Improve assume_pure docs and tests (#9038)

Author: tengyifei

docs/source/perf/assume_pure.md

@@ -122,8 +122,13 @@ a fixed up-front cost, and then later runs will reuse the cached XLA computation
 ## Limitations
 
 Currently, all operations in a function wrapped with `@assume_pure` must be
-PyTorch upstream operations (e.g. `torch.einsum`, `torch.sin`, ...). More
-PyTorch/XLA operations (e.g. `mark_sharding`) will be supported in the future.
+PyTorch upstream operations (e.g. `torch.einsum`, `torch.sin`, ...), or these
+PyTorch/XLA operations:
+  * `torch_xla.experimental.assume_pure` (recursive `assume_pure`)
+  * `torch_xla.distributed.spmd.mark_sharding`
+
+More PyTorch/XLA operations (e.g. `flash_attention`) will be supported in the
+future.
 
 <!-- xrefs -->
 

infra/ansible/config/pip.yaml

@@ -28,7 +28,7 @@ pip:
       - tqdm
       - typing_extensions
       - sympy
-      - yapf==0.30.0
+      - yapf==0.40.2
 
     build_amd64:
       - mkl

test/run_tests.sh

@@ -212,6 +212,7 @@ function run_xla_op_tests2 {
   XLA_USE_SPMD=1 run_test "$CDIR/test_callback.py"
   run_test "$CDIR/test_jax_interop.py"
   run_test "$CDIR/test_assume_pure.py"
+  run_test "$CDIR/test_assume_pure_spmd.py"
 }
 
 # All the new xla op tests should go to run_xla_op_tests3

test/scan/test_scan_spmd.py

@@ -8,7 +8,6 @@
 import torch_xla
 import torch.nn as nn
 from torch_xla.distributed.spmd.xla_sharding import apply_xla_patch_to_nn_linear, Mesh
-from torch_xla.experimental.assume_pure import assume_pure
 from torch_xla.experimental.scan import scan
 from torch_xla.experimental.scan_layers import scan_layers
 from torch_xla.distributed.spmd import mark_sharding, mark_sharding_with_gradients, set_global_mesh, get_1d_mesh, get_global_mesh
@@ -231,27 +230,6 @@ def check_dots_in_model(self, model, x, expect_pattern):
   def count_regex(self, hlo_text, regex_str):
     return len(re.findall(regex_str, hlo_text))
 
-  @unittest.skipIf(
-      torch.cuda.is_available() or os.environ.get('PJRT_DEVICE') == 'CUDA',
-      "TODO(https://github.com/pytorch/xla/issues/9017): Get these tests working on GPU"
-  )
-  def test_assume_pure_works_with_mark_sharding(self):
-    x = torch.randn((3, 4, 5, 128), device='xla')
-    assume_pure(mark_sharding)(x, self.spmd_mesh, ("model", None, None, None))
-    # assert not throwing
-
-  @unittest.skipIf(
-      torch.cuda.is_available() or os.environ.get('PJRT_DEVICE') == 'CUDA',
-      "TODO(https://github.com/pytorch/xla/issues/9017): Get these tests working on GPU"
-  )
-  def test_convert_to_jax_mesh(self):
-    jax_mesh = self.spmd_mesh.maybe_convert_and_get_jax_mesh()
-    self.assertEqual(jax_mesh.devices.shape, self.spmd_mesh.mesh_shape)
-    np.testing.assert_equal(
-        np.array([dev.id for dev in jax_mesh.devices.flatten()]),
-        self.spmd_mesh.device_ids)
-    # assert not throwing
-
 
 if __name__ == '__main__':
   test = unittest.main()

