hendrik
/
matrix-pygmalion-bot
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity
Browse Source

implement pseudo-streaming

master
Hendrik Langer 2 years ago
parent
7e9918a06d
commit
9d5f2de7a5
2 changed files with 55 additions and 38 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Unified View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 61
      matrix_pygmalion_bot/ai/koboldcpp.py
  2. 32
      matrix_pygmalion_bot/core.py

61
matrix_pygmalion_bot/ai/koboldcpp.py
View File

@ -37,7 +37,7 @@ async def generate_sync(
"Content-Type": "application/json", "Content-Type": "application/json",
} }
max_new_tokens = 120 max_new_tokens = 200
prompt_num_tokens = await num_tokens(prompt) prompt_num_tokens = await num_tokens(prompt)
# Define your inputs # Define your inputs
@ -55,28 +55,45 @@ async def generate_sync(
logger.info(f"sending request to koboldcpp") logger.info(f"sending request to koboldcpp")
# Make the request
try:
r = requests.post(endpoint, json=input_data, headers=headers, timeout=360) TIMEOUT = 360
except requests.exceptions.RequestException as e: DELAY = 5
raise ValueError(f"<HTTP ERROR> {e}") tokens = 0
r_json = r.json() complete_reply = ""
logger.info(r_json) for i in range(TIMEOUT//DELAY):
input_data["max_length"] = 16 # pseudo streaming
if r.status_code == 200: # Make the request
reply = r_json["results"][0]["text"] try:
idx = reply.find(f"\nYou:") r = requests.post(endpoint, json=input_data, headers=headers, timeout=360)
if idx != -1: except requests.exceptions.RequestException as e:
reply = reply[:idx].strip() raise ValueError(f"<HTTP ERROR> {e}")
r_json = r.json()
logger.info(r_json)
if r.status_code == 200:
partial_reply = r_json["results"][0]["text"]
input_data["prompt"] += partial_reply
complete_reply += partial_reply
tokens += input_data["max_length"]
await typing_fn()
if not partial_reply or partial_reply.find('<|endoftext|>') != -1 or partial_reply.find("\nYou:") != -1 or tokens >= max_new_tokens:
idx = complete_reply.find(f"\nYou:")
if idx != -1:
complete_reply = complete_reply[:idx].strip()
else:
complete_reply = complete_reply.removesuffix('<|endoftext|>').strip()
complete_reply = complete_reply.replace(f"\n{bot.name}: ", " ")
complete_reply = complete_reply.replace(f"\n<BOT>: ", " ")
complete_reply = complete_reply.replace(f"<BOT>", "{bot.name}")
complete_reply = complete_reply.replace(f"<USER>", "You")
return complete_reply.strip()
else:
continue
elif r.status_code == 503:
#model busy
await asyncio.sleep(DELAY)
else: else:
reply = reply.removesuffix('<|endoftext|>').strip() raise ValueError(f"<ERROR>")
reply = reply.replace(f"\n{bot.name}: ", " ")
reply = reply.replace(f"\n<BOT>: ", " ")
reply = reply.replace(f"<BOT>", "{bot.name}")
reply = reply.replace(f"<USER>", "You")
return reply.strip()
else:
raise ValueError(f"<ERROR>")
async def generate_image(input_prompt: str, negative_prompt: str, api_url: str, api_key: str, typing_fn): async def generate_image(input_prompt: str, negative_prompt: str, api_url: str, api_key: str, typing_fn):

