Files
DB-GPT/pilot/model/chatglm_llm.py
2023-05-21 16:18:33 +08:00

47 lines
1.3 KiB
Python

#!/usr/bin/env python3
# -*- coding:utf-8 -*-
import torch
@torch.inference_mode()
def chatglm_generate_stream(model, tokenizer, params, device, context_len=2048, stream_interval=2):
"""Generate text using chatglm model's chat api """
prompt = params["prompt"]
temperature = float(params.get("temperature", 1.0))
top_p = float(params.get("top_p", 1.0))
stop = params.get("stop", "###")
echo = params.get("echo", False)
generate_kwargs = {
"do_sample": True if temperature > 1e-5 else False,
"top_p": top_p,
"repetition_penalty": 1.0,
"logits_processor": None
}
if temperature > 1e-5:
generate_kwargs["temperature"] = temperature
# TODO, Fix this
hist = []
messages = prompt.split(stop)
# Add history chat to hist for model.
for i in range(1, len(messages) - 2, 2):
hist.append((messages[i].split(":")[1], messages[i+1].split(":")[1]))
query = messages[-2].split(":")[1]
print("Query Message: ", query)
output = ""
i = 0
for i, (response, new_hist) in enumerate(model.stream_chat(tokenizer, query, hist, **generate_kwargs)):
if echo:
output = query + " " + response
else:
output = response
yield output
yield output