 on030c16f0 passed formatting/native Python/audits, then Clippy rejected identical adjacent invalid_request branches. Commit9d044921 combines their conditions without changing admission or disabling the lint. Exact-source Quality38672934132 is now running. Native sources/Python/workflow remain byte-identical to b8. The Code trial now targets9d044921, but no actual inference retry is dispatched before the required source checks. |
|
Exact-source follow-up audited:
Ledger updated at 85b5ae5. The earlier full all-type sanitizer failure remains failed; supported BF16/F32 and output-constraint checks do not erase it. Next narrow source change retains closed task-boundary errors currently lost before supervision or after report validation, without changing execution rights, wire responses, model/prompt/budgets or coding acceptance. No repeat actual trial dispatched. |
|
The closed task-boundary diagnostic is now verified on exact core Code Exactly one actual native follow-up 38681471469 is now running against that Code/core pair with the unchanged explicit Qwen4B profile, native backend, prompt/output/resource/time limits, task and acceptance rules. It is not a success claim or a presumed fix for the six execution failures. Core ledger-only commit 11bffbb records this evidence; the tested execution pin remains 5607f11. |
|
Native follow-up is now c969eb9, tree 9496e44f29efdad19202fe2d2880ccf4e2d556f4. Commit 714ddb6 adds the reviewed cache hint; merge parent 87f8c31 brings in current main's security, TCP and quote paths. PR212 is mergeable again. The 59 native-only paths and 190 main-only paths are preserved; the four shared resolutions retain both Quality gates, both provenance/evidence histories and the existing native execution path. The last actual Code run 38681471469 (Code e3926734 / core 5607f11) completed one read, then recorded six compute_model_memory_headroom refusals. Artifact SHA-256 29f93f53de70c6ce7dbbac261916289010743cd3b4cb3fae6dce6a7493561810. This proves admission refusal, not a file-cache cause or an OOM kill. Candidate scope: bind an owner-authorized, read-only non-following GGUF descriptor before execution; after successful supervisor cleanup/result validation, and only with known insufficient headroom, issue one bounded POSIX_FADV_DONTNEED advice for that unchanged file. Descriptor/path/parent identity, owner, permissions, link count, size and timestamps are rechecked. No model rehash after advice, global reclaim, retry, permission change or telemetry schema change. Advice is best effort; it guarantees neither reclaim nor the next admission, and does not change the original result. Existing admission thresholds, profiles, task limits and cleanup behavior remain unchanged. Eleven new Rust tests cover authorization, file identity/replacement, failure preservation, cleanup gating and unchanged admission. They have not yet executed. Sixty existing inert native Python tests pass after the merge. Root and independent source review found no blocker. Automatic Quality 38685931749 now runs focused native/resource tests before the normal gates; no duplicate manual dispatch and no new actual model trial. Code's pin remains e3926734/5607f117 until exact-source hosted acceptance. |
|
Published source-verification lifecycle candidate 3327115 (tree 98f164b1e5042e76aa6a1aa3f3238674366c26fd). Original actual 38726624377 remains incomplete: one observed completed read; second admitted execution cancelled by the existing memory guard before any retained prompt_encode observation; five later headroom refusals, no edit/test. The original audit is on Code PR11, comment6191437311. The worker now retains only three descriptors from the actual full source-shard hashing pass. After all source/config/dataset/constraint and owner checks, it makes at most one best-effort file-bound cache advisory per unchanged descriptor, then closes the retained descriptors before backend/GGUF preparation. This adds no content reread, cgroup policy, admission credit, output schema or private diagnostic. All original hashes, deadlines, guards, existing post-supervisor hooks and C++/GGUF sources remain unchanged. Metadata binding is not a substitute for content verification or proof of reclaim. Independent source review and 191 targeted inert cases pass with zero skips; the new real-small-file lifecycle sequence also passes independently at root and fails as expected against original 2dd. The ledger distinguishes those contracts from still-unproved runtime improvement. No model ran locally. One combined core Quality gate is now running at exact332: https://github.com/VOLPAROSSA/volparossa/actions/runs/38733078778 . CodeQL38732695543 remains the original current handle. No new model actual has been started. The existing main merge conflicts are unchanged and will not be hidden by these branch-only checks. |
