fix to segmentation fault when empty tensor is passed to choose_qpara… (#161966)

arkadip-maitra · pytorchmergebot · commit 1aa7476885e8 · 2025-09-03T20:26:26.000Z
…ms_optimized Fixes #153326 Minimal code to reproduce error: ``` import torch tensor = torch.tensor([]) torch.choose_qparams_optimized( tensor, 0, 200, 0.16, 8 ) ``` Previous Output: `Segmentation fault` Now Output: ``` Traceback (most recent call last): File "/home/amaitra/work/tests/issue_153326.py", line 5, in <module> torch.choose_qparams_optimized( RuntimeError: input tensor is empty and has no data ``` Caused because `const float* input_row =input_tensor.const_data_ptr<float>();` becomes null Pull Request resolved: #161966 Approved by: https://github.com/Skylion007
diff --git a/aten/src/ATen/native/quantized/QTensor.cpp b/aten/src/ATen/native/quantized/QTensor.cpp
@@ -335,14 +335,16 @@ std::tuple<Tensor, Tensor> choose_qparams_optimized(
     const int64_t n_bins,
     const double ratio,
     int64_t bit_width) {
+  const float* input_row = input_tensor.const_data_ptr<float>();
+  TORCH_CHECK_VALUE(input_row != nullptr, "input tensor is empty and has no data");
 
   if (numel < 0 || numel > input_tensor.numel()) {
     TORCH_CHECK(false, "numel is out of the bound of input tensor");
   }
 
   TORCH_CHECK(numel <= input_tensor.numel(), "numel ", numel,
       " greater than input_tensor.numel() ", input_tensor.numel());
-  const float* input_row = input_tensor.const_data_ptr<float>();
+
   float xmin = *std::min_element(input_row, input_row + numel);
   float xmax = *std::max_element(input_row, input_row + numel);
   float n_bins_float = static_cast<float>(n_bins);
diff --git a/test/quantization/core/test_quantized_tensor.py b/test/quantization/core/test_quantized_tensor.py
@@ -1409,6 +1409,9 @@ def test_choose_qparams_optimized(self):
             self.assertEqual(y[0].numpy(), ref[0])
             self.assertEqual(y[1].numpy(), ref[1])
 
+        with self.assertRaisesRegex(ValueError, "input tensor is empty and has no data"):
+            torch.choose_qparams_optimized(torch.tensor([]), numel=0, n_bins=200, ratio=0.16, bit_width=8)
+
     def _test_pickle_checkpoint_qtensor(self, device):
         with TemporaryFileName() as fname:
             class M(torch.jit.ScriptModule):
