Support the XLA GPU compilation flags in Orbax

PiperOrigin-RevId: 879684156

Author: Orbax Authors

export/orbax/export/modules/obm_module_test.py

@@ -384,6 +384,36 @@ def test_obm_module_bfloat16_conversion(self, enable_bf16_optimization):
     with self.subTest('test_weights_b_dtype'):
       self.assertEqual(module.model_params['b'].dtype, expected_dtype)
 
+  def test_obm_module_gpu_xla_flags_integration_stable(self):
+    param_shape = (2, 5)
+    param_dtype = jnp.dtype(jnp.float32)
+    param_spec = jax.ShapeDtypeStruct(shape=param_shape, dtype=param_dtype)
+    model_function_name = 'simple_add'
+
+    jax2obm_options = obm_configs.Jax2ObmOptions(
+        checkpoint_path='checkpoint_path',
+        native_serialization_platforms=('cuda',),
+        xla_flags_per_platform={
+            'cuda': ['--xla_gpu_enable_latency_hiding_scheduler=true']
+        },
+    )
+
+    orbax_model_module = obm_module.ObmModule(
+        params=param_spec,
+        apply_fn={model_function_name: simple_add},
+        jax2obm_options=jax2obm_options,
+    )
+
+    xla_compile_options_map = (
+        orbax_model_module.xla_compile_options_per_platform
+    )
+    self.assertIsNotNone(xla_compile_options_map)
+    build_options_cuda = xla_compile_options_map.map['cuda']
+    self.assertIn(
+        'xla_gpu_enable_latency_hiding_scheduler',
+        build_options_cuda.env_option_overrides,
+    )
+
 
 class GetSharedValueTest(parameterized.TestCase):
 

model/orbax/experimental/model/core/python/compile_options_util.py

@@ -27,6 +27,7 @@
 from .google.protobuf import any_pb2
 from .platforms.xla.service.jellyfish import tpu_compilation_environment_pb2 as tpu_comp_env_pb2
 from .platforms.xla.service.jellyfish.python import tpu_compilation_environment as tpu_comp_env
+from .third_party.neptune.model._src.core import xla_gpu_flag_validation
 from tensorflow.compiler.xla import xla_data_pb2  # pylint: disable=g-direct-tensorflow-import
 from tensorflow.compiler.xla import xla_pb2  # pylint: disable=g-direct-tensorflow-import
 from tensorflow.compiler.xla.pjrt.proto import compile_options_pb2  # pylint: disable=g-direct-tensorflow-import
@@ -165,6 +166,10 @@ def generate_xla_compile_options(
   tpu_platform_name = manifest_pb2.Platform.Name(
       manifest_pb2.Platform.TPU
   ).lower()
+  cuda_platform_name = manifest_pb2.Platform.Name(
+      manifest_pb2.Platform.CUDA
+  ).lower()
+
   compile_options_map = manifest_pb2.CompileOptionsProtoMap()
   if native_serialization_platforms is None:
     # If no native serialization platforms are specified, we will set the
@@ -195,24 +200,97 @@ def generate_xla_compile_options(
         )
 
   for platform in platforms:
-    if platform.lower() == tpu_platform_name:
-      if xla_flags_per_platform:
-        xla_flags_overrides = xla_flags_per_platform.get(platform, None)
+    if xla_flags_per_platform:
+      xla_flags_overrides = xla_flags_per_platform.get(platform, None)
+      if xla_flags_overrides:
         _validate_xla_flags_setting(xla_flags_overrides, persist_xla_flags)
-      else:
-        xla_flags_overrides = None
+    else:
+      xla_flags_overrides = None
+
+    platform_lower = platform.lower()
+    if platform_lower == tpu_platform_name:
       compile_environment = _generate_tpu_compilation_env(xla_flags_overrides)
+    elif platform_lower == cuda_platform_name:
+      # GPU Trick: Empty proto to bypass 'None' check and enable jax_mesh
+      # serialization.
+      compile_environment = xla_pb2.CompilationEnvironmentsProto()
     else:
+      # CPU Path: Leave as None to preserve legacy portable execution behavior.
       compile_environment = None
-    compile_options_map.map[platform.lower()].CopyFrom(
-        _generate_compilation_options(compile_environment, jax_mesh)
+
+    compile_options = _generate_compilation_options(
+        compile_environment, jax_mesh
     )
+
+    # Inject env_option_overrides natively for GPU using a dedicated helper.
+    if platform_lower == cuda_platform_name and xla_flags_overrides:
+      _apply_gpu_compilation_env_options(compile_options, xla_flags_overrides)
+
+    compile_options_map.map[platform_lower].CopyFrom(compile_options)
+
   if not persist_xla_flags:
     for compile_options in compile_options_map.map.values():
       compile_options.executable_build_options.comp_envs.Clear()
   return compile_options_map
 
