Disable autocast cache for tensor views as fix for pytorch#48049 (pytorch#48696) (pytorch#48936)

Co-authored-by: pbialecki <pbialecki@nvidia.com>

Author: malfet

aten/src/ATen/autocast_mode.cpp

@@ -68,7 +68,7 @@ Tensor cached_cast(at::ScalarType to_type, const Tensor& arg) {
   if (is_eligible(arg) && (arg.scalar_type() != to_type)) {
     // Heuristic:  Do what Apex does, and cache fp16 casts of fp32 model weights (leaves).
     // See cached_casts declaration above for detailed strategy.
-    bool can_try_cache = (to_type == at::kHalf && arg.scalar_type() == at::kFloat && arg.requires_grad() && arg.is_leaf());
+    bool can_try_cache = (to_type == at::kHalf && arg.scalar_type() == at::kFloat && arg.requires_grad() && arg.is_leaf() && !arg.is_view());
     if (can_try_cache) {
       auto it = cached_casts.find(arg.unsafeGetTensorImpl());
       if (it != cached_casts.end()) {

test/test_cuda.py

@@ -2706,6 +2706,22 @@ def test_autocast_rnn(self):
                 for grad, grad_control in zip(grads, grads_control):
                     self.assertEqual(grad.half(), grad_control)
 
+    def test_autocast_cache_leak(self):
+        # Reported at https://github.com/pytorch/pytorch/issues/48049
+        # Test is used to check, if autocast recaches the same parameters
+        # when executed in a `torch.no_grad()` block.
+
+        linear = torch.nn.Linear(10, 10).to('cuda')
+        data = torch.randn(1, 10, device='cuda')
+
+        with torch.cuda.amp.autocast():
+            with torch.no_grad():
+                out = linear(data)
+                first_iter_mem = torch.cuda.memory_allocated()
+                for _ in range(3):
+                    out = linear(data)
+                self.assertTrue(first_iter_mem == torch.cuda.memory_allocated())
+
     @slowTest
     @unittest.skipIf(not TEST_LARGE_TENSOR, "not enough memory")
     def test_max_large_axis(self):