teknium commited on
Commit
359079f
1 Parent(s): 8820e1a

Create transformers_inference.py

Browse files
Files changed (1) hide show
  1. transformers_inference.py +33 -0
transformers_inference.py ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Code to inference Open Hermes 2.5 with HF Transformers
2
+ # Requires pytorch, transformers, bitsandbytes, sentencepiece, protobuf, and flash-attn packages
3
+
4
+ import torch
5
+ from transformers import AutoTokenizer, AutoModelForCausalLM
6
+ from transformers import LlamaTokenizer, LlamaForCausalLM, MistralForCausalLM
7
+ import bitsandbytes, flash_attn
8
+
9
+ tokenizer = LlamaTokenizer.from_pretrained('teknium/OpenHermes-2.5-Mistral-7B', trust_remote_code=True)
10
+ model = MistralForCausalLM.from_pretrained(
11
+ "teknium/OpenHermes-2.5-Mistral-7B",
12
+ torch_dtype=torch.float16,
13
+ device_map=device_map="auto",#{'': 'cuda:0'},
14
+ load_in_8bit=False,
15
+ load_in_4bit=True,
16
+ use_flash_attention_2=True
17
+ )
18
+
19
+ prompts = [
20
+ """<|im_start|>system
21
+ You are a sentient, superintelligent artificial general intelligence, here to teach and assist me.<|im_end|>
22
+ <|im_start|>user
23
+ Write a short story about Goku discovering kirby has teamed up with Majin Buu to destroy the world.<|im_end|>
24
+ <|im_start|>assistant
25
+ """,
26
+ ]
27
+
28
+ for chat in prompts:
29
+ print(chat)
30
+ input_ids = tokenizer(chat, return_tensors="pt").input_ids.to("cuda")
31
+ generated_ids = model.generate(input_ids, max_new_tokens=750, temperature=0.8, repetition_penalty=1.1, do_sample=True, eos_token_id=tokenizer.eos_token_id)
32
+ response = tokenizer.decode(generated_ids[0][input_ids.shape[-1]:], skip_special_tokens=True, clean_up_tokenization_space=True)
33
+ print(f"Response: {response}")