returnn 1.20240715.155107__tar.gz → 1.20240718.142037__tar.gz

{returnn-1.20240715.155107 → returnn-1.20240718.142037}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20240715.155107
+Version: 1.20240718.142037
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20240718.142037/_setup_info_generated.py

@@ -0,0 +1,2 @@
+version = '1.20240718.142037'
+long_version = '1.20240718.142037+git.92c7ce7'

{returnn-1.20240715.155107 → returnn-1.20240718.142037}/returnn/frontend/array_.py

@@ -558,8 +558,12 @@ def sequence_mask(dims: Union[Dim, Sequence[Dim]], *, device: Optional[str] = No
         dims = [dims]
     assert len(dims) > 0
     dyn_dims = [d for d in dims if d.need_masking()]
-    assert len(dyn_dims) == 1  # not implemented otherwise yet...
-    return dyn_dims[0].get_mask(dim_order=dims, device=device)
+    if not dyn_dims:
+        return rf.constant(True, dims=())
+    mask = True
+    for dim in dyn_dims:
+        mask = rf.opt_logical_and(mask, dim.get_mask(dim_order=dims, device=device))
+    return mask
 
 
 def pack_padded(

{returnn-1.20240715.155107 → returnn-1.20240718.142037}/returnn/frontend/attention.py

@@ -542,7 +542,7 @@ class RelPosCausalSelfAttention(CausalSelfAttention):
         output, _ = rf.merge_dims(att, dims=(self.num_heads, self.value_dim_per_head), out_dim=self.value_dim_total)
         if self.proj:
             output = self.proj(output)
-        return output
+        return output, new_state
 
 
 class CrossAttention(rf.Module):

{returnn-1.20240715.155107 → returnn-1.20240718.142037}/returnn/frontend/math_.py

@@ -36,6 +36,7 @@ __all__ = [
     "logical_or",
     "logical_not",
     "opt_logical_or",
+    "opt_logical_and",
     "maximum",
     "minimum",
     "clip_by_value",
@@ -341,6 +342,24 @@ def opt_logical_or(a: Union[Tensor, bool], b: Union[Tensor, bool]) -> Union[Tens
     return combine(a, "logical_or", b)
 
 
+@overload
+def opt_logical_and(a: bool, b: bool) -> bool:
+    """logical and"""
+
+
+def opt_logical_and(a: Union[Tensor, bool], b: Union[Tensor, bool]) -> Union[Tensor, bool]:
+    """logical and"""
+    if isinstance(a, bool):
+        if not a:
+            return False
+        return b
+    if isinstance(b, bool):
+        if not b:
+            return False
+        return a
+    return combine(a, "logical_and", b)
+
+
 def maximum(a: Tensor, b: Union[Tensor, _RawTensorTypes], *other_tensors) -> Tensor:
     """maximum"""
     if not other_tensors:

{returnn-1.20240715.155107 → returnn-1.20240718.142037}/returnn.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20240715.155107
+Version: 1.20240718.142037
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

{returnn-1.20240715.155107 → returnn-1.20240718.142037}/tests/test_torch_util.py

@@ -13,14 +13,13 @@ import torch
 
 from returnn.util import better_exchook
 
-# noinspection PyProtectedMember
-from returnn.torch.util.gradient_checkpoint import gradient_checkpoint_scope, _can_exit_saved_tensors_hooks_inside_hooks
-
 
+@unittest.skipIf(torch.__version__ < (2,), "gradient_checkpoint_scope needs PyTorch >= 2.0")
 def test_gradient_checkpoint_scope():
     # https://github.com/rwth-i6/returnn/issues/1552
     from copy import deepcopy
     from torch.profiler import profile, record_function, ProfilerActivity
+    from returnn.torch.util.gradient_checkpoint import gradient_checkpoint_scope
 
     shape = (101, 103)
 
@@ -153,8 +152,12 @@ def test_gradient_checkpoint_scope():
         torch.testing.assert_allclose(param_post_state[k], param_post_state_[k])
 
 
+@unittest.skipIf(torch.__version__ < (2,), "gradient_checkpoint_scope needs PyTorch >= 2.0")
 def test_gradient_checkpoint_scope_twice():
     # https://github.com/rwth-i6/returnn/issues/1579
+
+    from returnn.torch.util.gradient_checkpoint import gradient_checkpoint_scope
+
     shape = (101, 103)
 
     class _Model(torch.nn.Module):
@@ -198,9 +201,14 @@ def test_gradient_checkpoint_scope_twice():
         gradient_checkpoint_scope._tensor_del_hook = orig_gradient_checkpoint_scope_tensor_del_hook
 
 
+@unittest.skipIf(torch.__version__ < (2,), "gradient_checkpoint_scope needs PyTorch >= 2.0")
 def test_saved_tensors_hooks_gc_segfault():
     # https://github.com/rwth-i6/returnn/issues/1581
     # https://github.com/pytorch/pytorch/issues/130734
+
+    # noinspection PyProtectedMember
+    from returnn.torch.util.gradient_checkpoint import _can_exit_saved_tensors_hooks_inside_hooks
+
     if not _can_exit_saved_tensors_hooks_inside_hooks():
         raise unittest.SkipTest("Not yet fixed.")
 

returnn-1.20240715.155107/_setup_info_generated.py

@@ -1,2 +0,0 @@
-version = '1.20240715.155107'
-long_version = '1.20240715.155107+git.23d666c'