[Yasei-no-otoko]

Problem and result

This PR originally provided a manual Windows HIP image-encoding path when that path was missing. Its main feature is now implemented independently on current main. Keep this PR closed and unmerged; no reopening or replacement PR is requested.

On 2026-10-22, inspected main at 011c9e600a93bbc9fd451c563501b5831c32bbca, whose documentation identifies Strata 0.2.2. Main supports a separate HIP picture reader, strata-vision-hip.exe, alongside the CPU reader, and its Windows HIP build script builds and packages both. This source inspection does not establish that the current release works on this RX 6900 XT.

Related discussion

• Tracker reset and reopening requirements.
• The original proposal and review remain in this PR. The author is retaining it as historical evidence because the core capability now exists upstream.

Approach

Current upstream references:

• HIP encoder options, architecture selection and output name.
• Windows build and packaging of the CPU and HIP readers.
• Current AMD image-reader behavior and package documentation.

This does not mean every hunk of the old patch was incorporated. The standalone helper and the explicit warm-up-failure guard differ from main. Source comparison shows that main logs its warm-up token count and then reports READY; the original PR also refused readiness after failed warm-up. This is an unconfirmed static-analysis observation, not a reproduced failure on current main, and no additional fix or review is requested here.

A local trial rebase was inspected but was not pushed. The PR branch remains at historical head b4d759058c14bd92bf57455d9cbe14fb53a3a67b; its old deployment instructions and helper are historical, not instructions for current main.

Validation

Passed:

• 2026-10-22 source/metadata checks only: git fetch upstream main, git rev-parse upstream/main, git show upstream/main:tools/vision/CMakeLists.txt, git show upstream/main:tools/hip/build_windows.bat, and git show upstream/main:docs/AMD_HIP.md confirmed the source integration described above at 011c9e60. Read current CONTRIBUTING.md, AGENTS.md and the PR template.
• Historical GPU/API results, 2026-10-07, pre-merge implementation recorded at 35d8c590ab2f21988c42c6dfe7d9c400b9aefc3a: physical Windows 11 Pro build 26300, Threadripper 3990X, RX 6900 XT 16 GiB (gfx1030), driver 32.0.21045.5002; ROCm 10.2.0a20260930, Clang 24, VS2022/CMake/Ninja, llama.cpp 3cf03257f219afbe7334045ff7c6a06ac68c627d. Official Strata 0.1.40 language engine, Huihui Swift IQ3_XXS, matching BF16 projector, 300-image-token cap and 131072 context. set STRATA_HIP_ARCHS=gfx1030, set LLAMA_DIR to that pinned checkout, then tools\vision\build_windows_hip.bat: HIP build and incremental rebuild passed. CPU and GPU encodes produced finite 240 × 2560 embeddings; masked-GPU --gpu correctly failed without READY. GPU logs identified ROCm0 / AMD Radeon RX 6900 XT; DLL paths and GPU-memory counters supported HIP execution.
• Historical API results from the same pre-merge deployment: distinct uncached image fixtures were recognized through Chat Completions (HIP2026) and Responses (HIP2027), including the red circle and blue square. Health, model list, version, web root and text chat passed. Dated report and structured results, requests and outputs.
• Historical follow-up at bea867945e7f23c5e1b25ab86711d8311f246ffb: the actual Python selection/venv block passed Windows cmd.exe checks with only py, only python, both, and neither available. The first three created runnable environments; the last produced the expected missing-Python error, as recorded in the resolved review discussion.
• Historical post-merge checks, b4d759058c14bd92bf57455d9cbe14fb53a3a67b, after merging main e8ca9af: Windows HIP Release build/link passed; CPU encode produced finite 240 × 2560 embeddings; hidden-GPU --gpu failed without READY. Portable/default, native reconfiguration and CUDA/HIP exclusion configure checks passed. python tools/test_setup_choices.py passed 40 tests and python tools/test_setup_amd.py passed 30. Preserved local logs: logs/conflict-1287-20261007/{hip-build.log,encoder-smoke.json,cmake-matrix/summary.txt,review.txt}. These records were re-inspected; the commands were not rerun today.
• git diff --check e8ca9afd03d839d4f8dbbe82dffce7f8a3bafd7a...HEAD on the retained historical PR head passed on 2026-10-22.

