Albert Zeyer

Download the attached zipped pickle file: [dlm_traindata_debug_snapshot_002.pickle.zip](https://github.com/user-attachments/files/20816357/dlm_traindata_debug_snapshot_002.pickle.zip) (via @dorian-K) Unzip it. Drag & drop the pickle file to here: https://docs.pytorch.org/memory_viz That looks like this: None of the functions/ops in RF...

I'm seeing this now the first time ([here](https://github.com/rwth-i6/returnn/actions/runs/14908412440/job/41876349116)). It's maybe/probably non-determinstic and rare? (Not sure if some hiccup could cause this.) Anyway, reporting here now. ``` File "/home/runner/.local/lib/python3.10/site-packages/_pytest/python.py", line 159,...

Drop older Python support

2

Previous: * require Python >=3.8: https://github.com/rwth-i6/returnn/issues/1326 * require Python >=3.7: https://github.com/rwth-i6/returnn/issues/487 Bump again? To what? See also related recent Sisyphus discussion: https://github.com/rwth-i6/sisyphus/issues/195 Also the question came up w.r.t. automatic formatting...

``` ... File "/rwthfs/rz/cluster/home/az668407/setups/combined/2021-05-31/tools/returnn/returnn/util/file_cache.py", line 334, in FileCache.handle_cached_files_in_config.._handle_value line: res = self.get_file(value.filename) locals: res = self = self.get_file = value = CachedFile(filename='/rwthfs/rz/cluster/home/az668407/setups/combined/2021-05-31/work/i6_core/returnn/oggzip/BlissToOggZipJob.5ad18raRAWhr/output/ou t.ogg.zip') value.filename = '/rwthfs/rz/cluster/home/az668407/setups/combined/2021-05-31/work/i6_core/returnn/oggzip/BlissToOggZipJob.5ad18raRAWhr/output/out.ogg.zip', len = 136 File...

HDFDataset startup for huge dataset is slow

1

It takes maybe 20 mins or so. (I did not measure that yet.) The HDF file is 38G large and has 40M seqs with 4.613B frames. Via `dump-dataset.py`: ``` Returnn...

RF combine inconcistent between native and pure Python

2

``` ... PyExtModCompiler call: g++ -shared -O2 -std=c++11 -fno-strict-overflow -Wsign-compare -DDYNAMIC_ANNOTATIONS_ENABLED=1 -DNDEBUG -O2 -g -pipe -Wall -Werror=format-security -Wp,-D_FORTIFY_SOURCE=2 -Wp,-D_GLIBCXX_ASSERTIONS -fexceptions -fstack-protector-strong -grecord-gcc-switches -m64 -mtune=generic -fasynchronous-unwind-tables -fstack-cla sh-protection -fcf-protection -O2 -g...

MixingDataset needed

3

It would be similar as `ConcatDataset` in that it combines (concatenates) multiple datasets over the sequences (unlike `MetaDataset`, or `ConcatSeqsDataset`, or so, which combine within a sequence, or multiple sequences...

PostprocessingDataset with multi-processing

2

This could be an alternative to `MultiProcDataset`. In most cases (`OggZipDataset` etc), the data loading part of the dataset is not really the bottleneck, but any postprocessing is the bottleneck,...

There are various cases, e.g. whether we import for train, or for recog, or also just randomly initialize. Maybe it depends on the case whether it is ok to ignore...

Currently `rf.set_default_device` is called *after* the model was created. Specifically, we only use `rf.set_default_device_ctx(self._device)` around the run step, and not otherwise. The model creation happens on CPU, and then we...