returnn 1.20250416.172956__py3-none-any.whl → 1.20250418.115249__py3-none-any.whl

returnn/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20250416.172956
+Version: 1.20250418.115249
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn/_setup_info_generated.py

@@ -1,2 +1,2 @@
-version = '1.20250416.172956'
-long_version = '1.20250416.172956+git.505963b'
+version = '1.20250418.115249'
+long_version = '1.20250418.115249+git.3cef372'

returnn/datasets/generating.py

@@ -792,10 +792,10 @@ class DummyDataset(GeneratingDataset):
         i1 = seq_idx
         i2 = i1 + seq_len * self.num_inputs
         features = numpy.array(
-            [((i % self.input_max_value) + self.input_shift) * self.input_scale for i in range(i1, i2)]
+            [((i % self.input_max_value) + self.input_shift) * self.input_scale for i in range(i1, i2)], dtype="float32"
         ).reshape((seq_len, self.num_inputs))
         i1, i2 = i2, i2 + seq_len
-        targets = numpy.array([i % self.num_outputs["classes"][0] for i in range(i1, i2)])
+        targets = numpy.array([i % self.num_outputs["classes"][0] for i in range(i1, i2)], dtype="int32")
         return DatasetSeq(seq_idx=seq_idx, features=features, targets=targets)
 
 

returnn/datasets/postprocessing.py

@@ -141,13 +141,14 @@ class PostprocessingDataset(CachedDataset2):
         self._map_seq_stream = map_seq_stream
         if map_seq_stream_preserves_num_seqs is None and map_seq_stream is not None:
             map_seq_stream_preserves_num_seqs = getattr(map_seq_stream, "preserves_num_seqs", None)
-        self._map_seq_stream_preserves_num_seqs = map_seq_stream_preserves_num_seqs or False
+        assert map_seq_stream_preserves_num_seqs is None or isinstance(map_seq_stream_preserves_num_seqs, bool)
+        self._map_seq_stream_preserves_num_seqs = map_seq_stream_preserves_num_seqs
         self._map_outputs = map_outputs
         self._rng = RandomState(self._get_random_seed_for_epoch(0))
         self._seq_list_for_validation: Optional[List[str]] = None
 
         self._dataset = init_dataset(self._dataset_def, parent_dataset=self)
-        if self._map_seq_stream is None or self._map_seq_stream_preserves_num_seqs:
+        if self._map_seq_stream is None or self._map_seq_stream_preserves_num_seqs is True:
             # if the stream mapper is set, the num_seqs may change and the estimation is less accurate
             self._estimated_num_seqs = self._dataset.estimated_num_seqs
         self._data_iter: Optional[Iterator[Tuple[int, TensorDict]]] = None
@@ -210,7 +211,7 @@ class PostprocessingDataset(CachedDataset2):
         self._data_iter = enumerate(self._build_mapping_iter())
         self._data_iter_produced_num_seqs = 0
         self._seq_list_for_validation = seq_list
-        if self._map_seq_stream is None or self._map_seq_stream_preserves_num_seqs:
+        if self._map_seq_stream is None or self._map_seq_stream_preserves_num_seqs is True:
             # If we don't have an iterable mapper (or the user explicitly specifies this),
             # we know the number of segments exactly equals the number of segments in the wrapped dataset
             try:

returnn/frontend/audio/specaugment.py

@@ -67,7 +67,9 @@ def specaugment(
             )
         return x_masked
 
-    return rf.cond(rf.get_run_ctx().train_flag | (not only_on_train), _mask_branch, lambda: x)
+    return rf.cond(
+        rf.get_run_ctx().is_train_flag_enabled(func=specaugment) | (not only_on_train), _mask_branch, lambda: x
+    )
 
 
 def random_mask(

returnn/frontend/dropout.py

@@ -60,7 +60,7 @@ def dropout(
         return _dropout(source, keep_prob, noise_dims=noise_dims)
 
