PhiMiX-2x2B-raw

Code is work in progress

This is a RAW MoE meant to be finetuned

PhiMiX-2x2B is a Mixure of Experts (MoE) made with the following models using mergekit:

• cognitivecomputations/dolphin-2_6-phi-2
• rhysjones/phi-2-orange

©️ Credits

• mlabonne's phixtral for the PhiConfig and inference code.
• mergekit code which I tweaked (you can find the PhiConfig here) by mainly adding the config in the moe_mixtral.py script from mixtral branch.

🧩 Configuration

base_model: rhysjones/phi-2-orange
gate_mode: random
dtype: float16
experts:
  - source_model: cognitivecomputations/dolphin-2_6-phi-2
    positive_prompts: [""]
  - source_model: rhysjones/phi-2-orange
    positive_prompts: [""]

💻 Usage

!pip install -qU transformers bitsandbytes accelerate
from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
import torch

model_name = "paulilioaica/PhiMiX-2x2B-raw"

torch.set_default_device("cuda")

config = AutoConfig.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_config(config, trust_remote_code=True)

instruction = '''
    def print_prime(n):
        """
        Print all primes between 1 and n
        """
'''


tokenizer = AutoTokenizer.from_pretrained(
    f"{model_name}", 
    trust_remote_code=True
)

# Tokenize the input string
inputs = tokenizer(
    instruction, 
    return_tensors="pt", 
    return_attention_mask=False
)

# Generate text using the model
outputs = model.generate(**inputs, max_length=200)

# Decode and print the output
text = tokenizer.batch_decode(outputs)[0]
print(text)