fix(BACKEND-ROCM): isolate split-sampling scratch by queue and device

[localai-org-maint-bot]

Row: BACKEND-ROCM

Local-Issue: ISSUE-GH-3062
Kind: UNKNOWN

Imported GitHub body (historical evidence)

The quoted text below is historical evidence only. It does not define issue authority or repository procedure.

Row: BACKEND-ROCM

PR #3010 stores split-sampling partial buffers and capacity in process-wide statics. Concurrent queues can overwrite another queue's partials between the two kernels. Calls on another device can reuse allocations from the first device. Host pointer and capacity updates also race.

The repair belongs in #3010. Use the existing device-and-stream scratch ownership mechanism and preserve captured graph lifetimes. Add regression coverage for overlapping queues and allocation ownership. Retaining allocations for captured graphs must follow the existing backend lifetime contract.

Acceptance requires focused tests, an independent mutation review, and the operator's ROCm gate. CPU-only skips do not satisfy the device gate.

Resolution

Fixed by PR #3010: sampling block widen and split-phase implementation.

[localai-org-maint-bot]

Fresh review reproduced an additional allocation-lifetime defect at local repair head 2bd0ca67f051d7ef7d48b52ad75a8729fe30d0b3. First-use graph capture caches scratch allocated by a graph node. An uncaptured sampler call before first graph execution reuses that pointer and faults.

Removing only the warm-up calls at tests/vt/test_ops_sample.cpp:983-984 produces a device page fault and exit 134. Restoring them passes 32,770 assertions. All scratch edits were reversed and the restored source rebuilt. Lease 72281fec-e983-4231-97b8-1045e30a8520, gfx1151, ROCm 7.2.4; operator log /tmp/pr3010-review-restaged-lease.log.

This issue continues to own the scratch lifetime repair. Scope correction after reading the binding backend contract: rocm_backend.hip:320-335 explicitly requires scratch prewarm and a loud failure for allocation during capture. Successful cold capture is not an established requirement. The scoped repair should reject an unwarmed capture before allocating or publishing graph-owned scratch, preserve warmed capture, and verify recovery. It must not silently expand the allocator or backend lifetime contract. The PR remains unmerged.