     return rf.cond(
-        pred=rf.get_run_ctx().train_flag,
+        pred=rf.get_run_ctx().is_train_flag_enabled(func=dropout),
         true_fn=lambda: _dropout(source, keep_prob, noise_dims=noise_dims),
         false_fn=lambda: source,
     )

returnn/frontend/normalization.py

@@ -226,8 +226,9 @@ class BatchNorm(rf.Module):
 
         if use_mask:
             # Generic implementation which supports masking.
-            use_current_batch_stats = self.running_mean is None or rf.get_run_ctx().train_flag
-            update_running_stats = self.running_mean is not None and rf.get_run_ctx().train_flag
+            train_flag = rf.get_run_ctx().is_train_flag_enabled(func=BatchNorm.__call__)
+            use_current_batch_stats = self.running_mean is None or train_flag
+            update_running_stats = self.running_mean is not None and train_flag
             need_current_batch_stats = rf.opt_logical_or(use_current_batch_stats, update_running_stats)
 
             mean_cur_batch, variance_cur_batch = rf.cond(

returnn/frontend/parametrizations.py

@@ -48,7 +48,7 @@ class WeightDropout:
                 # on_forward=True because we already checked for train_flag
                 return rf.dropout(param, drop_prob=self.drop_prob, on_forward=True)
 
-        return rf.cond(rf.get_run_ctx().train_flag, _on_train, lambda: param)
+        return rf.cond(rf.get_run_ctx().is_train_flag_enabled(func=WeightDropout.__call__), _on_train, lambda: param)
 
 
 def weight_noise(module: rf.Module, param_name: str, *, std: float) -> rf.Module:
@@ -84,4 +84,4 @@ class WeightNoise:
                 noise = rf.random_normal(param.dims, dtype=param.dtype, stddev=self.std)
                 return param + noise
 
-        return rf.cond(rf.get_run_ctx().train_flag, _on_train, lambda: param)
+        return rf.cond(rf.get_run_ctx().is_train_flag_enabled(func=WeightNoise.__call__), _on_train, lambda: param)

returnn/frontend/rec.py

@@ -218,7 +218,7 @@ def _zoneout(*, prev: Tensor, cur: Tensor, factor: float, out_dim: Dim, dropout_
     if factor == 0.0:
         return cur
     return rf.cond(
-        rf.get_run_ctx().train_flag,
+        rf.get_run_ctx().is_train_flag_enabled(func=ZoneoutLSTM.__call__),
         lambda: (1 - factor) * rf.dropout(cur - prev, factor, axis=dropout_broadcast and out_dim) + prev,
         lambda: (1 - factor) * cur + factor * prev,
     )

returnn/frontend/reduce.py

@@ -251,7 +251,11 @@ class RunningMean(rf.Module):
             x_ = rf.reduce_mean(x, axis=[d for d in x.dims if d not in self.shape])
             self.mean.assign_add(self.alpha * (x_ - self.mean))
 
-        rf.cond((not self.update_only_in_train) or rf.get_run_ctx().train_flag, _update_running_stats, lambda: None)
+        rf.cond(
+            (not self.update_only_in_train) or rf.get_run_ctx().is_train_flag_enabled(func=RunningMean.__call__),
+            _update_running_stats,
+            lambda: None,
+        )
         return self.mean
 
 

returnn/frontend/run_ctx.py

@@ -7,7 +7,8 @@ or forwarding loop.
 """
 
