Repository navigation
Specify initial MellowGPU contracts and vendor foundations - #4
Merged
Merged
Conversation
Add metadata-only Intel/AMD/NVIDIA interfaces, checked GFX9 PM4 encoding and a NVIDIA queue lifecycle adapter that retains uncertain native ownership. Fix the independently reproduced A-to-B-to-A stale channel history case by retaining one channel identity per adapter and a monotonic released generation. Publish the original safe runtime metadata helper and sanitized observations from the designated x86_64 15IGL7. Extend host/Intel CI with the real source tests. Validation: 26 host suites; 122 vendor-interface checks; 324 AMD packet checks; 13 NVIDIA lifecycle groups; 7 metadata fixture cases. Focused runs use fatal ASan/UBSan. Transport callbacks remain simulated: no native GPU command, Metal3 registration, WindowServer qualification, kernel loading or boot change.
Add nine formal documents, three ADRs, nineteen requirement groups and a versioned exact dependency/schema index for the actual 15IGL7 GLK target. Keep native GPU, Metal3 and WindowServer qualification explicitly false. Harden private source staging and full Metal declaration recapture against confirmed integrity/provenance counterexamples. Reject contradictory support evidence and Rosetta fixture execution. Escalate NVIDIA transport faults from queue reset state, and reclaim never-issued rollback records only after complete recovery while preserving all visible tokens. Validation: 26 host suites, 29 specification tests, 90 intake tests, 29 Clang inventory groups, 134 vendor metadata checks, 324 AMD packet checks, 14 NVIDIA lifecycle groups, 7 native metadata cases and focused fault/race regressions. No native GPU execution, driver load or boot/security change.
NiSeullent
changed the title
Preserve historical import hashes while recording the three local NVIDIA queue adaptations. Make the four-of-thirteen passive event selection reproducible and clarify closed-admission drain status wording. Reuse the existing sanitizer runtime policy, validate the actual backend summary and use portable symbol inspection. Actual ASan/UBSan findings under zero-exit caller settings now fail; six runner regressions pass. Run queue and copy regressions under fatal sanitizers in Linux CI too. No native GPU, Metal3, WindowServer or hardware support claim is added.
NiSeullent
marked this pull request as ready for review
October 16, 2026 01:28
NiSeullent
changed the title
NiSeullent
changed the title
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.
MellowGPU needs a reproducible native-driver contract and source trail before any physical device can be admitted. This change adds the initial v1 specification, nine architecture/validation documents, three ADRs, versioned ABI/capability/support/requirement schemas, nineteen requirement groups and an exact generated index of twelve pinned reference components. The actual first device is the user-designated 15IGL7: Intel N4120 / GLK8086:3185 on native x86_64 macOS15.8.1/24H32, routed through i915/ANV research. CPU architecture and GPU vendor are separate keys.
The initial implementation adds metadata-only Intel/AMD/NVIDIA interfaces, checked GFX9 packet encoding and a portable NVIDIA queue/lifecycle adapter. It preserves uncertain ownership, distinguishes channel from whole-device quiescence, rejects stale binding history, escalates transport failures when the queue requires reset and reclaims never-issued rollback records only after complete proven recovery. Client-visible token acknowledgements remain intact. The original runtime-metadata helper and sanitized target observations record class/type encodings without enumerating Metal devices or submitting GPU work. Portable models remain separate from native driver authority.
Upstream intake now verifies selected Git objects and descriptor-relative file access, applies patches in private staging, and rejects known restricted inputs before publication. Metal declaration verification regenerates the full Clang census and binds compiler/SDK/command evidence. Profile consistency checks reject false native capabilities, CPU/policy-based qualification and contradictory target evidence. Runtime fixture execution rejects Rosetta translation; CI adds fatal sanitizer runs for AMD and NVIDIA units.
Local validation: 26 portable host suites; 29 specification regressions; 90 intake tests (one optional existing skip); 29 real-Clang inventory groups; 134 metadata-interface checks; 324 AMD packet checks; 14 NVIDIA lifecycle groups; seven original runtime-metadata fixture cases. Focused interface/vendor/metadata runs use fatal ASan/UBSan with instrumentation checked; the complete host suite uses its normal optimized build. The actual translated x86_64 metadata invocation rejects before output creation. All seven workflow runs on commit
5d31a6c4f8f55091be5cfae5e9da3d7f1513b885pass, including native Xcode build and Linux/macOS/Windows host tests. The fresh bounded Claude review found a source-adaptation receipt omission and minor test-evidence issues; those were corrected and verified. Six additional backend-runner regressions reject actual fatal sanitizer findings even with hostile inherited runtime options. CI now covers the actual NVIDIA queue and copy-pushbuffer units under fatal sanitizers. This remains an initial implementation: source/firmware closure is UNKNOWN; qualified GPU profiles are zero. Native GLK DMA/VM/interrupt/submission ownership, complete applicable Metal3 conformance, GPU IOSurface sharing, WindowServer, display, applications/games and performance/recovery qualification remain unfinished. No target driver was loaded, and no boot or security settings were changed during this work. Existing v0.5.1-dev release assets remain immutable.