Clear CUDA errors once reported and raise eval errors in buffer protocol

mlx/backend/cuda/event.cu

@@ -341,7 +341,8 @@ void Event::wait() {
   } else {
     event.atomic->wait(value());
   }
-  CHECK_CUDA_ERROR(cudaPeekAtLastError());
+  // Check for errors during kernel execution and reset the error state.
+  CHECK_CUDA_ERROR(cudaGetLastError());
   check_error();
 }
 

mlx/backend/cuda/utils.cpp

@@ -12,6 +12,8 @@ namespace mlx::core {
 
 void check_cuda_error(const char* name, cudaError_t err) {
   if (err != cudaSuccess) {
+    // Clear the error so it does not resurface in later checks.
+    cudaGetLastError();
     throw std::runtime_error(
         fmt::format("{} failed: {}", name, cudaGetErrorString(err)));
   }

python/src/buffer.h

@@ -88,9 +88,13 @@ extern "C" inline int getbuffer(PyObject* obj, Py_buffer* view, int flags) {
   std::memset(view, 0, sizeof(Py_buffer));
   auto a = nb::cast<mx::array>(nb::handle(obj));
 
-  {
+  // Exceptions can not propagate through the buffer protocol.
+  try {
     nb::gil_scoped_release nogil;
     a.eval();
+  } catch (const std::exception& e) {
+    PyErr_SetString(PyExc_RuntimeError, e.what());
+    return -1;
   }
 
   std::vector<Py_ssize_t> shape(a.shape().begin(), a.shape().end());

python/tests/test_array.py

@@ -2065,6 +2065,12 @@ def test_buffer_protocol_ref_counting(self):
         mv = None
         self.assertIsNone(wr())
 
+    def test_buffer_protocol_eval_error(self):
+        # Errors from evaluating the array are raised instead of aborting
+        a = mx.linalg.inv(mx.zeros((2, 2)), stream=mx.cpu)
+        with self.assertRaises(RuntimeError):
+            memoryview(a)
+
     def test_array_view_ref_counting(self):
         a = mx.arange(3)
         wr = weakref.ref(a)

tests/cuda_tests.cpp

@@ -109,3 +109,13 @@ TEST_CASE("test clear cache trims CUDA pool") {
       cudaSuccess);
   CHECK_LT(final_reserved, allocated_reserved);
 }
+
+TEST_CASE("test eval after cuda error") {
+  auto s = default_stream(Device::gpu);
+  // A failed allocation sets a CUDA error on the calling thread.
+  CHECK_THROWS(eval(zeros({1 << 20, 1 << 20}, float32, s)));
+  // The error must not resurface in later evaluations.
+  auto a = ones({4}, float32, s);
+  CHECK_NOTHROW(eval(a));
+  CHECK(array_equal(a, ones({4})).item<bool>());
+}