|
Original source gates pass for exact core 3327115, tree 98f164b1e5042e76aa6a1aa3f3238674366c26fd. Quality38733078778 includes 72 native Python cases (the 12 new source-lifecycle cases actually execute), 62 distinct targeted Rust cases, formatting, strict Clippy and the full workspace gates. Original log ZIP SHA-256 dbe843ee87550cf3b38fed5b839e27ef395f684b358b1f0ec625818015b684ac. CodeQL38732695543 passes four analyses on the exact source; ZIP8a1b4f329fb63da1b0be976684537338218488155f18fbce9e12f77f8a408930. Namespace/ignored/optional harness cases remain explicit skips. Code3a5c1f59 passes original source run38733022967 (410PASS/4SKIP). After confirming both remote heads and clean worktrees, with no active inference job, exactly one new actual was dispatched: https://github.com/VOLPAROSSA/volparossa-code/actions/runs/38737284583 . It uses exact Code3a5c1f59, this core332, the same Qwen4B/native backend, task and admission/resource limits. Previous actual38726624377 remains failed. The new trial is pending, not a successful edit/test or confidential peer execution claim. |
|
Original hosted trial 38737284583 is terminal FAILURE on Code 3a5c1f59b282952e378e9525f2cf47da43ee19b2 and core 3327115. The pinned source/build/preflight stages passed. During guest-inference the owned QEMU unit reported oom-kill / signal 9; SSH returned 255 and no guest-result.json was recovered. This does not establish which memory boundary caused the kill, which guest phase had been reached, or whether the new source-cache hint ran. It is not the same observed guest memory-guard failure as the preceding trial. QEMU join, removal of the private VM scratch/overlay, unchanged observed host routes/DNS, runner restoration and build cleanup are confirmed. Guest-internal worker join/private cleanup are not independently evidenced without the missing guest result. No successful edit/test/model-execution or confidential-execution claim. Original artifact 11974713652 contains 10 closed JSON records, SHA256 031678c85a91bc8210c5dff121b8a1037cdb31e9519bc5a19cfdd36ce39d0cf2 (API digest matches); log ZIP SHA256 358264ace79ac02eee6d355a8d095a0245b20d5ed14623a5bad14af48c480d39. Both ZIP CRC checks pass. Exact checkouts, 33 Code source blobs and the core Git archive were independently audited. No duplicate actual is running. Next scoped work is bounded host-cgroup diagnostics in a separate Code candidate, without changing model, task, resource limits, cache settings or privacy. |
|
The next actual OpenCode run on unchanged core 3327115 is now audited: provisioning/native build and two genuine model-driven read actions completed; the third request hit the compute deadline before an edit or test. The VM did not OOM in this run. Full task acceptance and protected peer execution remain false. The new QEMU observations show memory pressure, not a proven single cause of latency. No retry or resource increase is scheduled before the preparation/prefill diagnosis. Original evidence and cleanup scope: VOLPAROSSA/volparossa-code#11 (comment) |
|
New reviewed source candidate adds a connection-owned native model/prefix lifetime while preserving full ordinary conversation/tool inputs, source/GGUF checks, per-request CPU/wall/pressure limits and separate final joined cleanup. Backend3/build4 provenance and a session ABI test target are included. Inert evidence: 164 Python passes; root independently repeated all 11 session cases. Twenty-one new Rust cases and native C++ changes are not yet compiled here. Hosted Quality and normal/ASan source tests must establish that next; no actual-model acceleration or protected peer-execution claim. Paired Code candidate is 97dfa4adb910970128b6f47f2f65c28f6201abca; its actual core pin remains unchanged until the new source is verified. |
|