 from __future__ import annotations
-from typing import Optional, Union, Any, Sequence, Dict
+from typing import Optional, Union, Any, Callable, Sequence, Dict, List
+from types import FunctionType
 from dataclasses import dataclass
 from contextlib import contextmanager
 from returnn.tensor import Tensor, Dim, TensorDict, batch_dim
@@ -101,7 +102,7 @@ class RunCtx:
             - "forward_step", for mark_as_output
         """
         self._stage = stage
-        self._train_flag = train_flag
+        self._train_flags_stack: List[Dict[Optional[FunctionType], Union[Tensor, bool]]] = [{None: train_flag}]
         self._step = step
         self._epoch = epoch
         self.losses = {}  # type: Dict[str, Loss]
@@ -121,14 +122,17 @@ class RunCtx:
     @property
     def train_flag(self) -> Union[bool, Tensor]:
         """
-        :return: whether we are in training mode, i.e. the model is updated,
-            and we are supposed to use dropout and similar mechanisms.
-            In a graph-based backend, this can be dynamic.
+        :return: ``is_train_flag_enabled(func=None)``. See :func:`is_train_flag_enabled`.
         """
-        return self._train_flag
+        return self.is_train_flag_enabled(func=None)
 
     @contextmanager
-    def train_flag_ctx(self, train_flag: Union[bool, Tensor]):
+    def train_flag_ctx(
+        self,
+        train_flag: Union[bool, Tensor],
+        *,
+        func: Optional[Union[Sequence[Union[FunctionType, Callable]], FunctionType, Callable]] = None,
+    ):
         """
         Context manager to temporarily set the train_flag.
 
@@ -137,14 +141,51 @@ class RunCtx:
             with rf.get_run_ctx().train_flag_ctx(False):
                 ...
 
-        :param train_flag: whether we are in training mode
-        """
-        old_train_flag = self.train_flag
-        self._train_flag = train_flag
+        :param train_flag: whether we are in training mode.
+            In a graph-based backend, this can be dynamic (scalar Tensor, not just bool).
+        :param func: if given, the train flag is only enabled/disabled for this specific function(s)
+            (e.g. ``rf.dropout`` or ``rf.BatchNorm.__call__``).
+            (See https://github.com/rwth-i6/returnn/issues/1712 for some discussion.)
+            (Note: We expect a Python function, not just any general Callable. But typing seems to get this wrong.)
+        """
+        old_train_flags = self._train_flags_stack[-1]
+        new_train_flags = old_train_flags.copy()
+        if func is None:
+            new_train_flags[None] = train_flag
+        elif isinstance(func, FunctionType):
+            new_train_flags[func] = train_flag
+        elif isinstance(func, (list, tuple)):
+            for f in func:
+                if not isinstance(f, FunctionType):
+                    raise TypeError(f"Expected function, got {type(f)}")
+                new_train_flags[f] = train_flag
+        else:
+            raise TypeError(f"Expected function or sequence of functions, got {type(func)}")
+        self._train_flags_stack.append(new_train_flags)
         try:
             yield
         finally:
-            self._train_flag = old_train_flag
+            last = self._train_flags_stack.pop(-1)
+            assert last is new_train_flags
+            assert len(self._train_flags_stack) >= 1
+
+    def is_train_flag_enabled(self, *, func: Optional[Union[FunctionType, Callable]]) -> Union[bool, Tensor]:
+        """
+        :param func: function for which we want to check the train flag
+            (e.g. ``rf.dropout`` or ``rf.BatchNorm.__call__``),
+            or None for the global fallback.
+            (See https://github.com/rwth-i6/returnn/issues/1712 for some discussion.)
+        :return: Whether the train flag is enabled, either for the specific function, or globally.
+            Training is usually when the model is updated,
+            and we are supposed to use dropout and similar mechanisms.
+            This is either for the specified function, or globally.
+            In a graph-based backend, this can also be dynamic (scalar Tensor, not just bool).
+        """
+        train_flags = self._train_flags_stack[-1]
+        if func in train_flags:
+            return train_flags[func]
+        assert isinstance(func, FunctionType)
+        return train_flags[None]  # global fallback. this should always be defined, see __init__
 
