feat: add ROCm (AMD GPU) and Intel GPU support (v0.2.12)

New image variants:
- :rocm  — InsightFace via ROCmExecutionProvider, SigLIP via PyTorch ROCm 6.3
- :intel — InsightFace via OpenVINOExecutionProvider (onnxruntime-openvino);
           Intel GPU compute runtime auto-installed from Intel graphics repo;
           OPENVINO_DEVICE=GPU opts into Arc/iGPU inference (default: CPU)

Also adds:
- pyproject-rocm.toml + uv-rocm.lock, pyproject-intel.toml + uv-intel.lock
- compose.yml device passthrough snippets for AMD and Intel
- CI: build-rocm and build-intel jobs in docker-publish.yml; all four
  variants built and tagged in release.yml
- README reworked: cleaner structure, GPU variant quick-start examples,
  OPENVINO_DEVICE env var documented
- CHANGELOG entry and version bump to 0.2.12

Fix: IntPrompt in dict literal was eagerly evaluated in the no-embedding
fallback path of _get_strategy_choice, prompting users for a custom count
regardless of which strategy they picked.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

winnow/embeddings.py

@@ -85,15 +85,16 @@ def get_insightface_app():
         return _insightface_app
     _insightface_loaded = True
 
-    # Preload CUDA/cuDNN DLLs BEFORE any ORT InferenceSession is created
-    _preload_cuda_libs()
-
     ctx_id = -1
     insightface_home = os.environ.get("INSIGHTFACE_HOME", os.path.expanduser("~/.insightface"))
     try:
         import onnxruntime as ort
         from insightface.app import FaceAnalysis
 
+        # Preload CUDA/cuDNN DLLs before any ORT InferenceSession is created.
+        # Silently no-ops on ROCm/Intel builds where preload_dlls() is absent.
+        _preload_cuda_libs()
+
         # Disk cache check — lets the user know whether a download is coming
         buffalo_path = Path(insightface_home) / "models" / "buffalo_l"
         if buffalo_path.exists() and any(buffalo_path.iterdir()):
@@ -110,18 +111,35 @@ def get_insightface_app():
             "ROCmExecutionProvider",
             "MPSExecutionProvider",
             "CoreMLExecutionProvider",
+            "OpenVINOExecutionProvider",
         }
         has_gpu_provider = bool(gpu_providers & set(providers))
         ctx_id = -1 if _is_force_cpu() else (0 if has_gpu_provider else -1)
 
+        # For OpenVINO EP, inject device_type from env var (default CPU; set GPU for Intel Arc/iGPU)
+        has_openvino = "OpenVINOExecutionProvider" in providers
+        if has_openvino:
+            openvino_device = os.getenv("OPENVINO_DEVICE", "CPU")
+            providers = [
+                ("OpenVINOExecutionProvider", {"device_type": openvino_device})
+                if p == "OpenVINOExecutionProvider" else p
+                for p in providers
+            ]
+            logger.debug(f"OpenVINO EP: device_type={openvino_device}")
+
         if not has_gpu_provider and not _is_force_cpu():
             logger.warning(
                 "No GPU execution provider found — running InsightFace on CPU. "
-                "If you have an NVIDIA GPU, ensure the NVIDIA Container Toolkit is "
-                "installed and the container has GPU access (deploy.resources in compose)."
+                "Ensure the container has GPU access and the correct variant image is used "
+                "(gpu for NVIDIA, rocm for AMD, intel for Intel Arc/iGPU)."
             )
 
-        device_str = "GPU" if ctx_id >= 0 else "CPU"
+        if ctx_id < 0:
+            device_str = "CPU"
+        elif has_openvino:
+            device_str = f"OpenVINO ({os.getenv('OPENVINO_DEVICE', 'CPU')})"
+        else:
+            device_str = "GPU"
         logger.info(f"InsightFace Buffalo_L: loading into memory on {device_str}...")
 
         t0 = time.time()
@@ -223,11 +241,14 @@ def get_siglip_model():
 
         _siglip_model.eval()
 
-        # Move to GPU if available
+        # Move to GPU if available (ROCm builds expose torch.cuda.is_available() == True)
         if not _is_force_cpu():
             if torch.cuda.is_available():
                 _siglip_model = _siglip_model.cuda()
                 device_name = "CUDA GPU"
+            elif hasattr(torch, "xpu") and torch.xpu.is_available():
+                _siglip_model = _siglip_model.to("xpu")
+                device_name = "Intel XPU"
             elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
                 _siglip_model = _siglip_model.to("mps")
                 device_name = "Apple MPS"

winnow/image_processing.py

@@ -144,7 +144,14 @@ def process_object_mode(
     try:
         model = get_yolo_model()
         target_class = config.get("object_class", "dog")
-        device = "cpu" if os.getenv("FORCE_CPU", "").lower() in ("true", "1", "yes") else None
+        import torch
+
+        if os.getenv("FORCE_CPU", "").lower() in ("true", "1", "yes"):
+            device = "cpu"
+        elif hasattr(torch, "xpu") and torch.xpu.is_available():
+            device = "xpu"
+        else:
+            device = None  # YOLO auto-selects (CUDA/ROCm/CPU)
 
         results = model(img, verbose=False, device=device)
 

winnow/jobs.py

@@ -58,8 +58,11 @@ def _get_strategy_choice(has_embedding: bool, entity_type: str) -> tuple[int | s
     rprint("  [bold]4.[/bold] Skip")
 
     choice = Prompt.ask("Choice", choices=["1", "2", "3", "4"], default="1")
-    limits = {"1": 30, "2": 100, "3": IntPrompt.ask("Enter number of images", default=30)}
-    return limits.get(choice, 0), "time" if choice != "4" else "skip"
+    if choice == "4":
+        return 0, "skip"
+    if choice == "3":
+        return IntPrompt.ask("Enter number of images", default=30), "time"
+    return {"1": 30, "2": 100}.get(choice, 30), "time"
 
 
 def _resolve_strategy(strategy: str, has_embedding: bool) -> tuple[int | str, str]: