All my 3.80 wus has this error

File "axiom_streaming_client.py", line 1407, in train_loop
AttributeError: 'ExpertWorker' object has no attribute 'sgd_step'
[Trainer] Error: 'ExpertWorker' object has no attribute 'sgd_step'
Traceback (most recent call last):
[BOINC] Training stopped. 0 total samples
[BOINC] Self-testing on 32768 bytes holdout...
[SelfTest] Error: Unable to allocate 6.00 MiB for an array with shape (8, 64, 3072) and data type float32
[BOINC] Self-test: bpc=1.0000 acc=50.0% (0 samples)
[BOINC] Timeout waiting for tasks
[SelfTest] Error: Unable to allocate 6.00 MiB for an array with shape (8, 64, 3072) and data type float32
[BOINC] Cleanup self-test: bpc=1.0000 acc=50.0% (0 samples)
[BOINC] Complete: syncs=0, samples=0, bpc=1.0
2026-02-11 14:08:43 (3608): axiom_worker_380_windows_x86_64.exe exited; CPU time 1999.640625
2026-02-11 14:08:43 (3608): called boinc_finish(0)

Unable to allocate 6mb (MiB, for precision) of ram??

It looks to me that the Admin is developing on the fly . There haven been several new versions already today and it looks like he puts them online without testing.

An other problem with the recent versions is the memory use. When the WU nears the end is starts useing all avaliable memory on my system. It has gone up to 32 GB and there was nothing left free.

In reply to Henk Haneveld's message of 11 Feb 2026:

An other problem with the recent versions is the memory use. When the WU nears the end is starts useing all avaliable memory on my system. It has gone up to 32 GB and there was nothing left free.

I noticed some lags of my pc, but I didn't look at the memory

And another new version (3.82)

Still contains an error and the absurd memory use.

It's AI in Python, memory usage is going to be absurd.

But yeah, nearly 10 revisions a day it seems. The admin did a test post here and not said a thing since. There used to be posts in discord and on the front page regarding revision updates but the last was 3.66

A new day and a new version, we are up to 3.85 .

This one looks a lot beter, no errors and the memory use is way down.