     @property
     def step(self) -> Union[int, Tensor]:

returnn/torch/frontend/_backend.py

@@ -1859,7 +1859,7 @@ class TorchBackend(Backend[torch.Tensor]):
             bias=beta.raw_tensor if affine else None,
             # training: means whether we should use the current batch statistics
             #   + update the running statistics (if given)
-            training=rf.get_run_ctx().train_flag or (running_mean is None),
+            training=rf.get_run_ctx().is_train_flag_enabled(func=rf.BatchNorm.__call__) or (running_mean is None),
             momentum=momentum,
             eps=epsilon,
         )
@@ -2236,7 +2236,7 @@ class TorchBackend(Backend[torch.Tensor]):
             has_biases=has_biases,
             num_layers=1,
             dropout=0.0,
-            train=rf.get_run_ctx().train_flag,
+            train=rf.get_run_ctx().is_train_flag_enabled(func=rf.LSTM.__call__),
             bidirectional=False,
         )
 

{returnn-1.20250416.172956.dist-info → returnn-1.20250418.115249.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20250416.172956
+Version: 1.20250418.115249
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

{returnn-1.20250416.172956.dist-info → returnn-1.20250418.115249.dist-info}/RECORD

@@ -1,9 +1,9 @@
-returnn/PKG-INFO,sha256=y5t_H0i0pLYw-W1KevNnfXbSxkrkqPT2uFYpKMGhS3o,5215
+returnn/PKG-INFO,sha256=vWxYPYgsvU3GDUCQEiHY5h2C09lxsFvZ-GUR6jfYUxM,5215
 returnn/__init__.py,sha256=biBtRsM0WZ406vShaeH-9WFoqJ8XwTbn6g0EeFJ7l8E,1012
 returnn/__main__.py,sha256=qBFbuB1yN3adgVM5pXt2-Yq9vorjRNchNPL8kDKx44M,31752
 returnn/__old_mod_loader__.py,sha256=nvsNY-xELdS_IPNkv66Q9Rmvg4dbGW0-EBRDcCmctos,7654
 returnn/__setup__.py,sha256=22kQn2fh11iPM0hLb2Fy5sLmoU1JGvmDxXRYuRgQkwU,4659
-returnn/_setup_info_generated.py,sha256=AcoND8SAX3c8L8MOr-ar4L8QEmAQF2emtFz2sd2BL5I,77
+returnn/_setup_info_generated.py,sha256=oEWWbm7JJwkhmRryz4mvP5IIrbCPVx8zczj4izvepn0,77
 returnn/config.py,sha256=3tmKhB6FnQZaNdtcYsiB61JnEY--iZ2qmJ4yq0b6tE0,29140
 returnn/forward_iface.py,sha256=A_OJiaXsX4MlXQRzST86ylyxSUZbC402PQL1REcqHjM,911
 returnn/learning_rate_control.py,sha256=ZvWryAn_tv9DhV8sh1LV3eE34Yltl3On3mYZAG4hR9s,34684
@@ -18,7 +18,7 @@ returnn/datasets/bundle_file.py,sha256=KQNrS1MSf-4_idlK0c0KFwON-f5sEK0sWU15WpoMY
 returnn/datasets/cached.py,sha256=DIRdWrxBmsZG8O_9eVxBO5mcdo4f5KU-Xb-4wVz59Io,25418
