returnn 1.20250902.114352__tar.gz → 1.20250903.215851__tar.gz

{returnn-1.20250902.114352/returnn.egg-info → returnn-1.20250903.215851}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20250902.114352
+Version: 1.20250903.215851
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20250903.215851/_setup_info_generated.py

@@ -0,0 +1,2 @@
+version = '1.20250903.215851'
+long_version = '1.20250903.215851+git.7651133'

{returnn-1.20250902.114352 → returnn-1.20250903.215851}/returnn/frontend/array_.py

@@ -1195,7 +1195,10 @@ def reverse_sequence(tensor: Tensor, *, axis: Dim, handle_dynamic_dims: bool = T
     if not handle_dynamic_dims or not axis.need_masking():
         # noinspection PyProtectedMember
         return tensor._raw_backend.flip_no_mask(tensor, axis=axis)
-    indices = rf.combine_bc(axis.get_size_tensor(), "-", rf.range_over_dim(axis)) - 1
+    indices = (
+        rf.combine_bc(axis.get_size_tensor(device=tensor.device), "-", rf.range_over_dim(axis, device=tensor.device))
+        - 1
+    )
     return rf.gather(tensor, indices=indices, axis=axis, clip_to_valid=True)
 
 

{returnn-1.20250902.114352 → returnn-1.20250903.215851}/returnn/tensor/_dim_extra.py

@@ -2184,7 +2184,7 @@ class _DimMixin:
             other = other.dimension  # makes matching easier
         if isinstance(other, int) and other == 1:
             return self
-        if self.is_constant_static_dim() and isinstance(other, _d.Dim):
+        if self.is_constant_static_dim() and isinstance(other, _d.Dim) and not other.is_constant_static_dim():
             return self.dimension * other  # use rmul
         cache_key = ("mul", other)
         cache = self.get_same_base()._make_extra().cache_dim_math
@@ -2571,14 +2571,19 @@ class _MathFindMatchingAdditive:
 
 
 def _math_find_matching_mult(start: Dim, other: Union[int, Dim], *, right: bool) -> Optional[Dim]:
-    if (isinstance(other, int) or other.is_constant_static_dim()) and start.is_constant_static_dim():
+    # we assume, if other is Dim, then it is not constant static dim
+    if isinstance(other, int) and start.is_constant_static_dim():
         return _math_get_dim_via_bin_op([start, other] if right else [other, start], "mul")
     c_op = start.derived_from_op
     if c_op and c_op.kind == "mul" and len(c_op.inputs) == 2:
         if right:
             return c_op.inputs[0] * (c_op.inputs[1] * other)
-        else:
-            return (other * c_op.inputs[0]) * c_op.inputs[1]
+        # Don't do right=False -> (other * c_op.inputs[0]) * c_op.inputs[1],
+        # because this can lead to infinite recursions,
+        # and also we don't have a proper normalized form for multiplication.
+        # However, if both left-most factors are constant static dims, then we can merge it.
+        elif isinstance(other, int) and c_op.inputs[0].is_constant_static_dim():
+            return (other * c_op.inputs[0].dimension) * c_op.inputs[1]
     return None
 
 

{returnn-1.20250902.114352 → returnn-1.20250903.215851/returnn.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20250902.114352
+Version: 1.20250903.215851
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

{returnn-1.20250902.114352 → returnn-1.20250903.215851}/tests/test_rf_encoder_conformer.py

@@ -6,6 +6,7 @@ from __future__ import annotations
 import _setup_test_env  # noqa
 from typing import Sequence
 from returnn.tensor import Tensor, Dim, TensorDict, batch_dim
+import returnn.frontend as rf
 from rf_utils import run_model
 
 
@@ -315,8 +316,6 @@ def test_e_branchformer():
     def _tensor(x: torch.Tensor, name: str, dims: Sequence[Dim]) -> Tensor:
         return rf.convert_to_tensor(x, name=name, dims=dims)
 
-    from returnn.frontend.conversions.espnet_e_branchformer import _reorder_rel_pos_emb_espnet_to_rf
-
     check_py_traces_rf_to_pt_equal(
         trace_rf.captured_locals,
         trace_espnet.captured_locals,
@@ -407,3 +406,31 @@ def test_e_branchformer():
     assert enc_seq_lens_raw.max() > 0
     assert torch.mean(enc_out.raw_tensor**2) > 0.1
     print("All matching!")
+
+
+def test_e_branchformer_custom_ff():
+    rf.select_backend_torch()
+
+    from returnn.frontend.encoder.conformer import ConformerEncoder
+    from returnn.frontend.encoder.e_branchformer import EBranchformerLayer
+
+    encoder = ConformerEncoder(
+        in_dim=Dim(7, name="in"),
+        input_layer=None,
+        num_layers=2,
+        out_dim=16,
+        encoder_layer=rf.build_dict(
+            EBranchformerLayer,
+            ff=rf.build_dict(
+                rf.encoder.conformer.ConformerPositionwiseFeedForward,
+                activation=rf.build_dict(rf.relu_square),
+                with_bias=False,
+            ),
+            num_heads=8,
+        ),
+    )
+    assert encoder.layers[0].cgmlp.linear_ff.out_dim == 2 * encoder.out_dim * 3
+    time_dim = Dim(11, name="time")
+    x = rf.random_normal((time_dim, encoder.in_dim))
+    y, _ = encoder(x, in_spatial_dim=time_dim)
+    y.verify_out_shape({time_dim, encoder.out_dim})

{returnn-1.20250902.114352 → returnn-1.20250903.215851}/tests/test_tensor.py

@@ -48,6 +48,15 @@ def test_dim_math_double_neg():
     assert --a == a
 
 
+def test_dim_math_mul_rmul():
+    a = Dim(None, name="a")
+    b = a * 3
+    c = 2 * b
+    assert c == 2 * (a * 3)
+    assert c != a * 6
+    assert c == (2 * a) * 3
+
+
 def test_dim_math_mul_div():
     a = Dim(None, name="a")
     b = Dim(None, name="b")

returnn-1.20250902.114352/_setup_info_generated.py

@@ -1,2 +0,0 @@
-version = '1.20250902.114352'
-long_version = '1.20250902.114352+git.87030fa'