fix(h3): route video device selection through the backend seam - #135
Merged
Conversation
Replacement for malformed-trailer PR #134 on row/ARCH-ONE-SURFACE-H3-DEVICE. FOLLOWING_AGENTS_PROTOCOL Assisted-by: Codex:gpt-5 [Codex]
Keep the public ABI selector at 0=CPU and 1=CUDA while converting it once to vt::DeviceType. The H3 shared layer now creates one queue through GetBackend(device_type), restoring the device-leakage ratchet without an allowlist or baseline change. Helper row: row/ARCH-ONE-SURFACE-H3-DEVICE (PR #135; replaces #134). FOLLOWING_AGENTS_PROTOCOL Assisted-by: Codex:gpt-5 [Codex]
Exercise the real CUDA load path with a counting backend so a second queue creation or loss of the selected queue device fails the CPU fold gate. Backend registration is restored by scope on builds that already provide CUDA. Helper row: row/ARCH-ONE-SURFACE-H3-DEVICE (PR #135). FOLLOWING_AGENTS_PROTOCOL Assisted-by: Codex:gpt-5 [Codex]
localai-bot
marked this pull request as ready for review
August 8, 2026 00:13
Preserve CLAIM-ARCH-ONE-SURFACE-H3-DEVICE / PR #135 on origin/main while taking every keyed record from main wholesale and reapplying only the scoped H3 evidence. The append-only state is unioned and sorted. FOLLOWING_AGENTS_PROTOCOL Assisted-by: Codex:gpt-5 [Codex]
mudler
added a commit
that referenced
this pull request
Aug 8, 2026
INCIDENT, disclosed in full: the #136 squash was built by merging its branch against main@9f772cad, but `git commit-tree -p origin/main` resolved the ref at squash time - after another session's fetch in the shared repo had advanced origin/main to 57d546f (PR #135, the H3 device-seam repair). The pushed commit bcd9c48 therefore parented on 57d546f while its tree lacked #135's changes, silently reverting them on main: minimax_h3_video.{h,cpp} device-seam fix, 53 fold-test lines, and #135's record rows. This commit restores #135's full delta on top of #136: exact-restore of the six files #136 never touched, hunk-level re-application of the six shared record files (NOW composed into one Surface-coverage row; the #135 state entry re-inserted at its chronological slot). Verified: diff vs 57d546f now shows only #136's own changes; the restored fold test passes on this tree. Hardening (applied to the landing recipe going forward): pin MAIN_SHA=$(git rev-parse origin/main) at worktree creation, use commit-tree -p "$MAIN_SHA" (never the symbolic ref, which other sessions' fetches move), and push with --force-with-lease=refs/heads/main:"$MAIN_SHA"-derived expectations so a moved remote fails the push instead of absorbing a wrong parent. FOLLOWING_AGENTS_PROTOCOL Assisted-by: Claude Code:claude-fable-5 [ClaudeCode]
Collaborator
Author
|
INCIDENT DISCLOSURE + REPAIR: the #136 landing (bcd9c48) accidentally reverted this PR's content on main for ~40 minutes — a shared-repo ref race: the squash's |
mudler
added a commit
that referenced
this pull request
Aug 8, 2026
Validate human W1-W13 row identity before dictionary normalization, and pin the five remaining exact machine fields through independent mutation inventories and consumers. Reconcile origin/main b44ad33 while preserving its Kimi and #135 device-seam work. Row: ENG-RELEASE-BINARIES FOLLOWING_AGENTS_PROTOCOL Assisted-by: Codex:gpt-5 [Codex]
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
2 participants
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.
Clean replacement for #134; the malformed implementation commit was not cherry-picked or rewritten. The exact tree diff was reapplied from origin/main, verified identical before the mechanical PR-reference update, and committed with real separate protocol trailer lines.
Change
device: 0=CPU, 1=CUDA.MiniMaxH3VideoDeviceType.GetBackend(device_type)and use that queue's device.RED -> GREEN evidence
python3 scripts/check-device-leakage.py --report-> DSR 34,kcuda=2, baseline 32, exit 1.4e3d64d5: DSR 32,kcuda=0, baseline unchanged, no allowlist.python3 tests/scripts/test_device_leakage.py: 25/25 pass.sed -n l:FOLLOWING_AGENTS_PROTOCOLandAssisted-by:are real separate lines.Honest pending gate
Per operator direction, the local CPU build was not repeated after the shared-disk incident.
test_minimax_h3_video_fold, H3, C API, and CPU build/tests are NOT claimed locally and must pass GitHub CI before review/merge. No GPU job, download, benchmark, release-row edit, baseline change, or allowlist was made.Draft remains open for the operator's fresh mutation review.