 returnn/datasets/cached2.py,sha256=_6pza3IG68JexaExhj1ld3fP6pE7T-G804driJ9Z_qo,12141
 returnn/datasets/distrib_files.py,sha256=_UlcrnaU1rA9v6D3H3X4dPhcA--09fNeVnWs9VNo0yg,27656
-returnn/datasets/generating.py,sha256=O1fs9dhk1Um2E3ZeOTfDHS5FlwvqFImfGcMlJP-xAQM,99814
+returnn/datasets/generating.py,sha256=E_6KpnSu8ChqG3pb4VTChWDsBTonIwFFAj53SI9NSow,99846
 returnn/datasets/hdf.py,sha256=yqzr-nzqlt02QZoW2uFowKT19gd5e-9mJpHCKSQxW8o,67643
 returnn/datasets/lm.py,sha256=5hSdBgmgTP0IzO2p-JjiWtny0Zb0M20goXtjlw4JVR4,99206
 returnn/datasets/map.py,sha256=kOBJVZmwDhLsOplzDNByIfa0NRSUaMo2Lsy36lBvxrM,10907
@@ -26,7 +26,7 @@ returnn/datasets/meta.py,sha256=0wQzRzjShLSYNFoGo_MdR5IT8arxHr9gFjUlEqb2rbY,9496
 returnn/datasets/multi_proc.py,sha256=aVjsLt2qjHnHOrEYCgIPCwNYE-f1fiGP6eZ8NGAr3A4,22583
 returnn/datasets/normalization_data.py,sha256=wOHrbO3612uWXpzLHHxksDw0qeVmQ42w7byBL9QMh9Q,14618
 returnn/datasets/numpy_dump.py,sha256=wl8bKIKAlff2HPJPtuu5wBg3TLOf16d2wLVB4lLAwTM,5158
-returnn/datasets/postprocessing.py,sha256=dXlRiZNuC8921RZ3lkgX1Lt88K61lKsG47nVIH5x7nI,23395
+returnn/datasets/postprocessing.py,sha256=Ug2fvzbutnwJSvErIK2Ft-bd0pz79ZBmEg12pLEo1f0,23514
 returnn/datasets/raw_wav.py,sha256=M7eTHp4CTtLQf3yPTiJY-mSJYgZNxkGV9IFN9J1dq_4,9144
 returnn/datasets/sprint.py,sha256=YhhdNbBTuL_HCc3asgK3o6vgq5h5nMPH5nBFvsuwVjA,55464
 returnn/datasets/stereo.py,sha256=PkowC91bZWihIYuIZgyGgPcNwgq5jBvyxxu1nER-VhM,17633
@@ -91,7 +91,7 @@ returnn/frontend/control_flow_ctx.py,sha256=v17CsNwRnZYe8GdMtGJt2ftibfxMCGK1i0l-
 returnn/frontend/conv.py,sha256=Q0q90-uu9d6qV-v8_DlFGxpZtc6FjfXVpfkkXmv1Alk,31959
 returnn/frontend/device.py,sha256=K7Y1qoQcO4GIHgLkPLQWK-GVT8gKL8GwyQrmPo8LgBE,1438
 returnn/frontend/dims.py,sha256=aH5FQ_m0xMD6Rj-BUWGx8lB-HkCuwZfMBf6mZbGGW5E,12611
-returnn/frontend/dropout.py,sha256=rsx3p5b0NblBfXXSQZTQFJ8jUUS3fj4Qzc39iffBMCA,5006
+returnn/frontend/dropout.py,sha256=bH0keqKcBzkC_SPlQoir9HPxTtgoVCp61YD5ZvEOuA4,5031
 returnn/frontend/dtype.py,sha256=Ooc5BrcNrTp6XShuFEV9g5V6-niuy4ImP_Lt_Qgq3jE,1886
 returnn/frontend/gradient.py,sha256=G-Qv4gKGHYEeB92Zwco9ao4qjd6umZPUzQC4J-fbYWo,4033
 returnn/frontend/graph.py,sha256=PIv901WZ1rfTV0QGkyzBv6UxfWk9NsLGxdoJ5x9-8Xg,1818
@@ -105,15 +105,15 @@ returnn/frontend/math_.py,sha256=KlJxdIib8ENlid7cc4lcwHv5e21tzTjTEV8VgEDAijo,169