Follow-up d6d7c4d applies only the eight formatting hunks reported by the original Quality run 38759968511. The five resulting files were compared byte-for-byte with the preserved formatter output; no runtime behavior changed. That original run failed at formatting and skipped compilation/tests. New Rust compilation remains pending. Native run 38759834612 is being audited separately; its overall sanitizer failure is not being presented as a pass. |
|
Native source-contract audit for exact core
The sanitizer failure has the previously recorded source location and error class, but its new 388-byte output is not byte-identical to the historical record. No suppression, retry or dependency change was used. These are source/synthetic contract checks, not a loaded-model test, actual KV-reuse/acceleration proof, completed coding task, confidential peer execution, or full-suite approval. Original retained ZIPs passed CRC and artifact API digest checks: normal |
|
Original Quality38762467619 failed while compiling the first native Rust target: one unresolved supervisor import and three denied redundant qualifications. No new Rust tests ran. Its checkout d58a823 has exactly the candidate48e2 tree691158c4; the original log ZIP was preserved (SHA256 7d9adec6c697125a8e77213ad4631cb02fcb39bdc45c63e11354b0da8f2bfb28, CRC verified). Follow-up00a4d73b corrects only that import and the three compiler-reported qualifications in three files. No behavior, resource limits, model or native C++ inputs changed. Automatic Quality38763979390 and CodeQL38763974149 now test this source; no model trial has been repeated. |
|
Native-session source-gate update (not a model-trial result)
No manual duplicate workflow, new C++ build or model trial was started. Model/pins, full normal tool context, resource/admission limits and cleanup authority remain unchanged. Warm model/prefix reuse and full coding-loop performance remain unproven; the earlier C++ sanitizer boundary remains as recorded in the preceding audit. Final source audit for Explicit limits: seven Rust cases remain ignored (two 1-GiB proofs, hwsim, host-LAN, nft/veth, and two crash-child-only entries). Three separately reported namespace probes return explicit skipped-proof markers because user namespaces are denied; their successful test wrappers are not live datapath proof. The final Python harness also skips one real local CLI-identity compatibility case without |
|
Follow-up to failed actual Code run 38771413590: the cause remains unproven. The supervisor previously hid the original failure behind its post-cleanup wrapper, while correlated worker status:error results lost their closed code. Commit a37f5cd preserves only closed diagnostic facts, without exposing Error::source or granting received errors exit, cleanup, quarantine or retry authority. Deadlines, budgets, model, tools, cancellation and join/cleanup behavior are unchanged. Exact-source run 38779519182 passes 33 targeted Rust tests (including six new), 32 Python collector tests, strict CLI/test Clippy and formatting, with no skips. Original artifact SHA-256: 92c8cfa62486c9ba8b1e3892aaedde955f6052dd51198866bf618558f1719505; original log ZIP: b8860c039284dff39837f77b97ef789646dd4a06c3c985762f7492c9d07131ed. The preceding run 38778420599 failed formatting only; that failure is retained. This is source evidence, not a successful model turn or repaired actual coding loop. Full PR checks and the paired Code source checks must pass before a further actual run; the earlier sanitizer issue remains open. |
|
Current integration checkpoint — 15 October The tested follow-up candidate is core 6d65bb1, not yet this draft branch head or main. It retains the bounded native-session context allocation and adds the exact aligned Q1 scalar-load correction. Original source run 38802248241 passes the normal and full ASan/UBSan suites, including the unchanged upstream all-type tensor checks. Those are source/native-library checks, not completed coding acceptance. The preceding actual Code run 38797816570 completed three model calls and two reads, then rejected the forwarded edit permission. Its private request payload was not retained. A separate pinned-source regression identified the non-Git worktree-relative edit-pattern mismatch; Code PR17 corrects that exact target presentation under the unchanged