returnn 1.20250719.212120__tar.gz → 1.20250725.163756__tar.gz

{returnn-1.20250719.212120/returnn.egg-info → returnn-1.20250725.163756}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20250719.212120
+Version: 1.20250725.163756
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20250725.163756/_setup_info_generated.py

@@ -0,0 +1,2 @@
+version = '1.20250725.163756'
+long_version = '1.20250725.163756+git.9859629'

{returnn-1.20250719.212120 → returnn-1.20250725.163756}/returnn/datasets/basic.py

@@ -19,7 +19,6 @@ import os
 import math
 import numpy
 import functools
-import typing
 from typing import TYPE_CHECKING, Optional, Any, Set, Tuple, Union, Type, Dict, Sequence, List, Callable
 
 from returnn.log import log
@@ -428,9 +427,9 @@ class Dataset:
         """
         raise OptionalNotImplementedError
 
-    def get_num_timesteps(self):
+    def get_num_timesteps(self) -> Union[int, NumbersDict]:
         """
-        :rtype: int
+        :return: how much frames we have in total.
         """
         assert self._num_timesteps > 0
         return self._num_timesteps
@@ -559,7 +558,7 @@ class Dataset:
             for i in range(1, num):
                 seq_index[i::num] += i * (num_seqs // num)
         elif seq_ordering_method == "reverse":
-            seq_index = range(num_seqs - 1, -1, -1)  # type: Union[range, typing.Sequence[int]]
+            seq_index = range(num_seqs - 1, -1, -1)  # type: Union[range, Sequence[int]]
         elif seq_ordering_method in ["sorted", "sorted_reverse"]:
             assert get_seq_len
             reverse = -1 if seq_ordering_method == "sorted_reverse" else 1
@@ -748,12 +747,11 @@ class Dataset:
         """
         self.epoch = None
 
-    def get_current_seq_order(self):
+    def get_current_seq_order(self) -> Sequence[int]:
         """
         :return: many datasets use self.get_seq_order_for_epoch. this function would return the current seq order
           for the current epoch, after self.init_seq_order was called.
           Not all datasets implement this.
-        :rtype: typing.Sequence[int]
         """
         raise OptionalNotImplementedError
 
@@ -902,7 +900,7 @@ class Dataset:
         if self.seq_ordering == "default" and self.partition_epoch == 1:
             return seq_idx
         assert self.have_corpus_seq_idx()
-        raise NotImplemented
+        raise NotImplementedError
 
     def have_get_corpus_seq(self) -> bool:
         """
@@ -1061,7 +1059,7 @@ class Dataset:
         if key in self.num_outputs:
             if self.num_outputs[key][1] <= 1:
                 return []
-            res_shape = [None] * (self.num_outputs[key][1] - 1)  # type: typing.List[typing.Union[None,int]]
+            res_shape: List[Union[None, int]] = [None] * (self.num_outputs[key][1] - 1)
             if not self.is_data_sparse(key):
                 res_shape[-1] = self.get_data_dim(key)
             return res_shape

{returnn-1.20250719.212120 → returnn-1.20250725.163756}/returnn/datasets/cached2.py

@@ -4,18 +4,10 @@ Provides :class:`CachedDataset2`.
 
 from __future__ import annotations
 import numpy
-import typing
-from typing import Optional
+from typing import Optional, List
 from threading import Condition
 from .basic import Dataset, DatasetSeq
 
-try:
-    # noinspection PyCompatibility
-    from _thread import interrupt_main
-except ImportError:
-    # noinspection PyUnresolvedReferences,PyCompatibility
-    from thread import interrupt_main
-
 
 class CachedDataset2(Dataset):
     """
@@ -36,7 +28,7 @@ class CachedDataset2(Dataset):
         self._num_timesteps = None
         self.epoch = None
         self.reached_final_seq = False
-        self.added_data = []  # type: typing.List[DatasetSeq]
+        self.added_data: List[DatasetSeq] = []
         self.expected_load_seq_start = 0
         self._num_timesteps_accumulated = 0
 

{returnn-1.20250719.212120 → returnn-1.20250725.163756}/returnn/datasets/distrib_files.py

@@ -626,7 +626,7 @@ def _worker_proc_loop(
 ):
     if sys.platform == "linux":
         with open("/proc/self/comm", "w") as f:
-            f.write(f"CFD worker {epoch}")
+            f.write(f"DFD worker {epoch}")
     better_exchook.setup_all()
 
     assert isinstance(epoch, int) and isinstance(buffer_size, int)

{returnn-1.20250719.212120 → returnn-1.20250725.163756}/returnn/datasets/lm.py

@@ -24,7 +24,6 @@ from typing import (
     cast,
     Generator,
 )
-import typing
 import os
 from io import IOBase
 import sys
@@ -1563,7 +1562,6 @@ class TranslationDataset(CachedDataset2):
             import returnn.util.better_exchook
 
             returnn.util.better_exchook.install()
-            from returnn.util.basic import AsyncThreadRun
 
             # First iterate once over the data to get the data len as fast as possible.
             data_len = 0

{returnn-1.20250719.212120 → returnn-1.20250725.163756}/returnn/frontend/decoder/transformer.py

