Re-introduce "XLA_USE_32BIT_LONG" flag (#8589)

pytorch · Jan 17, 2025 · fccd395 · fccd395
1 parent 1dd4969
commit fccd395
Show file tree

Hide file tree

Showing 7 changed files with 75 additions and 57 deletions.
diff --git a/test/neuron/run_tests.sh b/test/neuron/run_tests.sh
@@ -160,9 +160,7 @@ function run_xla_op_tests1 {
   run_test "$CDIR/dynamo/test_graph_input_matcher.py"
   run_test "$CDIR/dynamo/test_dynamo_config.py"
   run_save_tensor_ir run_test "$CDIR/dynamo/test_dynamo_graph_dump.py"
-  #run_test "$CDIR/test_data_type.py"
-  run_use_bf16 "$CDIR/test_data_type.py"
-  run_downcast_bf16 "$CDIR/test_data_type.py"
+  run_test "$CDIR/test_data_type.py"
   #run_test "$CDIR/test_fp8.py"
   run_xla_ir_debug "$CDIR/test_env_var_mapper.py"
   run_xla_hlo_debug "$CDIR/test_env_var_mapper.py"

diff --git a/test/neuron/test_neuron_data_types.py b/test/neuron/test_neuron_data_types.py
@@ -27,10 +27,10 @@ def test_datatypes(self):
                   (torch.double, "f32", torch.floor_divide),
                   (torch.int16, "s32", torch.add),
                   (torch.int32, "s32", torch.add),
-                  (torch.int64, "s32", torch.add),
+                  (torch.int64, "s64", torch.add),
                   (torch.uint16, "u32", torch.add),
                   (torch.uint32, "u32", torch.add),
-                  (torch.uint64, "u32", torch.add)]
+                  (torch.uint64, "u64", torch.add)]
 
     for dtype, op_xla_dtype, op in test_cases:
       with self.subTest(dtype=dtype, op_xla_dtype=op_xla_dtype, op=op):

diff --git a/test/run_tests.sh b/test/run_tests.sh
@@ -178,8 +178,6 @@ function run_xla_op_tests1 {
   run_test "$CDIR/dynamo/test_dynamo_config.py"
   run_save_tensor_ir run_test "$CDIR/dynamo/test_dynamo_graph_dump.py"
   run_test "$CDIR/test_data_type.py"
-  run_use_bf16 "$CDIR/test_data_type.py"
-  run_downcast_bf16 "$CDIR/test_data_type.py"
   run_test "$CDIR/test_fp8.py"
   run_xla_ir_debug run_test "$CDIR/test_env_var_mapper.py"
   run_xla_hlo_debug run_test "$CDIR/test_env_var_mapper.py"

diff --git a/test/test_data_type.py b/test/test_data_type.py
@@ -1,73 +1,82 @@
 import os
+import sys
+import unittest
 
 import torch
 import torch_xla
 import torch_xla.core.xla_model as xm
 import torch_xla.utils.utils as xu
-import unittest
 
 
-def check_env_flag(name, default=''):
-  return os.getenv(name, default).upper() in ['ON', '1', 'YES', 'TRUE', 'Y']
+class XlaDataTypeTest(unittest.TestCase):
 
+  def setUp(cls):
+    cls.original_env = {
+        'XLA_USE_BF16': os.environ.get('XLA_USE_BF16'),
+        'XLA_DOWNCAST_BF16': os.environ.get('XLA_DOWNCAST_BF16'),
+        'XLA_USE_32BIT_LONG': os.environ.get('XLA_USE_32BIT_LONG')
+    }
 
-class XlaDataTypeTest(unittest.TestCase):
+  def tearDown(self):
+    for key, value in self.original_env.items():
+      if value is None:
+        os.environ.pop(key, None)
+      else:
+        os.environ[key] = value
 
-  def test_datatype_f32(self):
-    t1 = torch.tensor([2.0, 3.0], dtype=torch.float, device=xm.xla_device())
-    t2 = torch.tensor([2.0, 3.0], dtype=torch.float, device=xm.xla_device())
-    t3 = torch.div(t1, t2, rounding_mode='floor')
-    assert t3.dtype == torch.float
+  def _set_env(self, **kwargs):
+    for key, value in kwargs.items():
+      os.environ[key] = value
 