Failed:

• None observed in today's source/metadata checks. Historical no-GPU and mutually exclusive backend rejections above were expected negative-test results.

Not run:

• Current-main / 0.2.2 HIP, CUDA, SYCL or Metal builds, current-main GPU execution, and current-main image/API inference. This update records source status only and requests no code review. The installed service was not restarted or replaced.
• GPU/API inference on the historical post-merge head b4d7590 was not repeated. Pre-merge measurements are not validation of that head or current main.
• Other GPUs, fine-detail OCR, grounding accuracy, long-context image stress and sustained load.

Limits

The retained runtime evidence applies only to the dated Strata 0.1.40 / RX 6900 XT configuration. No performance improvement or compatibility of current packages is claimed. In particular, main labels the new RX 6000 packages as built but not yet tested on a real card. This PR remains closed as a historical proposal; it was not merged.

[Copilot]

Copilot review overview

🟡 Changes recommended

Python launcher selection can prevent building, and deployment instructions omit configuration required by stock Windows HIP installations.

Review effort: Balanced
Findings: 2

Open (2)

• Document complete vision setup and engine flag for stock Windows HIP · New
• Keep Python fallback assignment inside the conditional · New

What changed in this PR

Adds a manual Windows HIP build path for Strata’s optional image encoder without changing stock installation or release packaging.

Changes:

• Adds HIP build configuration and a Windows build script.
• Requires GPU selection and successful warm-up before reporting readiness.
• Documents deployment and RX 6900 XT validation.

File	Description
tools/​vision/​strata_vision.cpp	Adds explicit GPU selection, logging, and warm-up failure handling.
tools/​vision/​CMakeLists.txt	Adds mutually exclusive HIP and CUDA options.
tools/​vision/​build_windows_hip.bat	Builds the encoder using a private ROCm environment.
docs/​AMD_HIP.md	Documents manual build and deployment.
bench/​results/​2026-10-07-windows-hip-vision-rx6900xt/​verification.json	Records hardware, runtime, and API verification results.
bench/​results/​2026-10-07-windows-hip-vision-rx6900xt/​README.md	Summarizes validation evidence and limitations.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Copilot review overview

🔵 Needs a closer look

Windows toolchain and GPU-runtime integration warrant human review of the recorded hardware validation, although no blocking defects were identified.

Review effort: Balanced
Findings: None

Resolved since last review (2)

• Keep Python fallback assignment inside the conditional
• Document complete vision setup and engine flag for stock Windows HIP

[lorum-ai]

I came across this, thank you for working on HIP vision. I’ve done similar work on lorum-ai/Strata’s codex/hip-vision branch with adding opt-in Windows setup integration, including the HIP encoder build, projector download, and runtime configuration.
I’m looking forward to seeing this in main and would be glad to help however I can.

[Niko1221]

Thanks for the pull request.

We're resetting the tracker for 0.2.0-beta. There are more than 1,300 open issues and pull requests, too many for us to review properly, so we're closing everything that isn't being worked on right now. Closed here does not mean rejected.

To reopen it takes about two minutes:

1. Rebase on main if it no longer applies, and re-run your checks on 0.2.0-beta.
2. Start the title with its area: [Engine], [GUI], [TUI], [Benchmark], [Docs] or [Other].
3. Fill in the pull request template: the problem and result, and the validation (what you ran, on which revision, what you saw).
4. Comment /reopen. A check runs and either reopens the pull request or lists what is still missing.

The rules are in CONTRIBUTING.md.