mistralai

/

Mixtral-8x7B-Instruct-v0.1

like 4.18k

Follow

Mistral AI_ 2,416

Adding the failure:

</>
from transformers import AutoTokenizer, AutoModelForCausalLM

In [3]: tokenizer = AutoTokenizer.from_pretrained("mistralai/Mixtral-8x7B-Instruct-v0.1")
...: model = AutoModelForCausalLM.from_pretrained("mistralai/Mixtral-8x7B-Instruct-v0.1")
Loading checkpoint shards: 26%|█████████████████████▊ | 5/19 [02:21<06:40, 28.60s/it]Killed
(pytorch_p38) ubuntu@ip-172-23-1-218:~$ ipython
Python 3.8.12 | packaged by conda-forge | (default, Oct 12 2021, 21:59:51)
Type 'copyright', 'credits' or 'license' for more information
IPython 7.31.1 -- An enhanced Interactive Python. Type '?' for help.

In [1]: # Use a pipeline as a high-level helper
...: from transformers import pipeline
...:
...: pipe = pipeline("text-generation", model="mistralai/Mixtral-8x7B-Instruct-v0.1")
Loading checkpoint shards: 26%|█████████████████████▊ | 5/19 [02:11<06:00, 25.75s/it]Killed
</>

Have you also tried loading it in half-precision by adding torch_dtype=torch.float16 in your pipeline? Something like this pipe = pipeline("text-generation", model="mistralai/Mixtral-8x7B-Instruct-v0.1", torch_dtype=torch.float16)

Have you considered deploying on other cloud platforms. I am using Runpod and it's working great. I have put together a guide here if you are interested: https://github.com/aigeek0x0/radiantloom-ai/blob/main/mixtral-8x7b-instruct-v-0.1-runpod-template.md

@aigeek0x0 have you performed fine-tuning using one A100 80GB runpod?

@bweinstein123 yes, i have. you can finetune this model with 4-bit quantization on A100. Even RTX A6000 would suffice if you use smaller batch size.

@aigeek0x0 how was the performance of mixtral instruct after fine-tuning? Any insights which I can borrow, thanks

@mlkorra based on vibes-check evaluations, it is looking promising. I haven't run benchmarks on it yet, but I will do so after a few more modifications.

Heyyy, same issue, i am trying to run it on a g5.8xlarge machine, and it gets killed RIGHT at the 26% checkpoint. Did you come across any solution?