Issues / #1521
#1521 Experiment: HIP vision encoder on RX 9070 XT (Windows; 2x Xeon E5-2680 v4)
open · @XTony-MontanaX · 0 Kommentare · Auf GitHub
BenchmarksSetup & installMulti-GPUAMD / HIPNVIDIA / CUDAModels & quantsWindows
Beschreibung
## Objective Test whether Strata's standalone vision helper can produce image embeddings through HIP on an AMD GPU under Windows. This is distinct from the CPU image-encoder integration. Issue #304 completed the AMD CPU path and described HIP/GPU vision as a later step. ## Test setup - Host CPU: 2 x Intel Xeon E5-2680 v4 @ 2.40 GHz; 14 cores / 28 logical processors each, 28 cores / 56 logical processors total - GPU: AMD Radeon RX 9070 XT, gfx1201, 16 GB VRAM - OS/driver: Windows 11 Pro build 26100; Adrenalin 32.0.31041.1004 - ROCm 7.1; Strata/engine v0.1.40.3 - Model/projector: Qwen3.8-Flash-Next IQ3_XXS and mmproj-Qwen3.8-Flash-Next-BF16.gguf The vision helper was built separately with CMake/Ninja Release, GGML_HIP=ON, STRATA_VISION_CUDA=OFF, and STRATA_PORTABLE=OFF. LLVM-MinGW clang.exe was the host C compiler; a small wrapper invoked ROCm clang++.exe for gfx1201. The helper received HIP_VISIBLE_DEVICES=1 and the ROCm/LLVM-MinGW DLL directories in PATH. The sanitized config and build/runtime notes are in the attached ZIP. ## Observations - Vision files loaded in 2.7 s; warmup at a 1,024-token image budget took 2.1 s. - The helper log identified the RX 9070 XT (gfx1201) and HIP runtime; /health reported status=ok, loaded=true, images=true. - 1,549 MiB VRAM remained free with everything loaded. - One image (n=1) returned 990 image tokens and a 10,137,620-byte embedding payload. ## Interpretation and limits This is evidence that the manually built HIP encoder path returned embeddings on this setup. It is not a CPU-vs-GPU benchmark: there were no repeated trials, CPU baseline, end-to-end timing, OCR/answer-quality assessment, or GPU-utilization trace. The Xeon details identify the host and are not a performance claim. No source image or transcription is attached because the input was a private document. ## Request Please consider an officially supported Windows AMD GPU-vision option: documented HIP build/package steps for the helper, explicit device and ROCm DLL handling, and a repeatable encoder smoke test. (Obrigado pelo otimo trabalho fazendo o Strata @Niko1221 ) [strata-rx9070xt-hip-vision-evidence.zip](https://github.com/user-attachments/files/33195723/strata-rx9070xt-hip-vision-evidence.zip)
Mehr auf der Site
Links zu Install, Modellen, Releases.