Evgeny Pavlov

`pip-compile` doesn't lock the requirements for `bicleaner-ai==3.0.1`: ``` pip-compile -r pipeline/bicleaner/requirements/bicleaner-ai.in ERROR: Cannot install -r pipeline/bicleaner/requirements/bicleaner-ai.in (line 1) because these package versions have conflicting dependencies. Discarding bicleaner-hardrules==2.7.0 (from -r pipeline/bicleaner/requirements/bicleaner-ai.txt...

landing #561 and setting up dashboards for CPU machines can help with that

I see all the new models translate this example and everything else related to numbers correctly. I think `num_mismatch` discarded all the training examples with currency conversions. Closing.

@bhearsum I wonder if there's a way to pull a secret based on the author of the task. Then instead of using our service token we could use personal tokens...

Ok, ideally we should do tricks with tokens because it turned out W&B has some nice filters like "show only my work" based on this Author field and I don't...

@La0 @vrigal ok, we see this consistently, it seems the parser just can't parse: ``` [task 2024-05-09T20:18:17.301Z] [2024-05-09 20:18:17] [config] Loaded model has been created with Marian v1.12.14 2d067af 2024-02-16...

Also, this blocks landing https://github.com/mozilla/firefox-translations-training/pull/580/files as we want to make sure W&B continues tracking correctly after the training is restarted.

> Haven't seen reports of this elsewhere. > > @eu9ene - have you seen this on GPU workers only? Or also on the CPU workers? I'm not sure but I...

Ok, it happens every single pipeline run for me now and does not restart. Marking as blocker.

Did this go away by itself?