Fall back to CPU — onnxruntime-rocm PyPI build lacks RDNA2 (gfx1031) kernels

onnxruntime-rocm is compiled for CDNA datacenter GPUs (gfx908/gfx90a/gfx942).
RX 6700 XT (gfx1031/RDNA2) has no matching HIP kernel binary, causing
hipErrorNoBinaryForGpu at inference time. CPU execution provider works
correctly; GPU path can be revisited with a custom onnxruntime build.

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

engine.py

@@ -26,30 +26,19 @@ _cond_cache: Dict[str, Dict[str, np.ndarray]] = {}
 
 
 def detect_device() -> str:
-    try:
-        import onnxruntime as ort
-        available = ort.get_available_providers()
-        if "ROCMExecutionProvider" in available or "MIGraphXExecutionProvider" in available:
-            return "rocm"
-    except Exception:
-        pass
     return "cpu"
 
 
 def _get_providers(device: str) -> list:
-    if device not in ("rocm", "cuda"):
-        return ["CPUExecutionProvider"]
-
+    # onnxruntime-rocm from PyPI is compiled for CDNA datacenter GPUs only
+    # (gfx908/gfx90a/gfx942). RDNA2 (gfx1031) has no compatible HIP kernel
+    # binary, producing hipErrorNoBinaryForGpu at inference time even though
+    # the session creates successfully. CPU provider used until a ROCm build
+    # that includes RDNA2 targets is available.
     import onnxruntime as ort
-    available = set(ort.get_available_providers())
-    providers = []
-    # MIGraphXExecutionProvider excluded — symbol mismatch between onnxruntime-rocm
-    # and apt migraphx; ROCMExecutionProvider covers GPU execution adequately.
-    if "ROCMExecutionProvider" in available:
-        providers.append(("ROCMExecutionProvider", {"device_id": 0}))
-    providers.append("CPUExecutionProvider")
-    logger.info(f"Available ORT providers: {available}")
-    return providers
+    available = ort.get_available_providers()
+    logger.info(f"Available ORT providers: {available} — using CPUExecutionProvider (RDNA2 not supported by this onnxruntime-rocm build)")
+    return ["CPUExecutionProvider"]
 
 
 def _ort_type_to_np(ort_type: str):