owner policy. Hosted source run 38805641941 passes 504 unique tests with five explicit source-fixture skips. One new actual trial is running: https://github.com/VOLPAROSSA/volparossa-code/actions/runs/38806073909 on Code cce5a211a0624fdcc05a787bd4ff31422b30ae65 / core 6d65bb1, with the same full 4B/BF16 model, toolset and resource limits. No read-edit-test or confidential peer-execution success is claimed yet. Read-only merge preview against current main e03f0a9 finds only an implementation-ledger conflict, not an executable conflict. After the actual outcome is audited, integrate the exact candidate with current main, preserve both evidence histories, and validate that final combined source before merging this draft. The running trial sources remain unchanged. |
|
Integration checkpoint: the native candidate is now joined to current main in 985774f (parents 6d65bb1 and e917837; tree 887840a5a8ad3fa5d2fae6259969d04d621420fb). This is a normal fast-forward of the existing draft branch, not a merge into main. Only docs/IMPLEMENTATION_STATUS.md overlapped. Independent review verifies that every other changed blob/mode is identical to its respective parent, with no dependency/toolchain delta. The native historical ledger and current main evidence are preserved; the plan now names the immutable guest / retained worker / guest-terminated TLS integration without claiming protected execution. All 98 inert Python source cases pass locally; no local Rust/native/model build was run. The original combined source gates are now running: Quality 38811047918, CodeQL 38811040412, and native normal/ASan-UBSan 38811039580. Their acceptance is still pending. The automatically reported datapath job 38811047947 is explicitly skipped, not runtime proof. Original actual Code trial 38806073909 remains on Code cce5a211 / core 6d65bb1, and Storage 38804192575 remains on fc75a42. Neither was restarted or repinned. The coding milestone, source integration and protected remote work remain separate acceptance obligations. |
1 participant
Latest actual trial: invalid tool-call JSON, no completed coding loop
Run 37683661957/job 113005829633, attempt 1 on Code
45510055bfe4f8ead010e403e701e5c710efe499/ core9e9086ee873f7162f498e8647e6f6f438273e6ce, failed. Build, conversion and provisioning complete. The task ends after 475,006 ms within the unchanged execution deadline. Core records oneresult_observedand twoconversation_native_json_syntaxdiagnostics: repeated validation of one report, not two executions. One request returns one cleanup-confirmed incomplete result (invalid_output/invalid_model_output). Deadline, token-limit and wire-truncation counters are zero.No usable native turn, tool call, approved read/edit/test or command completes. The fixture is unchanged and independent verification is not reached. Core cgroup peak is 1,535,148,032 bytes within 11 GiB, with zero OOM kills and swap; this is not a complete inference-worker peak. Reported runtime/project, guest-unit/private-data, QEMU/scratch and owned CI-host cleanup pass. Normalized route/DNS observations match; raw byte hashes differ, so neither byte-identical host state nor a full firewall audit is claimed. Raw model output is not exported.
Source review confirms the new mandatory verification scope is checked before
result_observed. The syntax rejection is inside the recognized tool-call body. Exact malformed bytes and a unique generation/template cause remain unknown. Native generation is greedy without grammar constraints; this is a source fact, not a proven unique cause. Parser/permissions, model/prompt, resource limits and deadlines are unchanged. No repeat model trial is authorized by this result.Original artifact 11511695654: 9,982 bytes, ZIP SHA-256
93edd2b2b92db78a807a1de11a431afb38ede10443b965c9c4fe523eac75573f. Original job log: 68,937 bytes, SHA-256e0b7f8f2c8073ade7ee168c2ef9a4efddc2f5b710f9686c8621fbdf68eb2cd17. Original archives/logs and all consumed one-shot receipts remain unchanged.Exact-source checks pass: Code source-contracts 37680333983, core Quality 37680238438 and core CodeQL 37680230981. The ordinary topology job was skipped, not runtime proof. Both PRs remain draft.
Retained earlier checkpoints
The wording below records older checkpoints, not current CI or trial status; the latest actual outcome is above.
