v2ray

/

GPT4chan-8B

like 2

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id="v2ray/GPT4chan-8B"

model = AutoModelForCausalLM.from_pretrained(model_id, 
                                             torch_dtype=torch.bfloat16).to("cuda") .eval)_                                            
tokenizer = AutoTokenizer.from_pretrained(model_id)

# jinja2 template for tokeniser formatting
template = """g
{% for message in messages -%}
<|start_header_id|>{{ loop.index }}<|end_header_id|>
{{ message['content'] }}
{% endfor -%}
{% if add_generation_prompt and messages[-1]['role'] != 'assistant' %}
<|start_header_id|>{{ messages | length +1 }}<|end_header_id|>
{% endif %}
"""

messages = [
    {"role": "user", "content": "speculate thread\nwhat will ai land be like in 2025"},
]

tokenizer.chat_template = template
s = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
print("Formatted message\n", s)

inputs = tokenizer([s], return_tensors="pt", padding="max_length", max_length=1024, truncation=True, padding_side='left').to("cuda")

with torch.no_grad():
  inputs = tokenizer([s], return_tensors="pt",
                      ).to("cuda")
  r = model.generate(**inputs,  
                      max_length=1000000,
                    pad_token_id=tokenizer.eos_token_id, early_stopping=False, min_new_tokens=32,)

o = tokenizer.decode(r[0], skip_special_tokens=False)
o = o.replace("<|start_header_id|>", "\n<|start_header_id|>")
print("Input and generate\n", o)

 <|begin_of_text|>g

<|start_header_id|>1<|end_header_id|>
speculate thread
what will ai land be like in 2025

<|start_header_id|>2<|end_header_id|>
>AI will take over the world
>it will be bad
I don't care
<|start_header_id|>3<|end_header_id|>>>2
i do care
<|start_header_id|>4<|end_header_id|>>>1 (OP)
>what will ai land be like in 2025
A bunch of b