Albert Zeyer

Some fix for invalid broadcasting

5

Also see #1749.

I had this bug: ```python log_prob = ... # [B,T+1,D] targets = ... # [B,T] -> D loss = rf.cross_entropy(target=targets, estimated=log_prob, ...) loss.mark_as_loss(...) ``` What you get here is *no*...

In some Python file, we still have this, which basically survived for many years: ``` __author__ = "Patrick Doetsch" __copyright__ = "Copyright 2015" __credits__ = ["Patrick Doetsch", "Paul Voigtlaender"] __license__...

Sharding for multi-GPU training

2

Now, as a follow-up of #1630: a very nice next step/feature would be if we can use this sharding feature in general for any kind of multi GPU training. Similar...

Slurm engine, handle AssocMaxSubmitJobLimit

8

At the RWTH ITC cluster, I get this: ``` [2025-02-12 16:37:20,610] ERROR: Error: sbatch: error: AssocMaxSubmitJobLimit [2025-02-12 16:37:20,610] ERROR: SBATCH command: sbatch -J i6_core.returnn.forward.ReturnnForwardJobV2.oEi02THsWnau.run --mail-type=None --mem=6G --gres=gpu:1 --cpus-per-task=2 --tim e=1440...

skip _sis_runnable check when already finished

4

Via: https://github.com/rwth-i6/sisyphus/issues/167#issuecomment-2503658756 See the corresponding discussion there for why I did this. However, I'm not really sure that this is correct or wanted. I think this mostly makes a potential...

create_aliases never used when --ui is used

1

I'm not sure I'm missing it, but when looking through the code, it seems when using `--ui`, the aliases are not updated automatically? (`create_aliases` is not called automatically at startup,...

get_object_state for objects with __getnewargs__ (or ..._ex__), state unnecessary

10

In `get_object_state`, we already do this: ```python if hasattr(obj, "__getnewargs_ex__"): args = obj.__getnewargs_ex__() elif hasattr(obj, "__getnewargs__"): args = obj.__getnewargs__() else: args = None ... if args is None: return state...

Fix #215

sis_hash_helper/get_object_state fails for Numpy arrays

1

I think in earlier Sisyphus versions, it would not throw an exception, but dangerously just ignore the content of the Numpy array. Now in the latest Sisyphus version, it catches...