dwolfe66/text-generation-webui-space
1
1import os2from pathlib import Path3 4import numpy as np5from tokenizers import Tokenizer6 7import modules.shared as shared8from modules.callbacks import Iteratorize9 10np.set_printoptions(precision=4, suppress=True, linewidth=200)11 12os.environ['RWKV_JIT_ON'] = '1'13os.environ["RWKV_CUDA_ON"] = '1' if shared.args.rwkv_cuda_on else '0' # use CUDA kernel for seq mode (much faster)14 15from rwkv.model import RWKV16from rwkv.utils import PIPELINE, PIPELINE_ARGS17 18 19class RWKVModel:20 def __init__(self):21 pass22 23 @classmethod24 def from_pretrained(self, path, dtype="fp16", device="cuda"):25 tokenizer_path = Path(f"{path.parent}/20B_tokenizer.json")26 27 if shared.args.rwkv_strategy is None:28 model = RWKV(model=str(path), strategy=f'{device} {dtype}')29 else:30 model = RWKV(model=str(path), strategy=shared.args.rwkv_strategy)31 pipeline = PIPELINE(model, str(tokenizer_path))32 33 result = self()34 result.pipeline = pipeline35 return result36 37 def generate(self, context="", token_count=20, temperature=1, top_p=1, top_k=50, alpha_frequency=0.1, alpha_presence=0.1, token_ban=[0], token_stop=[], callback=None):38 args = PIPELINE_ARGS(39 temperature = temperature,40 top_p = top_p,41 top_k = top_k,42 alpha_frequency = alpha_frequency, # Frequency Penalty (as in GPT-3)43 alpha_presence = alpha_presence, # Presence Penalty (as in GPT-3)44 token_ban = token_ban, # ban the generation of some tokens45 token_stop = token_stop46 )47 48 return context+self.pipeline.generate(context, token_count=token_count, args=args, callback=callback)49 50 def generate_with_streaming(self, **kwargs):51 with Iteratorize(self.generate, kwargs, callback=None) as generator:52 reply = kwargs['context']53 for token in generator:54 reply += token55 yield reply56 57class RWKVTokenizer:58 def __init__(self):59 pass60 61 @classmethod62 def from_pretrained(self, path):63 tokenizer_path = path / "20B_tokenizer.json"64 tokenizer = Tokenizer.from_file(str(tokenizer_path))65 66 result = self()67 result.tokenizer = tokenizer68 return result69 70 def encode(self, prompt):71 return self.tokenizer.encode(prompt).ids72 73 def decode(self, ids):74 return self.tokenizer.decode(ids)75 