Pull requests / #685
#685 Add optional Windows HIP vision encoder builds and package validation
closed · @sifer · 0 commentaires · Sur GitHub
Setup & installServer & APIAMD / HIPModels & quantsDocumentationWindowsLinux
Description
Windows HIP setup currently turns image input off, and the Windows AMD package has no image encoder. This adds an optional CPU or HIP encoder to the existing Windows build and packaging path, allowing `--vision cpu` or `--vision gpu` when the installed package supports it. - Add `STRATA_HIP_VISION=none|cpu|gpu` to `tools/hip/build_windows.bat`; the default remains text-only. Use the pinned llama.cpp checkout and existing ROCm toolchain, and reconfigure the engine so its architectures agree with the encoder and package metadata. Explicitly pass the custom source override (or clear it), so a fresh-prompt rebuild uses the same checkout for both programs. - Include the encoder's recursive DLL dependencies, record its mode/architectures, validate package capabilities before installation, and preserve encoder capability during package updates. CPU mode hides HIP devices before backend initialization. - Add hardware-free setup/package/update regressions and document the Windows build and measured validation in the AMD and general installation guides. Include a checklist for other card owners, following the existing HIP port and community-report pattern. ### Validation report | Check | Result | | --- | --- | | Setup, package and Windows batch-build regression tests | 184 passed; no GPU or ROCm required | | Shared server/lifecycle tests | 126 passed | | Windows HIP engine + encoder build/package; device self-test | Passed for gfx1100 | | Standalone CPU/HIP image embeddings | Finite and distinct for red/blue images; OCR embeddings finite in both modes | | Image API checks: both APIs, ordered images, OpenAI streaming, repeated image, text after images | HIP: 6/6 at 8192 and 131072 configured context; CPU: 6/6 at 131072 | | Clean-text OCR through OpenAI-compatible API | 6/6 exact transcriptions in each encoder mode | | Independent subagent reviews | Green after fixes | Hardware checks: Windows 11, RX 7900 XTX 24 GiB (gfx1100), Ryzen 7 7800X3D, 32 GB RAM, driver 32.0.31041.1004, TheRock ROCm 10.2.0a20260930, Coder IQ1_M and matching BF16 projector. The text engine stayed on HIP in both encoder modes. CPU API checks used four threads and the standard 300-token image cap; HIP used 1024. Cache sizes differed, so API timings are not a controlled CPU/GPU comparison. Measurements and the other-card checklist are in [docs/AMD_HIP.md](https://github.com/sifer/Strata/blob/windows-hip-vision/docs/AMD_HIP.md#windows). Not validated: other AMD cards, an actual Linux run in this session, full context windows, sustained/large-image workloads, scans/handwriting, browser attachments or Claude Code attachment handling. API checks used short prompts. Machine-specific smoke scripts, models, binaries and runtime configs are excluded from this PR. Release follow-up: the standard Windows AMD package must enable the HIP encoder for its advertised architectures and publish the resulting ZIP. The default remains text-only. Validate the candidate package with bundled libraries on a PC without a ROCm development installation; additional card reports and sustained-image checks are still needed to broaden the measured support.
Sur le site
Liens install, modèles, releases.