#411 0.1.32 ready-made strata-vision.exe crashes with illegal instruction (0xC000001D) on a non-AVX-512 CPU (i9-9900KF)

本文

**Summary:** after updating to the 0.1.32 ready-made engine (`strata-windows-x64.zip`), `strata-vision.exe` dies at startup with `0xC000001D` (STATUS_ILLEGAL_INSTRUCTION). The process prints nothing and exits, so `serve/server.py` stops with `RuntimeError: the vision encoder did not start:` and the empty line it read. `strata.exe` from the same zip runs fine on the same PC: the model serves text on 0.1.32. The 0.1.30 `strata-vision.exe` ran on this machine without problems.

**PC**
- CPU: Intel i9-9900KF (Coffee Lake: AVX2, **no AVX-512**)
- GPUs: 2x RTX 5060 Ti 16 GB (sm_120) running the model as a layer split, plus an RTX 3060 12 GB (sm_86)
- Windows 10 Pro 19045
- `engine/BUILD.json`: `"version": "0.1.32", "source": "release", "archs": [75, 86, 89, 120], "vision": "gpu", "portable": true`

**Repro:** run the encoder directly, outside the server:
```
strata-vision.exe --mmproj mmproj-Qwen3.8-Flash-Next-BF16.gguf --model Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf --gpu
```
- with `--gpu` and every GPU visible: exit `0xC000001D`, no output
- with `--gpu` and `CUDA_VISIBLE_DEVICES` set to one card: same
- without `--gpu` (CPU): it prints `ggml_cuda_init: failed to initialize CUDA: no CUDA-capable device is detected`, then exits `0xC000001D`

The same happens under the server's own environment (`child_env`, PATH with the `nvidia/cu13` wheel bin): exit code `3221225501`.

**Guess:** the CPU-side ggml in the 0.1.32 vision build was compiled with AVX-512 (or `GGML_NATIVE` on the build machine) instead of the portable baseline the 0.1.30 build used. Running `strata-vision.exe --help` exits normally, so it seems to fail on the first ggml CPU code path, not at load.

**Workaround for now:** vision is turned off in the config, and the engine itself runs on 0.1.32. Thanks for #340, by the way: a ~7.5K-token prompt went from ~30 s to 10.4 s on our two 5060 Tis.