Current: explicit native verification scope; fresh CI pending
Core candidate
9e9086ee873f7162f498e8647e6f6f438273e6ce, treeecda529c486e4f759f15d107b5e09a7710bcbfa6, contains reviewed correction84cfc466and main55bb4a2a. The only merge conflict was adjacent third-party provenance sections; both remain intact. Native correction bytes are unchanged by that merge.The worker now carries its actual complete initial shard/aggregate measurement as conversion-source provenance and no longer reads those unused 8,044,982,000 source bytes again after generation. The used GGUF remains fully hashed before native loading and after the native handle closes, with the same owner-control/deadline checks. Converter tensor checks, original model identity, PyTorch before/after checks, model, prompt, resources and governor are unchanged. Mandatory internal
verification_scopedistinguishes these checks; core rejects missing, extra, old or different scope. This is an explicit native contract change, not a claim of continuous owner-proof filesystem immutability.Independent read-only review found no scoped blocker. Implementation tests: 149 focused Python and 9 native/7 private-task Rust checks pass. Root independently repeated 131 directly relevant Python checks and, after integrating main, all 16 Rust checks on Rust 1.85. The matching Code source pin is
45510055bfe4f8ead010e403e701e5c710efe499. Fresh exact-source CI is pending. No model has run on this correction.Latest actual native trial: deadline failure, no verified turn
Run 37671030306/job 112962546447, attempt 1 on Code
2b2687fe/ core36d616fb, failed. Build, conversion and provisioning passed. The worker generated 44 tokens internally at 518,788ms and entered verification at 519,120ms, but no verified result reached OpenCode. At 603,331ms the request ended withcompute_deadline/ cleanup-confirmedexecution_budget_exceeded. Last worker state was paused underquiet_hold; all 31 controls were acknowledged. These observations do not establish total pause duration or a unique overrun cause.Zero native turns, tool calls, approved reads/edits/tests or commands completed; the fixture stayed unchanged and independent verification was not reached. Core cgroup peak 9,466,974,208 bytes remains within 11GiB, with zero OOM kill/swap. Reported runtime/project, guest-unit/private-data, QEMU/scratch and CI-host cleanup passed. Normalized route/DNS matched; raw hashes did not, and no full firewall-audit claim follows. No raw model output was exported.
Original artifact 11505764017 ZIP SHA256
935e115a86abf1da83fe0e3d68ebac90d37ecbe399c7dcc79f055ff26dc33e83; original 68,786-byte raw job log SHA256369407954fff82728d2ff41d5d0e229c78726ea8ae487bf2112964138ca492d8. Original failures, logs and consumed once-only receipts remain retained. No rerun or successful coding claim.Historical source and trial record
The entries below are retained contemporaneous checkpoints. Their “current”, “latest”, “pending” and “no trial” wording is historical; the newest exact-source status is above.
Current integrated diagnostic candidate — source CI pending
Candidate
36d616fb6d79a36fe7c8ee95c4ad84bd56b561ae, tree8875adb196f721786efd92fca7d8664f54c46132, mergesbccd7f3ee613f4132a506e32b1f098806d81ec44with maine9c6e7d7054bf0370044e6e748fb936ea388a276. The sole conflict was adjacent implementation-status sections; both are retained. Native compute/worker code is byte-identical to the reviewed diagnostic candidate. Root's Rust 1.85 cargo check and 34 shared-selector plus 6 filter-fixture checks pass.The unchanged parser now distinguishes only Serde's closed syntax/data/eof/io categories. No parser message, field name, position or raw output is exposed. Historical generic codes, report binding, incomplete results, model, prompt and resource limits are unchanged. Root previously repeated 17 conversation, 8 diagnostic and 22 Python fixture checks; strict targeted Clippy and independent review passed. All bccd CodeQL checks passed, but its conflict prevented a Quality run; those results are not substituted for fresh exact-source CI on this merge. Matching Code candidate
2b2687fec5376b6e49f8266dff71cc7935845e99passes 36 native checks with retained exact sources. No new model trial has run.Latest actual native trial: failed on invalid conversation output
Run37654339995/job112905525075, attempt1 on Code
