returnn 1.20250817.33823__py3-none-any.whl → 1.20250820.123936__py3-none-any.whl

returnn/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20250817.33823
+Version: 1.20250820.123936
 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 = '1.20250817.033823'
-long_version = '1.20250817.033823+git.4f41ff3'
+version = '1.20250820.123936'
+long_version = '1.20250820.123936+git.9c74169'

returnn/torch/frontend/_backend.py

@@ -993,7 +993,10 @@ class TorchBackend(Backend[torch.Tensor]):
         if clip_to_valid:
             if axis.dyn_size_ext is not None:
                 indices = rf.clip_by_value(
-                    indices, 0, axis.get_dyn_size_ext_for_device(indices.device) - 1, allow_broadcast_all_sources=True
+                    indices,
+                    0,
+                    rf.cast(axis.get_dyn_size_ext_for_device(indices.device), indices.dtype) - 1,
+                    allow_broadcast_all_sources=True,
                 )
             else:
                 indices = indices.copy()

returnn/torch/updater.py

@@ -5,11 +5,10 @@ and model param update logic in general.
 
 from __future__ import annotations
 
-from typing import Optional, Union, Any, Type, Callable, Sequence, Iterable, Iterator, Set, Dict, List, Tuple
+from typing import Optional, Union, Any, Type, Callable, Sequence, Iterable, Set, Dict, List, Tuple
 import os
 import gc
 import torch
-import typing
 
 import returnn
 from returnn.log import log
@@ -130,8 +129,8 @@ class Updater:
             else:
                 raise NotImplementedError("not implemented for not callable dynamic_learning_rate")
 
-        self._optimizer_opts = None
-        self.optimizer = None  # type: typing.Optional[torch.optim.Optimizer]
+        self._optimizer_opts: Optional[Dict[str, Any]] = None
+        self.optimizer: Optional[torch.optim.Optimizer] = None
 
         self._grad_clip = self.config.float("gradient_clip", 0.0)
         self._grad_clip_global_norm = self.config.float("gradient_clip_global_norm", 0.0)
@@ -481,7 +480,7 @@ class Updater:
 
     def _get_optimizer_param_groups(
         self, optim_class: Type[torch.optim.Optimizer], optimizer_opts: Dict[str, Any]
-    ) -> Union[List[Dict[str, Any]], Iterator[torch.nn.Parameter]]:
+    ) -> Union[Iterable[Dict[str, Any]], Iterable[torch.nn.Parameter]]:
         """
         The weight_decay parameter from AdamW affects the weights of layers such as LayerNorm and Embedding.
         This function creates a blacklist of network modules and splits the optimizer groups in two:
@@ -514,10 +513,13 @@ class Updater:
         if custom_param_groups is not None:
             assert callable(custom_param_groups), f"invalid param_groups_custom {custom_param_groups!r}"
             rf_model = wrapped_pt_module_to_rf_module(self.network)
-            custom_param_groups = custom_param_groups(
+            custom_param_groups_ = custom_param_groups(
                 model=self.network, rf_model=rf_model, optimizer_class=optim_class, optimizer_opts=optimizer_opts
             )
-            return custom_param_groups
+            assert isinstance(custom_param_groups_, Iterable) and all(
+                isinstance(group, dict) for group in custom_param_groups_
+            ), f"invalid param_groups_custom {custom_param_groups!r} result {custom_param_groups_!r} type"
+            return custom_param_groups_
 
         network_params = self.network.parameters()
 
@@ -545,7 +547,7 @@ class Updater:
         # Parameters without weight decay: biases + LayerNorm/Embedding layers.
         wd_params = set()
         no_wd_params = set()
-        blacklist_wd_modules = optimizer_opts.pop("weight_decay_modules_blacklist", None)
+        blacklist_wd_modules: Any = optimizer_opts.pop("weight_decay_modules_blacklist", None)
         if blacklist_wd_modules is None:
             blacklist_wd_modules = (torch.nn.LayerNorm, torch.nn.Embedding)
         else:

{returnn-1.20250817.33823.dist-info → returnn-1.20250820.123936.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20250817.33823
+Version: 1.20250820.123936
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

{returnn-1.20250817.33823.dist-info → returnn-1.20250820.123936.dist-info}/RECORD