 
+def _apply_gpu_compilation_env_options(
+    compile_options: compile_options_pb2.CompileOptionsProto,
+    xla_flags_overrides: Sequence[str],
+) -> None:
+  """Applies XLA flag overrides generically for GPU platforms.
+
+  Args:
+    compile_options: The compilation options proto to be modified.
+    xla_flags_overrides: A sequence of XLA flags to apply as option overrides.
+  """
+  overrides_map = _parse_env_option_overrides(xla_flags_overrides)
+  for k, v in overrides_map.items():
+    compile_options.env_option_overrides[k].CopyFrom(v)
+
+
+def _parse_env_option_overrides(
+    xla_flags: Sequence[str],
+) -> dict[str, compile_options_pb2.OptionOverrideProto]:
+  """Parses a list of XLA flags into a dictionary of OptionOverrideProto."""
+  overrides = {}
+  for flag in xla_flags:
+    if not flag.startswith('--'):
+      raise ValueError(f"Flag {flag} must start with '--'")
+
+    try:
+      # Use the C++ ValidateXlaGPUFlag logic to ensure consistent policy
+      # enforcement across Python and C++ layers.
+      # The C++ function expects the flag with the '--' prefix.
+      xla_gpu_flag_validation.validate_xla_gpu_flag(flag, strict=True)
+    except Exception as e:
+      # pybind11_abseil appends the status code name to the exception string.
+      # Remove it to match exactly what users would see from the C++ binaries.
+      err_msg = str(e)
+      if err_msg.endswith(' [INVALID_ARGUMENT]'):
+        err_msg = err_msg.removesuffix(' [INVALID_ARGUMENT]')
+      raise ValueError(err_msg) from e
+
+    key, value = flag[2:].split('=', 1)
+    override_proto = compile_options_pb2.OptionOverrideProto()
+
+    # Infer type (True/False/Int/Float/String)
+    if value.lower() == 'true':
+      override_proto.bool_field = True
+    elif value.lower() == 'false':
+      override_proto.bool_field = False
+    elif value.isdigit() or (value.startswith('-') and value[1:].isdigit()):
+      override_proto.int_field = int(value)
+    else:
+      try:
+        override_proto.double_field = float(value)
+      except ValueError:
+        override_proto.string_field = value
+
+    overrides[key] = override_proto
+  return overrides
+
+
 def _validate_xla_flags_setting(
     xla_flags_overrides: Sequence[str] | None, persist_xla_flags: bool
 ) -> None:

model/orbax/experimental/model/core/python/compile_options_util_test.py

@@ -328,6 +328,121 @@ def test_generate_xla_compile_options_xla_flags_no_persist_raise_error(self):
           persist_xla_flags=False,
       )
 