8b5d5b6ad2d764a503cffb508d5c176711c36840with coref4cd8bd7ca310794dfc73cfcd096d6baedcf73bf, completed FAILED. Exact source/runtime/native build and full BF16 model provisioning succeeded. The guest reached the OpenCode task and retainedresult_observedplus twoconversation_native_jsondiagnostics. The provider submitted one request and returned one incomplete result (invalid_output/invalid_model_output), with zero completed turns, tool calls, permissions, edits or tests. No successful coding or full model-turn result is claimed; the underlying output-shape cause remains under source review. Raw model output was not exported.Core cgroup peak was4,555,878,400 bytes within11GiB, with zero observed OOM kill or swap. Reported owned units/processes/private-data/QEMU/scratch cleanup passed. Normalized route/DNS observations match, while raw IPv6-route bytes differ; byte-identical parent snapshots or a full firewall audit are not claimed. Task-memory and graceful-core-stop fields are absent, not passing or zero observations. Original artifact11499080343 ZIP9982B SHA256
c7d3ed9150e8eb8e7d58a65ed46b0179c997a74f2191c621cf0109552278b6dd; original decoded job log SHA256fd62b12408eee3e4bbf6c994a84a9cede4fd088806103525782cf428f1ff6610. Both and the consumed once-only helper/receipts are retained unchanged. No automatic repeat dispatch or main merge. Original model, prompt, resources, deadlines and acceptance gates remain unchanged.Previous source candidate: bounded validation and closed native diagnostics
Source
f4cd8bd7ca310794dfc73cfcd096d6baedcf73bf, tree169aa49f09120b5b443377b70edf60a6bec7ec78, parent773d6377e5263c24a9263f7b97b5c9ae739b3f74. Exact-source Quality37648810363/job112886443122 and CodeQL37648795706 are successful; the routine KVM check is skipped, not runtime evidence. The candidate's actual trial failed as reported above.Both pinned upstream tensor-validation paths now execute deferred work one tensor at a time, with the existing cancellation/progress callback between validation units. Every tensor is still validated on success; cancellation rejects the load. A single tensor remains non-interruptible. The build uses a byte-exact staged source overlay without modifying the retained upstream checkout, verifies the actual compile-database binding, and emits build-manifest v2 with original, effective and patch hashes. Old v1 manifests are rejected. Closed native load-stage diagnostics add no raw model output, paths or exception text. Model, original prompt, precision, thread budget, memory limits, deadlines and owner-control semantics remain unchanged.
Root repeated 125 Python worker/native/provision checks, 21 fixture checks and the Rust diagnostic privacy-canary check successfully. A fresh source-only build passes the ABI, complete bounded validation (including invalid tensors and cancellation/retry), BF16/F32 and unchanged all-type checks. Final build manifest SHA256
884bdbd00ad7cbcf1211b468d2835bf1e4b9edee2081cb6758712b3703ed0769; library SHA256ef8a9c6bd1df86db4118c4846077a2dbc71818cd809d79496adb66d816012618. Independent source/build-binding review found no blocker. Narrow validation harness sanitizer evidence does not replace the retained failed full-upstream sanitizer run.The matching Code consumer is published as
8b5d5b6awith v2 provenance and exact sibling imports; all89 focused tests and its source-contracts CI37650492562/job112892239423 pass. At that earlier source-only checkpoint the trial was undispatched; its subsequent failure is recorded above. This source correction is not a retrospective proof of the historical runtime failure's cause and is not inference or private-peer evidence.Previous actual trial: model loading failed
The one-shot trial37641148934/job112860150266, attempt1 on Code
d48801a6/core773d6377, completed FAILED. Build/conversion/provisioning pass with the same full BF16 model. All20 owner controls are acknowledged and preparation reachesmodel_load/begin, then reportsworker_other. The retained diagnostic does not identify the precise native failure; no model turn, tool call or coding task completed. Six failed runtime requests have confirmed cleanup. Core cgroup peak8,106,295,296 bytes is within11GiB, with zero swap/OOM kill.Reported owned units/processes/private-data/QEMU/scratch cleanup passes. Normalized route/DNS observations match; raw IPv6 route bytes differ in the explicitly excluded counter column, so byte-identical snapshots are not claimed. No raw model output was exported. Original artifact11492792961 ZIP SHA256
b2621c8918f1304848fcc4a7b9e736fab2379cdf19c29de9ab11606e5c9e9cfc; original job log SHA2564e32650af9e373e5e28ec93663777f1fe833feee419a5fe604fcfdfa6b7ec438. All prior failures and consumed helpers/receipts are retained; no repeat dispatch or main merge.Source review separately found unbounded asynchronous tensor-validation launches in the pinned loader despite the two-thread inference limit. The source-only candidate above now implements bounded, complete validation and closed failure diagnostics; its full-model runtime remains unproved. This is not a retrospective proof of the trial's cause, and no resource limit or validation requirement is being relaxed.
