Open tensimixt opened 5 days ago
@patrickvonplaten Hi do you know if mistral-inference works for lora+mixtral8x7b instruct v0.1? It does work for lora+mistral-7b v0.3 but getting error about LoRA weights file being loaded missing an expected key for one of the model layers when trying for lora+mixtral8x7b instruct v0.1
Is there something else required to make it work?
Thank you
Python Version
Pip Freeze
Reproduction Steps
clone repo download mixtral 8x7b instruct v0.1 ==> put in /mistral_models download v3 tokenizer and put into /mistral_models run util extend which generates /mistral_models_extended put v3 tokenizer into /mistral_models_extended directory put data into /data run data validation train (generates checkpoints. after 300 steps get /workspace/mistral-finetune/experiment5/checkpoints/checkpoint_000300/consolidated/lora.safetensors)
Finally run mistral-chat: torchrun --nproc-per-node 2 --no-python mistral-chat /workspace/mistral_models_extended --max_tokens 256 --temperature 0.7 --instruct --lora_path workspace/mistral-finetune/experiment5/checkpoints/checkpoint_000300/consolidated/lora.safetensors
This generates the following error
Expected Behavior
Expect Prompt to appear for interactive chat in terminal, but getting the above error instead.