@@ -1,9 +1,9 @@
-returnn/PKG-INFO,sha256=B57ObmyaxFzp_TjsXhRqtVvzxZWq3hS54vw34tZHZsQ,5214
+returnn/PKG-INFO,sha256=04mKkkm6MNJQzWwBZq5enBZVM1vpvfr-0W705kv7vsk,5215
 returnn/__init__.py,sha256=biBtRsM0WZ406vShaeH-9WFoqJ8XwTbn6g0EeFJ7l8E,1012
 returnn/__main__.py,sha256=lHyZcu_0yc9f7Vf_Kfdy9PmeU0T76XVXnpalHi5WKro,31740
 returnn/__old_mod_loader__.py,sha256=nvsNY-xELdS_IPNkv66Q9Rmvg4dbGW0-EBRDcCmctos,7654
 returnn/__setup__.py,sha256=22kQn2fh11iPM0hLb2Fy5sLmoU1JGvmDxXRYuRgQkwU,4659
-returnn/_setup_info_generated.py,sha256=BSqlencj66qyQkfuBO7fze_8cqW9p9M0Si-qLIGy1-k,77
+returnn/_setup_info_generated.py,sha256=UEqWo5ULLq_sbbrAc2dmq5fjH29qlfAXGYB_hw8vpZQ,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
@@ -208,7 +208,7 @@ returnn/torch/README.md,sha256=jzJ2FpOHW02vxN69yKaV97C9LI-hmvjBglKfdZXIDdc,85
 returnn/torch/__init__.py,sha256=MHEUyNHB20Vy89uKAqZoj6FxJKF1Gq3HW-i6ra1pNcI,24
 returnn/torch/distributed.py,sha256=_lyJR71HIoCHpMi5GztGM7YwrX54Am8zSkjnDkE1Lbk,7524
 returnn/torch/engine.py,sha256=JSsQZZiVs9TxRyFEJuR3iH-YZb9sRw7TzoIAIqmplZY,78275
-returnn/torch/updater.py,sha256=Vyh5w6ZFVc1hQvyyoWpeienQdlBVLZ2HYfjFZRQB3cQ,30035
+returnn/torch/updater.py,sha256=9Ju_LuEdXO7LMxt9rs9_6ReePG5y1h36N3coN696rVI,30285
 returnn/torch/data/__init__.py,sha256=6cLNEi8KoGI12PF6akN7mI_mtjlx-0hcQAfMYoExwik,132
 returnn/torch/data/extern_data.py,sha256=5al706ZaYtHWLp5VH2vS-rW69YXP3NHyOFRKY0WY714,7810
 returnn/torch/data/pipeline.py,sha256=HgIL0jQsPcgvh_SPC4wQ6BzclmrnpFja-UiboF_GPN4,29459
@@ -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=zzKN4_NJK3_I7Ehk8VlhhaXQ_jUEx8K73br8C0Q41p0,103081
+returnn/torch/frontend/_backend.py,sha256=1o6v9neXLTGVu_53QmoPn_2DbbuBC-iyojL9qe5DYBQ,103166
 returnn/torch/frontend/_rand.py,sha256=1JgIkV2XmpgJD86zXZ-NCAe-QuoP2swr6NaS1oz3Qa8,1830
 returnn/torch/frontend/bridge.py,sha256=c_mVBCBo29sjm8Bhxarv00szwGPgxjwoIqAHOmceGQw,7842
 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-1.20250817.33823.dist-info/LICENSE,sha256=ywBD_U2aD4vpuoIgNAsjIGBYydl0tVKll3De0Z8s77c,11041
-returnn-1.20250817.33823.dist-info/METADATA,sha256=B57ObmyaxFzp_TjsXhRqtVvzxZWq3hS54vw34tZHZsQ,5214
-returnn-1.20250817.33823.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
-returnn-1.20250817.33823.dist-info/top_level.txt,sha256=Lsn4WZc5Pbfk0-xDQOgnFCxOoqxL4CyeM3N1TFbJncw,8
-returnn-1.20250817.33823.dist-info/RECORD,,
+returnn-1.20250820.123936.dist-info/LICENSE,sha256=ywBD_U2aD4vpuoIgNAsjIGBYydl0tVKll3De0Z8s77c,11041
+returnn-1.20250820.123936.dist-info/METADATA,sha256=04mKkkm6MNJQzWwBZq5enBZVM1vpvfr-0W705kv7vsk,5215
+returnn-1.20250820.123936.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
+returnn-1.20250820.123936.dist-info/top_level.txt,sha256=Lsn4WZc5Pbfk0-xDQOgnFCxOoqxL4CyeM3N1TFbJncw,8
+returnn-1.20250820.123936.dist-info/RECORD,,