Scope
Previous candidate
773d6377e5263c24a9263f7b97b5c9ae739b3f74, tree1c536731a80ff245d6061e30b74e24cbf6220dc0, preserves narrow correction4f063742and merged main47a6573294f74b6ac4def51f3d1f5d98764ede94. Prior10fd and new773d exact-source checks passed. The subsequent actual trial failed during owner-controlled file verification, as recorded below.7fe450e19305b828c199d602c23a8337aaa1f03b; explicit owner manifest hash authorizes loading only inside the existing isolated worker.Latest original trial and diagnostic follow-up
Original Code trial37614132814/job112768290137, attempt1, on Code01c5c591/core7308371b is FAILED at model-provision/native_conversion, process status1. The actual guest source build and parent ready/release handshake completed with a verified library/manifest. Observed build peak is 1,031,565,312 bytes, with zero swap/OOM; these are build counters, not conversion counters. Wheel graph and runtime imports passed. The catch-all converter error did not retain the failed substage or underlying exception. No actual model provisioning, inference or coding task passed.
All reported guest units/process-group/private-data/QEMU/scratch cleanup gates passed, with scoped routes/DNS unchanged. Original ZIP SHA256
d4b32d66e90d4c76568b2fa989cbab5e49bbb736283868a03a81f9ae0303a254; original log SHA2566e561f75a15560a56f3d98c1a49e518c52ad543646ea1463c7f475e5ec2d0ce6. Original failures and once-only dispatch receipts remain preserved.The previous diagnostic candidate added only a versioned closed conversion-stage/category/child-exit diagnostic alongside the unchanged failure marker. Raw exceptions, paths, logs and tensors remain excluded. Root repeated 29 native/conversion Python checks, including inert substage failures, process cancellation/timeout and privacy canaries. Corresponding Code PR11 parses only the exact bounded failure contract. No model, converter source, resources, deadlines or acceptance gates changed. After independent source/helper review and a fresh all-green preflight, one new diagnostic trial37623495917/job112799334236, attempt1, was accepted at 2026-10-07T12:46:57Z on Code61d2c0c5/core77392063. The diagnostic trial completed FAILED at model provisioning. The converter exited with status 0; verification rejected its output (
stage=verify,failure=contract). The closed receipt does not identify the first failed assertion. No model was admitted and no inference/coding task ran. All reported guest/process/private-data/QEMU/scratch cleanup gates passed, with scoped routes/DNS unchanged. Original ZIP SHA25650ad3ba19fe9811e132f2dad7ee3c49407d5fcecc86a65a816d7a0637eab996e; original job112799334236 log SHA25624bcefeb3279b24e011008fecca2037a0808dac2af82587527c1d378759c86a3. Original once-only receipts remain retained; no repeat dispatch.Source-proven verifier correction
Two independent reviews found that the exact pinned Qwen3 converter omits
rope.dimension_count, while its loader defaults that value to the explicit attention key length. The correction verifies that effective value and key/value dimensions against the original head dimension; wrong explicit overrides and missing required parameters still fail. All tensor-value, tokenizer, model-pin and resource checks remain unchanged. Root repeated all 30 native/conversion tests, including wrong/missing metadata, cancellation and closed diagnostics; independent targeted review/tests found no blocker. This incompatibility is proven in source, not uniquely established as the first assertion in the failed trial. The subsequent reviewed trial below proves conversion and provisioning, but not inference.Earlier evidence
Six Rust native binding checks pass again after the equivalent permission-mask spelling fix. Earlier fifteen private-service, five sandbox and backend precision checks pass.
Root repeated 25 native Python contract/process checks plus 13 provision/retry checks: all pass. Earlier 121 existing worker/provision/conversation checks also passed; no actual model was loaded by these checks.
