File size: 2,081 Bytes
37c8faa 22b1752 37c8faa 73d8c0b 37c8faa 73d8c0b 37c8faa 73d8c0b 37c8faa 73d8c0b |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 |
# Code to inference Hermes with HF Transformers
# Requires pytorch, transformers, bitsandbytes, sentencepiece, protobuf, and flash-attn packages
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from transformers import LlamaTokenizer, MixtralForCausalLM
import bitsandbytes, flash_attn
class EndpointHandler:
def __init__(self, path=""):
self.tokenizer = LlamaTokenizer.from_pretrained(path, trust_remote_code=True)
self.model = MixtralForCausalLM.from_pretrained(
path,
torch_dtype=torch.float16,
device_map="auto",
load_in_8bit=False,
load_in_4bit=True,
use_flash_attention_2=True
)
def __call__(self, data: Any) -> List[List[Dict[str, float]]]:
sys_prompt=data["prompt"]
list=data["inputs"]
prompt=f"<|im_start|>system\n{sys_prompt}.<|im_end|>\n"
for item in list:
if item["role"]=="assistant":
content=item["content"]
prompt+=f"<|im_start|>assistant\n{content}<|im_end|>\n"
else:
content=item["content"]
prompt+=f"<|im_start|>user\n{content}<|im_end|>\n"
prompt+="<|im_start|>assistant\n"
#for chat in prompts:
#print(chat)
input_ids = self.tokenizer(prompt, return_tensors="pt").input_ids.to("cuda")
generated_ids = self.model.generate(input_ids, max_new_tokens=750, temperature=0.8, repetition_penalty=1.1, do_sample=True, eos_token_id=self.tokenizer.eos_token_id)
response = self.tokenizer.decode(generated_ids[0][input_ids.shape[-1]:], skip_special_tokens=True, clean_up_tokenization_space=True)
return response
"""
encodeds = self.tokenizer.encode(prompt, return_tensors="pt")
model_inputs = encodeds.to(device)
self.model.to(device)
generated_ids = self.model.generate(model_inputs, max_new_tokens=1000, do_sample=True)
decoded = self.tokenizer.decode(generated_ids[0])
return decoded
"""
|