 returnn/frontend/matmul.py,sha256=3QaGiZtSs9PriT40T7Vc3KnYKPgYSN4tCZytYeq9qMA,1945
 returnn/frontend/module.py,sha256=219rh5mE0CD0-NdxXLsKyhv3BNtOI9jSyiI1Rb8MOyU,10700
 returnn/frontend/nested.py,sha256=Hm4GT5ZI1OyWpYxv_SP5jlBztJsjGVMgtvKJnvQYa00,15068
-returnn/frontend/normalization.py,sha256=QIjXYg0C8BD2g_1lAkVO4Cara729uHC_bsQh99VsWeI,14061
+returnn/frontend/normalization.py,sha256=-lYJ9IWcheOQu1gXJehSOA76qgVtxd1C07Jqps6Qg1o,14116
 returnn/frontend/parameter.py,sha256=w6SN-uv87OyeWBt90_3UBbK0h6sftSOCxkqXPg76caY,10375
-returnn/frontend/parametrizations.py,sha256=hVbOlgm1pQAmZnAnNxq8Tk23rykr_iy3-6R1H6CwlMA,2798
+returnn/frontend/parametrizations.py,sha256=ptNgBw5IiPXVpB3QGse7AGAhdXp8X1rCqYUl2Mae8aI,2876
 returnn/frontend/parametrize.py,sha256=VhgTEP7ehON950Q4bkCy8rvg9641moEKAXn0XzomK6E,7216
 returnn/frontend/piecewise_linear.py,sha256=TdL6wzop8P1dcIZwkEbJFvSUZSI1cbhS3XKzlWQkEVI,1964
 returnn/frontend/rand.py,sha256=Levgf5VtOOBKDSgz0869Jf3VW4BWxYZuRXsa_fOxNI4,12969
-returnn/frontend/rec.py,sha256=la-VXR_hzvwNzpAgn4Okl-yDx3F4gOW-81EKm-jAAlg,7999
-returnn/frontend/reduce.py,sha256=-Zt-OH6Zbtb9uR6YEzurCyrowH-anIXvuga6Pla2V70,10220
-returnn/frontend/run_ctx.py,sha256=ItcZwuFItkZjYWrg715L1Za2Xg7__MQCrRCAwBeTUxA,21411
+returnn/frontend/rec.py,sha256=6YSsSG7fdtfvvg24vmexSg8R2aVCcKHBdGLh-Mgn9Co,8037
+returnn/frontend/reduce.py,sha256=xvxN_h3LsMJdmT0IbW4nOf8qFhckuAniIhD9PalO6j0,10305
+returnn/frontend/run_ctx.py,sha256=WxKmBpkb2b34AQbnQ9YIRYvNoy3Fa0yJIezNCWLpeWU,23785
 returnn/frontend/signal.py,sha256=XgOBL1iy-cJgulePH5HRPAwp2cScy60q4RItr7xzvGc,4412
 returnn/frontend/state.py,sha256=EePdrx6PtWL4mJ2XZmGlh5dl4nq6G9wZpqP4hdDEzfY,2935
 returnn/frontend/stepwise_scheduler.py,sha256=fMOTR7npGCDXrXDmSQ4VwmudoHEbY3Yr-QGyjFdQJSc,927
@@ -129,7 +129,7 @@ returnn/frontend/_native/tensor_ops.cpp,sha256=bYtwwn_NeJfAEHWYPEJlkoLDKt9baZ3RA
 returnn/frontend/_native/tensor_ops.hpp,sha256=dDqvUejRNHjItnmOP5aHyAQbAmXmXoDVXSe3tveEU8A,3732
 returnn/frontend/audio/__init__.py,sha256=8mahwucBje8qHKw0bOvoySlvvD0rFKxviSvcAHSjiJY,67
 returnn/frontend/audio/mel.py,sha256=VZdxf2mTLzLOXsLRzCvaad712Zf0c2iwdthrzeVfgxk,7885
-returnn/frontend/audio/specaugment.py,sha256=nw8PepKPPwmI13-QyGBm45QjnoY4I_FEjA-_X6KIwzM,5798
+returnn/frontend/audio/specaugment.py,sha256=w7YPEJ6zhCaG5AAaDd-HxsKwa_2vA7wFqHrEjxiUVPI,5841
 returnn/frontend/conversions/__init__.py,sha256=7plsDxWVYhASa-3qmqbdzSI34A9ujUH2iMkL3eRD0TI,84