test/test_assume_pure_spmd.py

@@ -0,0 +1,81 @@
+import os
+import sys
+import unittest
+
+import numpy as np
+import torch
+import torch_xla
+import torch_xla.runtime as xr
+from torch_xla.experimental.assume_pure import assume_pure
+from torch_xla.distributed.spmd import mark_sharding, set_global_mesh, get_1d_mesh, Mesh
+
+
+class AssumePureSpmdTest(unittest.TestCase):
+
+  @classmethod
+  def setUpClass(cls):
+    # Activate SPMD
+    xr.use_spmd()
+
+  def setUp(self):
+    # Set up a simple SPMD mesh for these tests.
+    self.spmd_mesh = get_1d_mesh(axis_name="model")
+    set_global_mesh(self.spmd_mesh)
+
+  @unittest.skipUnless(xr.global_runtime_device_count() > 1,
+                       "Multiple devices required")
+  @unittest.skipIf(
+      torch.cuda.is_available() or os.environ.get('PJRT_DEVICE') == 'CUDA',
+      "TODO(https://github.com/pytorch/xla/issues/9017): Get these tests working on GPU"
+  )
+  def test_assume_pure_works_with_mark_sharding(self):
+    x = torch.randn((8, 4, 5, 128), device='xla')
+    result = assume_pure(mark_sharding)(x, self.spmd_mesh,
+                                        ("model", None, None, None))
+    torch_xla.sync(wait=True)
+    N = xr.global_runtime_device_count()
+    self.assertIn(f'devices=[{N}',
+                  torch_xla._XLAC._get_xla_sharding_spec(result))
+
+  @unittest.skipUnless(xr.global_runtime_device_count() > 1,
+                       "Multiple devices required")
+  @unittest.skipIf(
+      torch.cuda.is_available() or os.environ.get('PJRT_DEVICE') == 'CUDA',
+      "TODO(https://github.com/pytorch/xla/issues/9017): Get these tests working on GPU"
+  )
+  def test_convert_to_jax_mesh(self):
+    jax_mesh = self.spmd_mesh.get_jax_mesh()
+    self.assertEqual(jax_mesh.devices.shape, self.spmd_mesh.mesh_shape)
+    np.testing.assert_equal(
+        np.array([dev.id for dev in jax_mesh.devices.flatten()]),
+        self.spmd_mesh.device_ids)
+
+  @unittest.skipUnless(xr.global_runtime_device_count() > 1,
+                       "Multiple devices required")
+  @unittest.skipUnless(os.environ.get('PJRT_DEVICE') == 'TPU', "TPU only test")
+  def test_convert_to_jax_mesh_shuffled(self):
+    """Test get_jax_mesh when the PyTorch/XLA mesh has a custom order."""
+
+    # Arrange
+    num_devices = xr.global_runtime_device_count()
+    device_ids = np.arange(num_devices)
+    device_ids = np.random.permutation(device_ids)
+    self.spmd_mesh = Mesh(
+        device_ids, mesh_shape=(num_devices,), axis_names=('model',))
+
+    # Act
+    jax_mesh = self.spmd_mesh.get_jax_mesh()
+
+    # Assert
+    torch_xla_devices = np.array(
+        [xr.global_runtime_device_attributes()[i] for i in device_ids])
+    self.assertEqual(jax_mesh.devices.shape, self.spmd_mesh.mesh_shape)
+    np.testing.assert_equal(
+        np.array([dev.coords for dev in jax_mesh.devices.flatten()]),
+        np.array([dev['coords'] for dev in torch_xla_devices.flatten()]),
+    )
+
+
+if __name__ == '__main__':
+  test = unittest.main()
+  sys.exit(0 if test.result.wasSuccessful() else 1)

test/tpu/run_tests.sh

@@ -37,6 +37,7 @@ python3 "$TEST_CDIR/scan/test_scan_pallas.py"
 python3 "$TEST_CDIR/scan/test_scan_layers.py"
 python3 "$TEST_CDIR/test_gru.py"
 python3 "$TEST_CDIR/test_assume_pure.py"
+python3 "$TEST_CDIR/test_assume_pure_spmd.py"
 python3 "$TEST_CDIR/test_as_stride_use_slice.py"
 run_xla_hlo_debug python3 "$TEST_CDIR/scan/test_scan_debug.py"
 python3 "$TEST_CDIR/test_pallas.py" -v

torch_xla/distributed/spmd/xla_sharding.py

@@ -182,7 +182,7 @@ def from_str(cls, mesh_str: str) -> Optional["Mesh"]:
       return None
 
   @requires_jax
-  def maybe_convert_and_get_jax_mesh(self):
+  def get_jax_mesh(self):
     # Construct a JAX mesh object with the same device ids shape and ordering
     # from torch_xla device mesh.
     import jax
@@ -611,7 +611,7 @@ def mark_sharding(t: Union[torch.Tensor, XLAShardedTensor], mesh: Mesh,
   if tx is not None and isinstance(t, tx.tensor.Tensor):
     from jax.sharding import PartitionSpec as P, NamedSharding
     op_sharding = tuple(str(i) if i is not None else i for i in partition_spec)
-    jmesh = mesh.maybe_convert_and_get_jax_mesh()
+    jmesh = mesh.get_jax_mesh()
     t.shard_(NamedSharding(jmesh, P(*op_sharding)))
     return t
 

torch_xla/experimental/assume_pure.py

@@ -15,9 +15,12 @@ def assume_pure(fn):
 
   Limitations:
   - The decorated function can only use upstream PyTorch operators e.g.
-    `torch.einsum`, `torch.nn.functional.layer_norm`. Custom PyTorch/XLA
-    operations such as `mark_sharding` are not supported. This limitation
-    may be lifted in the future.
+    `torch.einsum`, `torch.nn.functional.layer_norm`, and a few PyTorch/XLA operators:
+    * `torch_xla.experimental.assume_pure` (recursive `assume_pure`)
+    * `torch_xla.distributed.spmd.mark_sharding`
+
+  - Other custom PyTorch/XLA operations such as `flash_attention` are not
+    supported. This limitation may be lifted in the future.
   """
   from torchax.interop import jax_view
   return j2t_autograd(jax_view(fn))

torch_xla/experimental/splash_attention.py

@@ -86,7 +86,7 @@ def splash_attention_jax_wrapper(
       splash_attention_kernel,
       splash_attention_mask,
   )
-  mesh = Mesh.from_str(config.mesh).maybe_convert_and_get_jax_mesh()
+  mesh = Mesh.from_str(config.mesh).get_jax_mesh()
   # input q,k,v shape: [batch, #head, seq_len, head_dim]
   if decoder_segment_ids is not None and not decoder_segment_ids.shape:
     decoder_segment_ids = None