@@ -268,6 +268,7 @@ class TransformerDecoderLayer(rf.Module):
         ] = None,
         self_att_opts: Optional[Dict[str, Any]] = None,
         att_dropout: float = 0.1,
+        cross_att: Optional[Dict[str, Any]] = None,
         norm: Union[type, Dict[str, Any], rf.Module, Callable] = rf.LayerNorm,
     ):
         """
@@ -333,10 +334,10 @@ class TransformerDecoderLayer(rf.Module):
             raise TypeError(f"unexpected self_att type {self_att!r}")
         self.self_att_layer_norm = make_norm(norm, out_dim)
 
-        self.cross_att = None
+        self.cross_att: Optional[rf.CrossAttention] = None  # type might be inaccurate, but we expect this interface
         self.cross_att_layer_norm = None
         if encoder_dim is not None:
-            self.cross_att = rf.CrossAttention(
+            cross_att_opts = dict(
                 encoder_dim=self.encoder_dim,
                 query_in_dim=out_dim,
                 proj_dim=out_dim,
@@ -345,6 +346,12 @@ class TransformerDecoderLayer(rf.Module):
                 num_heads=num_heads,
                 att_dropout=att_dropout,
             )
+            if cross_att is None:
+                self.cross_att = rf.CrossAttention(**cross_att_opts)
+            elif isinstance(cross_att, dict):
+                self.cross_att: Optional[rf.CrossAttention] = rf.build_from_dict(cross_att, **cross_att_opts)
+            else:
+                raise TypeError(f"unexpected cross_att type {cross_att!r}")
             self.cross_att_layer_norm = make_norm(norm, out_dim)
 
     def default_initial_state(self, *, batch_dims: Sequence[Dim]) -> rf.State:

{returnn-1.20250719.212120 → returnn-1.20250725.163756/returnn.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20250719.212120
+Version: 1.20250725.163756
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

{returnn-1.20250719.212120 → returnn-1.20250725.163756}/tests/pycharm-inspect.py

@@ -11,14 +11,12 @@ See here:
 """
 
 import os
-import stat
 import sys
 import re
 import time
 import shutil
 import subprocess
 import tempfile
-import typing
 from glob import glob
 import argparse
 from xml.dom import minidom
@@ -115,8 +113,7 @@ def install_pycharm():
     print("Install PyCharm into:", pycharm_dir)
     sys.stdout.flush()
 
-    pycharm_version = (2020, 2)
-    name = "pycharm-community-%i.%i" % pycharm_version
+    name = "pycharm-community-2020.2"
     fn = "%s.tar.gz" % name
 
     subprocess.check_call(
@@ -466,6 +463,37 @@ def run_inspect(pycharm_dir, src_dir, skip_pycharm_inspect=False):
 
     fold_start("script.inspect")
     if not skip_pycharm_inspect:
+        fold_start("script.inspect.pycharm_inspect.sh.content")
+        with open("%s/bin/inspect.sh" % pycharm_dir) as f:
+            content = f.read()
+            print("Content of inspect.sh:")
+            print(content)
+        fold_end()
+
+        fold_start("script.inspect.pycharm.sh.content")
+        with open("%s/bin/pycharm.sh" % pycharm_dir) as f:
+            content = f.read()
+            print("Content of pycharm.sh:")
+            print(content)
+        fold_end()
+
+        fold_start("script.inspect.vmoptions.content")
+        fns = glob("%s/bin/*.vmoptions" % pycharm_dir)
+        if fns:
+            for fn in fns:
+                print("Content of %s:" % fn)
+                with open(fn) as f:
+                    content = f.read().splitlines(keepends=True)
+                print("".join(content))
+                if any(line.startswith("-Xmx") for line in content):
+                    print("Note: Patching Xmx settings...")
+                    content = ["-Xmx4000m\n" if line.startswith("-Xmx") else line for line in content]
+                    with open(fn, "w") as f:
+                        f.write("".join(content))
+        else:
+            print("No *.vmoptions found, not printing content.")
+        fold_end()
+
         # Note: Will not run if PyCharm is already running.
         # Maybe we can find some workaround for this?
         # See here: https://stackoverflow.com/questions/55339010/run-pycharm-inspect-sh-even-if-pycharm-is-already-runn
@@ -723,8 +751,11 @@ def main():
             "PyTypeCheckerInspection",  # too much false alarms: https://youtrack.jetbrains.com/issue/PY-34893
             # Not critical.
             "SpellCheckingInspection",  # way too much for now...
+            "GrazieInspection",  # grammar
             "PyClassHasNoInitInspection",  # not relevant?
             "PyMethodMayBeStaticInspection",  # not critical
+            # Does not work correctly here?
+            "PyPackageRequirementsInspection",  # TODO only with newer PyCharm versions?
         },
         ignore_count_for_files=ignore_count_for_files,
     )

returnn-1.20250719.212120/_setup_info_generated.py

@@ -1,2 +0,0 @@
-version = '1.20250719.212120'
-long_version = '1.20250719.212120+git.dd75ba0'