 returnn/frontend/conversions/espnet_e_branchformer.py,sha256=Mmp3G6nySy0CqeHa-um-RAuUSnFH1DKNjBbqQB_Pomo,9018
 returnn/frontend/conversions/hf_llama.py,sha256=1WQOhQyUWwkAznaRqK2zpThP8XZbaomkaE8qMG_bZPY,9662
@@ -216,7 +216,7 @@ returnn/torch/data/queued_data_iter.py,sha256=PoOsGHdHVZjTmcyfq_ZOw--P6hyfTdmAWI
 returnn/torch/data/returnn_dataset_wrapper.py,sha256=2CaDapzrlqahANuq-nyVAtv5ENHuM8A7okORwYJDisg,8006
 returnn/torch/data/tensor_utils.py,sha256=-Teqi--LLbt6q_5mDRdoHZHmPgSdC83W706ukif_YiU,1284
 returnn/torch/frontend/__init__.py,sha256=AA48HZnC17ASuKA0EWy8loZ-Bib_yUtqF4T1wYvjst4,62
-returnn/torch/frontend/_backend.py,sha256=SKxxpIM0rXEcZ92p-Um5thfC7vmDoZmda13SMAXVYL0,101771
+returnn/torch/frontend/_backend.py,sha256=xX3xGe-TtGy_VnTyQyhDD3lfZ55gZajgxE9khOyQ4HE,101844
 returnn/torch/frontend/_rand.py,sha256=1JgIkV2XmpgJD86zXZ-NCAe-QuoP2swr6NaS1oz3Qa8,1830
 returnn/torch/frontend/bridge.py,sha256=Z2_UW8AagezC7zsXDc5PKcd8G9WwisV7j9SWGHU0m4U,7840
 returnn/torch/frontend/raw_ops.py,sha256=lF0h-KtYYsdaaqQADylVZp9qzPskOOXA4MfmYDyx5IU,296
@@ -253,8 +253,8 @@ returnn/util/sig_proc.py,sha256=Tjz0VOAVyqu2qDCF5HZ1JjALjcFsHcNkcd96WgZeKfE,7265
 returnn/util/task_system.py,sha256=y4sMVXQ25Qd2z0rx03uOlXlkE-jbCYC1Sjfn-XlraVU,26003
 returnn/util/train_proc_manager.py,sha256=Pjht28k6uz6BNQ47uW6Gf880iyq5q4wx7P_K2tmoAM8,3266
 returnn/util/watch_memory.py,sha256=BR5P2kvBN6UI81cE0_1WAA6Hd1SByLbBaiDxvLhPOew,4213
-returnn-1.20250416.172956.dist-info/LICENSE,sha256=ywBD_U2aD4vpuoIgNAsjIGBYydl0tVKll3De0Z8s77c,11041
-returnn-1.20250416.172956.dist-info/METADATA,sha256=y5t_H0i0pLYw-W1KevNnfXbSxkrkqPT2uFYpKMGhS3o,5215
-returnn-1.20250416.172956.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
-returnn-1.20250416.172956.dist-info/top_level.txt,sha256=Lsn4WZc5Pbfk0-xDQOgnFCxOoqxL4CyeM3N1TFbJncw,8
-returnn-1.20250416.172956.dist-info/RECORD,,
+returnn-1.20250418.115249.dist-info/LICENSE,sha256=ywBD_U2aD4vpuoIgNAsjIGBYydl0tVKll3De0Z8s77c,11041
+returnn-1.20250418.115249.dist-info/METADATA,sha256=vWxYPYgsvU3GDUCQEiHY5h2C09lxsFvZ-GUR6jfYUxM,5215
+returnn-1.20250418.115249.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
+returnn-1.20250418.115249.dist-info/top_level.txt,sha256=Lsn4WZc5Pbfk0-xDQOgnFCxOoqxL4CyeM3N1TFbJncw,8
+returnn-1.20250418.115249.dist-info/RECORD,,