Row: BACKEND-ROCM
Problem
Two public/row projections are stale after recent ROCm merges:
Issue #2598 was opened while #1876 was temporarily closed. It is now closed as a duplicate because reopened #1876 owns the exact external-hardware scope.
Scope
Gate
- The device-specific scan must execute successfully on the implementation base.
- No hard-coded ROCm registered-operation total may remain in
docs/FEATURES.md.
- Repository prose, link, record, and full preflight gates must pass or retain only precisely classified inherited/argument-dependent results.
This is a record-only repair. It closes when the corrected projections merge.
Row:
BACKEND-ROCMProblem
Two public/row projections are stale after recent ROCm merges:
.agents/specs/rocm-q8k-cooperative-quantizer.mdstill says the implementation is not upstream. Pull request perf(BACKEND-ROCM): cooperate across Q8_K activation blocks #2472 merged as9f96b74465441ebbee3651f4b316cdb0bf183715on 2026-09-02. GitHub briefly marked QuantizeQ8KK is thread-per-superblock: 95.5us/call against llama.cpp's 1.6us, and 41% of the ROCm decode gap #1876 closed because a negated phrase in the merge message contained an adjacent closing keyword and issue number;localai-org-maint-botreopened it at 2026-09-02T19:27:25Z. Issue QuantizeQ8KK is thread-per-superblock: 95.5us/call against llama.cpp's 1.6us, and 41% of the ROCm decode gap #1876 remains open and owns the pending gfx1200/gfx1201 arm.Row:line, but its body now begins withRow: BACKEND-ROCM.docs/FEATURES.mdstores two hard-coded ROCm operation counts of 52. The wrapped-call-safe, device-specific command documented by docs/ROCM.md's op recount command is line-based, so it misses every multi-line RegisterOp( call -- and the count beside it is stale by nine #2573 currently finds another value, and any stored count will drift again whenrocm_ops.hipchanges.Issue #2598 was opened while #1876 was temporarily closed. It is now closed as a duplicate because reopened #1876 owns the exact external-hardware scope.
Scope
## Now, issue-ownership text, outcome, and## Owedsection with merged perf(BACKEND-ROCM): cooperate across Q8_K activation blocks #2472, open/reopened QuantizeQ8KK is thread-per-superblock: 95.5us/call against llama.cpp's 1.6us, and 41% of the ROCm decode gap #1876, and its canonical row ownership.docs/FEATURES.mdcells. Describe the native surface and link readers todocs/ROCM.mdfor the wrapped-call-safe recount command instead of storing a new count.Gate
docs/FEATURES.md.This is a record-only repair. It closes when the corrected projections merge.