+  def test_generate_xla_compile_options_env_overrides(self):
+    compile_options_map = compile_options_util.generate_xla_compile_options(
+        native_serialization_platforms=['cuda'],
+        xla_flags_per_platform={
+            'cuda': [
+                '--xla_gpu_enable_latency_hiding_scheduler=true',
+                '--xla_gpu_autotune_level=0',
+            ]
+        },
+        persist_xla_flags=True,
+    )
+    self.assertIn('cuda', compile_options_map.map)
+    compile_options = compile_options_map.map['cuda']
+
+    overrides = compile_options.env_option_overrides
+    self.assertIn('xla_gpu_enable_latency_hiding_scheduler', overrides)
+    self.assertTrue(
+        overrides['xla_gpu_enable_latency_hiding_scheduler'].bool_field
+    )
+
+    self.assertIn('xla_gpu_autotune_level', overrides)
+    self.assertEqual(overrides['xla_gpu_autotune_level'].int_field, 0)
+
+  def test_generate_xla_compile_options_gpu_flags_experimental_rejection(self):
+    with self.assertRaisesRegex(
+        ValueError,
+        r'XLA GPU compilation flag --xla_gpu_experimental_flag=true is not'
+        r' supported. Please check field description at'
+        r' CompilationConfig::xla_gpu_flags',
+    ):
+      compile_options_util.generate_xla_compile_options(
+          native_serialization_platforms=['cuda'],
+          xla_flags_per_platform={'cuda': ['--xla_gpu_experimental_flag=true']},
+          persist_xla_flags=True,
+      )
+
+  @parameterized.named_parameters(
+      dict(
+          testcase_name='bool_true',
+          flag='--xla_gpu_enable_latency_hiding_scheduler=true',
+          expected_key='xla_gpu_enable_latency_hiding_scheduler',
+          expected_field='bool_field',
+          expected_value=True,
+      ),
+      dict(
+          testcase_name='bool_false',
+          flag='--xla_gpu_enable_latency_hiding_scheduler=false',
+          expected_key='xla_gpu_enable_latency_hiding_scheduler',
+          expected_field='bool_field',
+          expected_value=False,
+      ),
+      dict(
+          testcase_name='bool_uppercase_true',
+          flag='--xla_gpu_enable_latency_hiding_scheduler=TRUE',
+          expected_key='xla_gpu_enable_latency_hiding_scheduler',
+          expected_field='bool_field',
+          expected_value=True,
+      ),
+      dict(
+          testcase_name='int_positive',
+          flag='--xla_gpu_autotune_level=4',
+          expected_key='xla_gpu_autotune_level',
+          expected_field='int_field',
+          expected_value=4,
+      ),
+      dict(
+          testcase_name='int_negative',
+          flag='--xla_gpu_nccl_termination_timeout_seconds=-1',
+          expected_key='xla_gpu_nccl_termination_timeout_seconds',
+          expected_field='int_field',
+          expected_value=-1,
+      ),
+      dict(
+          testcase_name='float_positive',
+          flag='--xla_gpu_auto_spmd_partitioning_memory_budget_ratio=1.5',
+          expected_key='xla_gpu_auto_spmd_partitioning_memory_budget_ratio',
+          expected_field='double_field',
+          expected_value=1.5,
+      ),
+      dict(
+          testcase_name='float_negative',
+          flag='--xla_gpu_auto_spmd_partitioning_memory_budget_ratio=-0.5',
+          expected_key='xla_gpu_auto_spmd_partitioning_memory_budget_ratio',
+          expected_field='double_field',
+          expected_value=-0.5,
+      ),
+      dict(
+          testcase_name='string_value',
+          flag='--xla_gpu_cuda_data_dir=/usr/local/cuda',
+          expected_key='xla_gpu_cuda_data_dir',
+          expected_field='string_field',
+          expected_value='/usr/local/cuda',
+      ),
+  )
+  @mock.patch.object(
+      compile_options_util.xla_gpu_flag_validation, 'validate_xla_gpu_flag'
+  )
+  def test_generate_xla_compile_options_gpu_flags_type_inference(
+      self, mock_validate, flag, expected_key, expected_field, expected_value
+  ):
+    del mock_validate  # Unused, just patching for bypass
+    compile_options_map = compile_options_util.generate_xla_compile_options(
+        native_serialization_platforms=['cuda'],
+        xla_flags_per_platform={'cuda': [flag]},
+        persist_xla_flags=True,
+    )
+    self.assertIsNotNone(compile_options_map.map)
+    build_options_cuda = compile_options_map.map['cuda']
+    self.assertIn(expected_key, build_options_cuda.env_option_overrides)
+    override_proto = build_options_cuda.env_option_overrides[expected_key]
+    with self.subTest('test_oneof_field'):
+      self.assertEqual(override_proto.WhichOneof('value'), expected_field)
+    with self.subTest('test_value'):
+      self.assertEqual(getattr(override_proto, expected_field), expected_value)
+
   @parameterized.named_parameters(
       dict(
           testcase_name='1d_mesh',