Add autocast support for einsum (#8420)

test/test_autocast_xla.py

@@ -0,0 +1,53 @@
+import re
+import torch
+import torch_xla
+import torch_xla.core.xla_model as xm
+import unittest
+
+device = xm.xla_device()
+
+
+class TestAutocastXla(unittest.TestCase):
+
+  def test_cross_entropy_loss(self):
+    data = torch.randn(16, 10).to(torch.bfloat16).to(device)
+    target = torch.randn(16, 10).to(torch.bfloat16).to(device)
+
+    with torch.autocast("xla"):
+      loss = torch.nn.CrossEntropyLoss()(data, target)
+      hlo = torch_xla._XLAC._get_xla_tensors_hlo([loss])
+      self.assertRegex(hlo, r".*convert.*f32.*convert.*bf16")
+      self.assertRegex(hlo, r".*exponential.*f32.*exponential.*f32")
+      self.assertRegex(hlo, r".*log.*f32.*log.*f32")
+
+  def test_einsum(self):
+    # irrespective of input dtype, output dtype will depend on autocast policy.
+    # Tests for bf16 and f32 given below.
+
+    # input data of type bf16
+    data = torch.randn(16, 10).to(torch.bfloat16).to(device)
+    target = torch.randn(5, 10).to(torch.bfloat16).to(device)
+
+    with torch.autocast("xla"):
+      product = torch.einsum("...n,mn->...m", data, target)
+      # test the HLO to see if autocast works for einsum op, which would show up as a dot op in the HLO
+      hlo = torch_xla._XLAC._get_xla_tensors_hlo([product])
+      # Verify that dot op has bf16 output and not f32, i.e. the computation is performed in bfloat16 precision by autocast
+      self.assertRegex(hlo, r".*dot.*bf16")
+      self.assertNotRegex(hlo, r".*dot.*f32")
+
+    # input data of type fp32
+    data32 = torch.randn(16, 10).to(torch.float32).to(device)
+    target32 = torch.randn(5, 10).to(torch.float32).to(device)
+
+    with torch.autocast("xla"):
+      product = torch.einsum("...n,mn->...m", data32, target32)
+      # test the HLO to see if autocast works for einsum op, which would show up as a dot op in the HLO
+      hlo = torch_xla._XLAC._get_xla_tensors_hlo([product])
+      # Verify that dot op has bf16 output and not f32, i.e. the computation is performed in bfloat16 precision by autocast
+      self.assertRegex(hlo, r".*dot.*bf16")
+      self.assertNotRegex(hlo, r".*dot.*f32")
+
+
+if __name__ == "__main__":
+  unittest.main()

torch_xla/csrc/autocast_mode.cpp

@@ -48,6 +48,7 @@ TORCH_LIBRARY_IMPL(aten, AutocastXLA, m) {
   KERNEL_XLA(prelu, lower_precision_fp)
   KERNEL_XLA(relu, lower_precision_fp)
   KERNEL_XLA(max_pool2d, lower_precision_fp)
+  KERNEL_XLA(einsum, lower_precision_fp)
   // Disable `scaled_dot_product_attention` for now since it causes
   // undefined symbol with official torch whl.
   // KERNEL_XLA(scaled_dot_product_attention, lower_precision_fp)