amgadhasan/phi-2

Disclaimer

I do NOT own this model. It belongs to its developer (Microsoft). See the license file for more details.

Overview

This repo contains the parameters of phi-2, which is a large language model developed by Microsoft.

How to run

This model requires 12.5 GB of vRAM in float32.

Should take roughly 6.7 GB in float16.

1. Setup

install the needed libraries

pip install sentencepiece transformers accelerate einops

2. Download the model

from huggingface_hub import snapshot_download
model_path = snapshot_download(repo_id="amgadhasan/phi-2",repo_type="model", local_dir="./phi-2", local_dir_use_symlinks=False)

3. Load and run the model

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)

# We need to trust remote code since this hasn't been integrated in transformers as of version 4.35
model = AutoModelForCausalLM.from_pretrained(model_path, device_map="auto", trust_remote_code=True)

def generate(prompt: str, generation_params: dict = {"max_length":200})-> str : 
    inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
    outputs = model.generate(**inputs, **generation_params)
    completion = tokenizer.batch_decode(outputs)[0]
    return completion

result = generate(prompt)
result

float16

To load this model in float16, use the following code:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)

# We need to trust remote code since this hasn't been integrated in transformers as of version 4.35
# We need to set the torch dtype globally since this model class doesn't accept dtype as argument
torch.set_default_dtype(torch.float16)
model = AutoModelForCausalLM.from_pretrained(model_path, device_map="auto", trust_remote_code=True)

def generate(prompt: str, generation_params: dict = {"max_length":200})-> str : 
    inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
    outputs = model.generate(**inputs, **generation_params)
    completion = tokenizer.batch_decode(outputs)[0]
    return completion

result = generate(prompt)
result

Acknowledgments

Special thanks to Microsoft for developing and releasing this mode. Also, special thanks to the huggingface team for hosting LLMs for free!