Definitive normal source-only build passes: library SHA-256
4af8d584349d184049c5f6ffea3e5618839c69536676a6bb5c12ef33b03755a7; build manifestd51c06798cae305622cf91089692c323cc31c1af8ff63420b270aaf44830d27c.Owned ABI smoke and separate unchanged-function BF16/F32 tests pass under ASan/UBSan. The full upstream all-type test remains failed on Q1_0/Q8_0 alignment; original failure evidence is retained, not waived or relabeled.
The first candidate Quality run 37233543544 on
d2fe6728remains failed on Clippy'sverbose_bit_mask. The equivalenttrailing_zeros() >= 6expression preserves zero group/other permissions. Skipped later steps are not passing evidence.Original failed decoded job 111528060483 log SHA-256:
80dbe57d6b939c97c41785c4b7473fd2cc218daa197f017b4157b92501399685.Independent read-only binding review found no blocker. Owned code formatting/diff checks pass; two original SentencePiece notices intentionally retain upstream whitespace byte-for-byte.
Exact follow-up Quality run 37235233676 on
f7e2c3b9passes Clippy but fails four complete CLI-tree tests: duplicate default Clap group nameOptionsintroduced by the flattened native options. The other 578 binary tests pass; subsequent namespace/integration steps are skipped. All four CodeQL analyses pass. Original job 111532896805 log SHA-256227fc42e300b3515506ebed4f3abab453222bb70659dadef9ef53772b9e35126. Native model dispatch remains stopped; no retry was sent. The new candidate gives only the flattened native options an explicit, unique group name. Root repeated all nine native tests successfully; all four original full-command-tree failures also pass locally. New regressions cover the complete CLI tree, paired owner options, unchanged explicit execution and unavailable native flags on private-task. The command-tree regression was observed failing before this correction; exact-source CI is still pending.Not proved / why draft
Actual GGUF conversion and provisioning are now demonstrated by trial 37631532611. Model load, inference throughput, EOS completion and a successful native OpenCode coding scenario remain unproved. Original Code run 37228603308 remains failed. Disposable guest fixture wiring is being integrated separately; new source CI must be judged on this exact follow-up, not the failed previous run. No model executed on the development host and no host networking changed.
The corresponding strict Code consumer capability and disposable fixture update is Code PR #11. This PR must not be described as fixing the earlier 4B stall until actual execution proves it.
Previous metadata-correction trial: failed
The one exact-source trial 37631532611/job 112826781182, attempt 1, on Code
8a7034d5/ core10fd16dccompleted FAILED during the OpenCode task. Source build, conversion and model provisioning now pass: the admitted original BF16 GGUF is 8,051,283,488 bytes, SHA256d7934f6fa06f9781bc15540f0e2b83f89b1895e86440a4e97af4f0434311d1bd.The worker reaches
verify_files/begin, then failscompute_control_ack_deadline: capacity requests an I/O-pressure pause (19.12); two controls were issued and only the initial resume was acknowledged. Source review identifies missing execution-thread control checkpoints in the full original-weight hash loop, also present in subsequent native bundle/post-execution verification. The new candidate services controls on the execution thread between unchanged 1 MiB reads during preparation, bundle validation and post-execution verification. All 138 focused worker/native tests pass, independently repeated by root, including real-file/pipe pause, resume, cancel, deadline and tamper checks. No skipped digests, changed pressure thresholds or wider deadlines. A single blocking read can still exceed the original deadline; no new large-model trial is claimed.No native model turn, tool call or coding task completed. Six runtime requests failed with cleanup confirmed; those are OpenCode behavior within this one trial, not repeated workflow dispatches. Core peak is 2,303,844,352 bytes, with zero OOM kill/swap. All reported owned-unit/process/private-data/QEMU/scratch cleanup gates passed; routes/resolv.conf snapshots match and no raw model output was exported. This is conversion/provisioning progress, not inference, private peer execution or full coding proof.
Original artifact 11487417682 ZIP SHA256
78c4732774fc93dae9872df0460e2addf7f3849a6a190f803069fbc7d6a6e9c8; original job log SHA256ef7d5c7bbeeebf497ec9a42881746a782b36f7c7bddc9febef11d7af373cb44f. All earlier failures and consumed one-shot receipts remain preserved. No repeat dispatch or main merge.