-    hlo_text = torch_xla._XLAC._get_xla_tensors_text([t3])
-    device_data_hlo = hlo_text.split('\n')[1]
-    assert 'xla::device_data' in device_data_hlo, device_data_hlo
-    if check_env_flag('XLA_USE_BF16') or check_env_flag('XLA_DOWNCAST_BF16'):
-      assert 'bf16' in device_data_hlo, device_data_hlo
-    elif check_env_flag('XLA_USE_FP16') or check_env_flag('XLA_DOWNCAST_FP16'):
-      assert 'f16' in device_data_hlo, device_data_hlo
-    else:
-      assert 'f32' in device_data_hlo, device_data_hlo
-
-  def test_datatype_f64(self):
-    t1 = torch.tensor([2.0, 3.0], dtype=torch.double, device=xm.xla_device())
-    t2 = torch.tensor([2.0, 3.0], dtype=torch.double, device=xm.xla_device())
-    t3 = torch.div(t1, t2, rounding_mode='floor')
-    assert t3.dtype == torch.double
+  def _test_datatype(self, dtype, expected_type, op):
+    t1 = torch.tensor([2, 3], dtype=dtype, device=xm.xla_device())
+    t2 = torch.tensor([2, 3], dtype=dtype, device=xm.xla_device())
+    t3 = op(t1, t2)
+    self.assertEqual(t3.dtype, dtype)
 
     hlo_text = torch_xla._XLAC._get_xla_tensors_text([t3])
-    device_data_hlo = hlo_text.split('\n')[1]
-    assert 'xla::device_data' in device_data_hlo, device_data_hlo
-    if check_env_flag('XLA_USE_BF16'):
-      assert 'bf16' in device_data_hlo, device_data_hlo
-    elif check_env_flag('XLA_USE_FP16'):
-      assert 'f16' in device_data_hlo, device_data_hlo
-    elif check_env_flag('XLA_DOWNCAST_BF16') or check_env_flag(
-        'XLA_DOWNCAST_FP16'):
-      assert 'f32' in device_data_hlo, device_data_hlo
-    else:
-      assert 'f64' in device_data_hlo, device_data_hlo
+    device_data_hlo = hlo_text.split('\n')[2]
+    self.assertIn('xla::device_data', device_data_hlo)
+    self.assertIn(expected_type, device_data_hlo)
+
+  def test_datatype_use_bf16(self):
+    self._set_env(XLA_USE_BF16='1')
+    self._test_datatype(torch.double, 'bf16', torch.floor_divide)
+    self._test_datatype(torch.float, 'bf16', torch.floor_divide)
+
+  def test_datatype_downcast_bf16(self):
+    self._set_env(XLA_DOWNCAST_BF16='1')
+    self._test_datatype(torch.double, 'bf16', torch.floor_divide)
+    self._test_datatype(torch.float, 'bf16', torch.floor_divide)
+
+  def test_datatype_use_32bit_long(self):
+    self._set_env(XLA_USE_32BIT_LONG='1')
+    self._test_datatype(torch.int64, 's32', torch.add)
+    self._test_datatype(torch.uint64, 'u32', torch.add)
 
   def test_module_to_dtype(self):
     device = torch_xla.device()
     linear = torch.nn.Linear(
         5, 10, dtype=torch.float32).to(device).to(torch.bfloat16)
-    input = torch.randn(
-        10,
-        5,
-    ).to(device).to(torch.bfloat16)
+    input = torch.randn(10, 5).to(device).to(torch.bfloat16)
     xm.mark_step()
     res = linear(input)
 
     hlo_text = torch_xla._XLAC._get_xla_tensors_text([res])
     res_hlo = hlo_text.split('\n')[-3]
-    assert 'bf16' in res_hlo, res_hlo
+    self.assertIn('bf16', res_hlo)
 
     linear_weight_hlo = torch_xla._XLAC._get_xla_tensors_text([linear.weight
                                                               ]).split('\n')[-3]
-    assert 'bf16' in linear_weight_hlo, linear_weight_hlo
+    self.assertIn('bf16', linear_weight_hlo)
 
