internlm2-chat-20b-llama

internlm/internlm2-chat-20b weights formatted to match standard Llama modeling code. Model can be loaded directly, but for tokenizer use trust_remote_code

usage:

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "kiranr/internlm2-chat-20b-llama"

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

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.float16,
    device_map="auto",
    attn_implementation="flash_attention_2",
)
messages = [
    {"role": "user", "content": "what is the square root of banana?"}
]

model_input = tokenizer.apply_chat_template(messages, return_tensors="pt").to("cuda")

generated_ids = model.generate(
    model_input,
    max_new_tokens=1024,
    do_sample=True,
    eos_token_id=[92542, 2],  # <|im_end|> and </s>
)
output = tokenizer.decode(
    generated_ids[0][model_input.shape[-1] : -1], skip_special_tokens=True
)
print(output)