Validation: independent AMD Windows run on a supported HIP GPU

[AIwork4me]

Objective

Independently validate the public v1.6-windows-capability-2026-07-26 release on a Windows AMD Radeon GPU supported by the locked official llama.cpp HIP binary.

Release: https://github.com/AIwork4me/omnidocbench-amd-windows/releases/tag/v1.6-windows-capability-2026-07-26

This closes the remaining hardware evidence gap. The release already has physical CPU + WSL CDM evidence on Radeon 860M, but gfx1152 cannot execute the tested official Windows HIP binaries.

Required hardware

• Windows 11
• AMD Radeon supported by the locked llama.cpp HIP release, preferably Radeon 8060S or RX 7900 XT-class hardware
• At least 8 GB usable GPU memory recommended
• WSL2 Ubuntu 22.04
• Do not use Radeon 860M/gfx1152 for this acceptance run unless a separately identified gfx1152-compatible llama.cpp build is supplied and locked

Record before execution:

Get-CimInstance Win32_VideoController | Select-Object Name, DriverVersion, PNPDeviceID
Get-CimInstance Win32_ComputerSystem | Select-Object Model, TotalPhysicalMemory
wsl -l -v

Phase 1: independent release smoke

Start in a new directory on a machine other than the development/release machine:

git clone --branch v1.6-windows-capability-2026-07-26 --depth 1 `
  https://github.com/AIwork4me/omnidocbench-amd-windows.git
cd omnidocbench-amd-windows
powershell -ExecutionPolicy Bypass -File scripts\reproduce.ps1 `
  -Profile cpu-smoke-10

The unmodified release must pass before any HIP-specific experiment. -SeedFrom is allowed only when the source and destination lock verification passes. Do not copy predictions, scores, .venv, generated checkouts, .env.local, or run state.

Phase 1 acceptance:

• outputs/reproduction/cpu-smoke-10/state.json reports passed
• Exactly 10 non-empty UTF-8 predictions and 100% manifest coverage
• Windows mandatory metrics finite and non-negative
• WSL CDM finite and positive
• full-verify.ps1: 0 failed
• A second -Resume exits 0

Phase 2: HIP backend proof

Use the release's locked HIP path:

powershell -ExecutionPolicy Bypass -File scripts\preflight.ps1 -CdmPath Wsl -Variant hip
powershell -ExecutionPolicy Bypass -File adapters\paddleocr-vl-1.6\01-vlm-server\setup.ps1 `
  -Variant hip -Port 8122
powershell -ExecutionPolicy Bypass -File adapters\paddleocr-vl-1.6\01-vlm-server\verify.ps1 `
  -Port 8122

At the documented human-intervention point, capture evidence that the server process is using the AMD GPU. Record Task Manager GPU engine/memory or rocm-smi output where available.

Do not overwrite predictions\paddleocrvl_cpu_smoke_10. Use a unique HIP prediction directory and a matching explicit scoring config/result name. Until a tracked hip-smoke-10 profile exists, follow the manual adapter/scoring ownership chain in AGENTS.md and record every exact command.

Phase 2 acceptance:

• Locked HIP archive and server binary pass verify-upstream-lock.ps1
• /v1/models returns HTTP 200 and the expected served model ID
• GPU utilization is observed during inference; CPU fallback is not accepted as HIP proof
• Exactly 10 fresh non-empty predictions, with frozen error count 0 and coverage 100%
• Windows and WSL shared metrics have absolute delta 0 under deterministic single-worker scoring
• WSL CDM is finite and positive
• No timeout, error, or exception cases
• Exact artifact paths and SHA-256 values are recorded

Evidence to attach

• Machine/GPU/driver identity
• Release tag and peeled commit (8fb05fc19d134c139a4730103bd18ce9d1d6db65)
• Commands and exit codes
• Phase durations
• GPU-use evidence
• Prediction count/coverage/error count
• Windows and WSL metric values plus sample denominators
• Manifest, prediction-tree, Windows metric, WSL metric, and final-state SHA-256
• Any UAC, reboot, mirror, or driver intervention

Current blocker on the release machine

The release machine has Radeon 860M/gfx1152 (PCI\VEN_1002&DEV_1114, driver 32.0.22032.14003). Official Windows HIP llama.cpp b9637 and b10107 both reproduced ROCm error: invalid device function; no local HIP SDK/hipcc is installed. This is tracked upstream at ggml-org/llama.cpp#26127, so rerunning that known-failing path would not provide new acceptance evidence.