Fix RVM inference on GPU execution providers

CMakeLists.txt

@@ -267,8 +267,8 @@ set(
   data/models/mediapipe.with_runtime_opt.ort.license
   data/models/pphumanseg_fp32.with_runtime_opt.ort
   data/models/pphumanseg_fp32.with_runtime_opt.ort.license
-  data/models/rvm_mobilenetv3_fp32.with_runtime_opt.ort
-  data/models/rvm_mobilenetv3_fp32.with_runtime_opt.ort.license
+  data/models/rvm_mobilenetv3_fp32.onnx
+  data/models/rvm_mobilenetv3_fp32.onnx.license
   data/models/selfie_multiclass_256x256.with_runtime_opt.ort
   data/models/selfie_multiclass_256x256.with_runtime_opt.ort.license
   data/models/selfie_segmentation.with_runtime_opt.ort

src/consts.h

@@ -12,7 +12,9 @@ const char *const MODEL_SINET = "models/SINet_Softmax_simple.with_runtime_opt.or
 const char *const MODEL_MEDIAPIPE = "models/mediapipe.with_runtime_opt.ort";
 const char *const MODEL_SELFIE = "models/selfie_segmentation.with_runtime_opt.ort";
 const char *const MODEL_SELFIE_MULTICLASS = "models/selfie_multiclass_256x256.with_runtime_opt.ort";
-const char *const MODEL_RVM = "models/rvm_mobilenetv3_fp32.with_runtime_opt.ort";
+// Use the original ONNX model for RVM: the pre-optimized .ort variant is
+// CPU-EP-targeted and crashes GPU execution providers (e.g. CUDA).
+const char *const MODEL_RVM = "models/rvm_mobilenetv3_fp32.onnx";
 const char *const MODEL_PPHUMANSEG = "models/pphumanseg_fp32.with_runtime_opt.ort";
 const char *const MODEL_ENHANCE_TBEFN = "models/tbefn_fp32.with_runtime_opt.ort";
 const char *const MODEL_ENHANCE_URETINEX = "models/uretinex_net_180x320.with_runtime_opt.ort";

src/models/Model.hpp

@@ -15,6 +15,7 @@
 
 #include <opencv2/imgproc.hpp>
 #include <algorithm>
+#include <cstring>
 
 template<typename T> T vectorProduct(const std::vector<T> &v)
 {
@@ -282,8 +283,25 @@ class Model {
 			rawOutputNames.push_back(outputName.get());
 		}
 
-		session->Run(Ort::RunOptions{nullptr}, rawInputNames.data(), inputTensor.data(), inputNames.size(),
-			     rawOutputNames.data(), outputTensor.data(), outputNames.size());
+		// Let ORT allocate the output tensors itself, then copy the results
+		// back into the pre-allocated CPU buffers. Pre-binding CPU output
+		// tensors can misbehave on GPU execution providers for stateful
+		// multi-output models (e.g. RVM).
+		std::vector<Ort::Value> results = session->Run(Ort::RunOptions{nullptr}, rawInputNames.data(),
+							       inputTensor.data(), inputNames.size(),
+							       rawOutputNames.data(), outputNames.size());
+
+		for (size_t i = 0; i < results.size() && i < outputTensor.size(); i++) {
+			const float *src = results[i].GetTensorData<float>();
+			float *dst = outputTensor[i].GetTensorMutableData<float>();
+			const size_t srcCount = results[i].GetTensorTypeAndShapeInfo().GetElementCount();
+			const size_t dstCount = outputTensor[i].GetTensorTypeAndShapeInfo().GetElementCount();
+			if (srcCount != dstCount) {
+				obs_log(LOG_WARNING, "Output %d size mismatch: model produced %d elements, expected %d",
+					(int)i, (int)srcCount, (int)dstCount);
+			}
+			memcpy(dst, src, (srcCount < dstCount ? srcCount : dstCount) * sizeof(float));
+		}
 	}
 };