text-generation-webui/modules/llamacpp_model.py

95 lines
2.6 KiB
Python
Raw Normal View History

2023-03-19 07:42:10 +01:00
import os
from pathlib import Path
import modules.shared as shared
from modules.callbacks import Iteratorize
import llamacpp
class LlamaCppTokenizer:
"""A thin wrapper over the llamacpp tokenizer"""
def __init__(self, model: llamacpp.PyLLAMA):
self._tokenizer = model.get_tokenizer()
self.eos_token_id = 2
self.bos_token_id = 0
@classmethod
def from_model(cls, model: llamacpp.PyLLAMA):
return cls(model)
def encode(self, prompt):
return self._tokenizer.tokenize(prompt)
def decode(self, ids):
return self._tokenizer.detokenize(ids)
class LlamaCppModel:
def __init__(self):
self.initialized = False
@classmethod
def from_pretrained(self, path):
params = llamacpp.gpt_params(
str(path), # model
2048, # ctx_size
200, # n_predict
40, # top_k
0.95, # top_p
0.80, # temp
1.30, # repeat_penalty
-1, # seed
8, # threads
64, # repeat_last_n
8, # batch_size
)
_model = llamacpp.PyLLAMA(params)
result = self()
result.model = _model
tokenizer = LlamaCppTokenizer.from_model(_model)
return result, tokenizer
# TODO: Allow passing in params for each inference
def generate(self, context="", num_tokens=10, callback=None):
# params = self.params
# params.n_predict = token_count
# params.top_p = top_p
# params.top_k = top_k
# params.temp = temperature
# params.repeat_penalty = repetition_penalty
# params.repeat_last_n = repeat_last_n
# model.params = params
if not self.initialized:
self.model.add_bos()
self.model.update_input(context)
if not self.initialized:
self.model.prepare_context()
self.initialized = True
output = ""
is_end_of_text = False
ctr = 0
while not self.model.is_finished() and ctr < num_tokens and not is_end_of_text:
if self.model.has_unconsumed_input():
self.model.ingest_all_pending_input(False)
else:
text, is_end_of_text = self.model.infer_text()
if callback:
callback(text)
output += text
ctr += 1
return output
def generate_with_streaming(self, **kwargs):
with Iteratorize(self.generate, kwargs, callback=None) as generator:
reply = kwargs['context']
for token in generator:
reply += token
yield reply