llama-3-4x8b

This is an MOE of Llama-3-8b with 4 experts. This does not use semantic routing, as this utilizes the deepseek-moe architecture. There is no routing, and there is no gate - all experts are active on every token.

import torch
from transformers import AutoTokenizer, TextStreamer, AutoModelForCausalLM

model_path = "Crystalcareai/llama-3-4x8b"
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    device_map="auto",
    low_cpu_mem_usage=True,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
    attn_implementation="flash_attention_2",
)

tokenizer = AutoTokenizer.from_pretrained(model_path)
streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)

# Modify the prompt to match the Alpaca instruction template
prompt = """
Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
Sam is faster than Joe. Joe is faster than Jane. Is Sam faster than Jane? Explain your reasoning step by step.

### Input:

### Response:
"""

tokens = tokenizer(
    prompt, 
    return_tensors='pt'
).input_ids.cuda()

generation_output = model.generate(
    tokens, 
    streamer=streamer,
    max_new_tokens=512,
)