text-generation-webui/modules/RWKV.py

75 lines
2.4 KiB
Python
Raw Normal View History

2023-02-28 04:09:11 +01:00
import os
2023-02-28 03:50:16 +01:00
from pathlib import Path
2023-02-28 04:09:11 +01:00
2023-02-28 03:50:16 +01:00
import numpy as np
2023-03-06 12:45:49 +01:00
from tokenizers import Tokenizer
2023-02-28 04:09:11 +01:00
import modules.shared as shared
2023-03-08 06:50:49 +01:00
from modules.callbacks import Iteratorize
2023-02-28 04:09:11 +01:00
2023-02-28 03:50:16 +01:00
np.set_printoptions(precision=4, suppress=True, linewidth=200)
os.environ['RWKV_JIT_ON'] = '1'
os.environ["RWKV_CUDA_ON"] = '1' if shared.args.rwkv_cuda_on else '0' # use CUDA kernel for seq mode (much faster)
2023-02-28 03:50:16 +01:00
from rwkv.model import RWKV
from rwkv.utils import PIPELINE, PIPELINE_ARGS
2023-03-01 16:18:17 +01:00
2023-03-01 16:08:55 +01:00
class RWKVModel:
def __init__(self):
pass
2023-02-28 03:50:16 +01:00
2023-03-01 16:08:55 +01:00
@classmethod
def from_pretrained(self, path, dtype="fp16", device="cuda"):
tokenizer_path = Path(f"{path.parent}/20B_tokenizer.json")
2023-02-28 03:50:16 +01:00
2023-03-02 00:02:48 +01:00
if shared.args.rwkv_strategy is None:
model = RWKV(model=str(path), strategy=f'{device} {dtype}')
2023-03-02 00:02:48 +01:00
else:
model = RWKV(model=str(path), strategy=shared.args.rwkv_strategy)
pipeline = PIPELINE(model, str(tokenizer_path))
2023-02-28 03:50:16 +01:00
2023-03-01 16:08:55 +01:00
result = self()
2023-03-01 16:33:09 +01:00
result.pipeline = pipeline
2023-03-01 16:08:55 +01:00
return result
2023-03-31 19:45:17 +02:00
def generate(self, context="", token_count=20, temperature=1, top_p=1, top_k=50, repetition_penalty=None, alpha_frequency=0.1, alpha_presence=0.1, token_ban=[0], token_stop=[], callback=None):
2023-03-01 16:16:11 +01:00
args = PIPELINE_ARGS(
temperature = temperature,
top_p = top_p,
2023-03-07 21:24:28 +01:00
top_k = top_k,
2023-03-01 16:19:37 +01:00
alpha_frequency = alpha_frequency, # Frequency Penalty (as in GPT-3)
alpha_presence = alpha_presence, # Presence Penalty (as in GPT-3)
token_ban = token_ban, # ban the generation of some tokens
token_stop = token_stop
2023-03-01 16:16:11 +01:00
)
return self.pipeline.generate(context, token_count=token_count, args=args, callback=callback)
2023-03-06 12:45:49 +01:00
2023-03-07 22:17:56 +01:00
def generate_with_streaming(self, **kwargs):
with Iteratorize(self.generate, kwargs, callback=None) as generator:
reply = ''
for token in generator:
reply += token
yield reply
2023-03-07 22:17:56 +01:00
2023-03-06 12:45:49 +01:00
class RWKVTokenizer:
def __init__(self):
pass
@classmethod
def from_pretrained(self, path):
tokenizer_path = path / "20B_tokenizer.json"
tokenizer = Tokenizer.from_file(str(tokenizer_path))
2023-03-06 12:45:49 +01:00
result = self()
result.tokenizer = tokenizer
return result
def encode(self, prompt):
return self.tokenizer.encode(prompt).ids
def decode(self, ids):
return self.tokenizer.decode(ids)