 
 if __name__ == '__main__':
-  test = unittest.main()
-  sys.exit(0 if test.result.wasSuccessful() else 1)
+  suite = unittest.TestSuite()
+  suite.addTest(XlaDataTypeTest("test_datatype_use_bf16"))
+  suite.addTest(XlaDataTypeTest("test_datatype_downcast_bf16"))
+  suite.addTest(XlaDataTypeTest("test_datatype_use_32bit_long"))
+  suite.addTest(XlaDataTypeTest("test_module_to_dtype"))
+  runner = unittest.TextTestRunner(failfast=True)
+  result = runner.run(suite)
+  sys.exit(0 if result.wasSuccessful() else 1)
diff --git a/test/tpu/run_tests.sh b/test/tpu/run_tests.sh
@@ -45,6 +45,7 @@ python3 "$TEST_CDIR/torch_distributed/test_torch_distributed_multi_all_reduce_xl
 python3 "$TEST_CDIR/torch_distributed/test_torch_distributed_reduce_scatter_xla_backend.py"
 python3 "$TEST_CDIR/quantized_ops/test_dot_general.py"
 run_xla_ir_hlo_debug python3 "$TEST_CDIR/test_user_computation_debug_cache.py"
+python3 "$TEST_CDIR/test_data_type.py"
 
 # run examples, each test should takes <2 minutes
 python3 "$TEST_CDIR/../examples/data_parallel/train_resnet_spmd_data_parallel.py"

diff --git a/torch_xla/__init__.py b/torch_xla/__init__.py
@@ -178,9 +178,7 @@ def _setup_tpu_vm_library_path() -> bool:
 
 
 def _check_deprecated_env_var():
-  deprecated_env_vars = [
-      'XLA_USE_FP16', 'XLA_DOWNCAST_FP16', 'XLA_USE_32BIT_LONG'
-  ]
+  deprecated_env_vars = ['XLA_USE_FP16', 'XLA_DOWNCAST_FP16']
   for env_var in deprecated_env_vars:
     if os.environ.get(env_var):
       warnings.warn(f"The environment variable '{env_var}' is deprecated "

diff --git a/torch_xla/csrc/dtype.cpp b/torch_xla/csrc/dtype.cpp
@@ -30,6 +30,17 @@ bool ShouldDowncastToBF16() {
   return downcast_bf16;
 }
 
+bool ShouldUse32BitLong() {
+  bool use_32bit_long =
+      runtime::sys_util::GetEnvBool("XLA_USE_32BIT_LONG", false);
+  if (use_32bit_long) {
+    std::cout
+        << "XLA_USE_32BIT_LONG will be deprecated after the 2.6 release\n";
+    TF_LOG(INFO) << "Using 32bit integers for kLong values";
+  }
+  return use_32bit_long;
+}
+
 bool UseBF16() {
   static bool use_bf16 = ShouldUseBF16();
   return use_bf16;
@@ -40,6 +51,11 @@ bool DowncastBF16() {
   return downcast_bf16;
 }
 
+bool Use32BitLong() {
+  static bool use_32bit_long = ShouldUse32BitLong();
+  return use_32bit_long;
+}
+
 }  // namespace
 
 at::ScalarType TorchTypeFromXlaType(xla::PrimitiveType xla_type) {
@@ -143,11 +159,9 @@ xla::PrimitiveType MaybeDowncastToXlaDeviceType(
       return CheckNeuronDevice(hw_type) ? xla::PrimitiveType::S32
                                         : xla::PrimitiveType::S16;
     case xla::PrimitiveType::S64:
-      return CheckNeuronDevice(hw_type) ? xla::PrimitiveType::S32
-                                        : xla::PrimitiveType::S64;
+      return Use32BitLong() ? xla::PrimitiveType::S32 : xla::PrimitiveType::S64;
     case xla::PrimitiveType::U64:
-      return CheckNeuronDevice(hw_type) ? xla::PrimitiveType::U32
-                                        : xla::PrimitiveType::U64;
+      return Use32BitLong() ? xla::PrimitiveType::U32 : xla::PrimitiveType::U64;
     case xla::PrimitiveType::C128:
       return xla::PrimitiveType::C128;
     default: