zsh: killed python export.py

[aPaleBlueDot]

(coreml) hg@007 Mistral7B % python export.py
Loading checkpoint shards: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:48<00:00, 16.07s/it]
Converting PyTorch Frontend ==> MIL Ops: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████| 5575/5575 [00:02<00:00, 1958.56 ops/s]
Running MIL frontend_pytorch pipeline: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████| 5/5 [00:00<00:00, 6.52 passes/s]
Running MIL default pipeline: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 79/79 [11:06<00:00, 8.44s/ passes]
Running MIL backend_mlprogram pipeline: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████| 12/12 [00:00<00:00, 17.32 passes/s]
zsh: killed python export.py

[antmikinka]

Possible hardware limitation? I occured a lot of issues with zsh killing processes due to limited RAM or storage while converting models. Could you provide more details regarding your processes to help troubleshoot you?

[aPaleBlueDot]

Indeed, that's my guess as well, based on two runs ending at different times, where the longer lasting one had more disk space available. Also noticed another GH issue on the same topic where their run lasted much shorter. I'm on an M1 with 16GB RAM, and made 100GB disk space available on my 3rd try.

[aPaleBlueDot]

Are PyTorch checkpoints required or does safetsensors only HF repos work for the following line?

torch_model = StatefulMistralForCausalLM(MODEL_ID, max_context_size=max_context_size)