32
matrix_pygmalion_bot/core.py
View File

@ -16,8 +16,8 @@ import json
from .helpers import Event from .helpers import Event
from .chatlog import BotChatHistory from .chatlog import BotChatHistory
ai = importlib.import_module("matrix_pygmalion_bot.ai.runpod_pygmalion") image_ai = importlib.import_module("matrix_pygmalion_bot.ai.runpod_pygmalion")
ai = importlib.import_module("matrix_pygmalion_bot.ai.koboldcpp") text_ai = importlib.import_module("matrix_pygmalion_bot.ai.koboldcpp")
#ai = importlib.import_module("matrix_pygmalion_bot.ai.stablehorde") #ai = importlib.import_module("matrix_pygmalion_bot.ai.stablehorde")
#from .llama_cpp import generate, get_full_prompt, get_full_prompt_chat_style #from .llama_cpp import generate, get_full_prompt, get_full_prompt_chat_style
#from .runpod_pygmalion import generate_sync, get_full_prompt #from .runpod_pygmalion import generate_sync, get_full_prompt
@ -114,30 +114,30 @@ class Callbacks(object):
typing = lambda : self.client.room_typing(room.room_id, True, 15000) typing = lambda : self.client.room_typing(room.room_id, True, 15000)
if self.bot.service == "runpod": if self.bot.service == "runpod":
if num == 1: if num == 1:
output = await ai.generate_image1(prompt, negative_prompt, self.bot.runpod_api_key, typing) output = await image_ai.generate_image1(prompt, negative_prompt, self.bot.runpod_api_key, typing)
elif num == 2: elif num == 2:
output = await ai.generate_image2(prompt, negative_prompt, self.bot.runpod_api_key, typing) output = await image_ai.generate_image2(prompt, negative_prompt, self.bot.runpod_api_key, typing)
elif num == 3: elif num == 3:
output = await ai.generate_image3(prompt, negative_prompt, self.bot.runpod_api_key, typing) output = await image_ai.generate_image3(prompt, negative_prompt, self.bot.runpod_api_key, typing)
elif num == 4: elif num == 4:
output = await ai.generate_image4(prompt, negative_prompt, self.bot.runpod_api_key, typing) output = await image_ai.generate_image4(prompt, negative_prompt, self.bot.runpod_api_key, typing)
elif num == 5: elif num == 5:
output = await ai.generate_image5(prompt, negative_prompt, self.bot.runpod_api_key, typing) output = await image_ai.generate_image5(prompt, negative_prompt, self.bot.runpod_api_key, typing)
elif num == 6: elif num == 6:
output = await ai.generate_image6(prompt, negative_prompt, self.bot.runpod_api_key, typing) output = await image_ai.generate_image6(prompt, negative_prompt, self.bot.runpod_api_key, typing)
elif num == 7: elif num == 7:
output = await ai.generate_image7(prompt, negative_prompt, self.bot.runpod_api_key, typing) output = await image_ai.generate_image7(prompt, negative_prompt, self.bot.runpod_api_key, typing)
elif num == 8: elif num == 8:
output = await ai.generate_image8(prompt, negative_prompt, self.bot.runpod_api_key, typing) output = await image_ai.generate_image8(prompt, negative_prompt, self.bot.runpod_api_key, typing)
else: else:
raise ValueError('no image generator with that number') raise ValueError('no image generator with that number')
elif self.bot.service == "stablehorde": elif self.bot.service == "stablehorde":
if num == 1: if num == 1:
output = await ai.generate_image1(prompt, negative_prompt, self.bot.stablehorde_api_key, typing) output = await image_ai.generate_image1(prompt, negative_prompt, self.bot.stablehorde_api_key, typing)
elif num == 2: elif num == 2:
output = await ai.generate_image2(prompt, negative_prompt, self.bot.stablehorde_api_key, typing) output = await image_ai.generate_image2(prompt, negative_prompt, self.bot.stablehorde_api_key, typing)
elif num == 3: elif num == 3:
output = await ai.generate_image3(prompt, negative_prompt, self.bot.stablehorde_api_key, typing) output = await image_ai.generate_image3(prompt, negative_prompt, self.bot.stablehorde_api_key, typing)
else: else:
raise ValueError('no image generator with that number') raise ValueError('no image generator with that number')
else: else:
@ -216,8 +216,8 @@ class Callbacks(object):
# send, mail, drop, snap picture, photo, image, portrait # send, mail, drop, snap picture, photo, image, portrait
pass pass
full_prompt = await ai.get_full_prompt(chat_message.getTranslation("en"), self.bot, self.bot.chat_history.room(room.display_name)) full_prompt = await text_ai.get_full_prompt(chat_message.getTranslation("en"), self.bot, self.bot.chat_history.room(room.display_name))
num_tokens = await ai.num_tokens(full_prompt) num_tokens = await text_ai.num_tokens(full_prompt)
logger.debug(full_prompt) logger.debug(full_prompt)
logger.debug(f"Prompt has " + str(num_tokens) + " tokens") logger.debug(f"Prompt has " + str(num_tokens) + " tokens")
# answer = "" # answer = ""
@ -237,7 +237,7 @@ class Callbacks(object):
# print("") # print("")
try: try:
typing = lambda : self.client.room_typing(room.room_id, True, 15000) typing = lambda : self.client.room_typing(room.room_id, True, 15000)
answer = await ai.generate_sync(full_prompt, self.bot.runpod_api_key, self.bot, typing, api_endpoint) answer = await text_ai.generate_sync(full_prompt, self.bot.runpod_api_key, self.bot, typing, api_endpoint)
answer = answer.strip() answer = answer.strip()
await self.client.room_typing(room.room_id, False) await self.client.room_typing(room.room_id, False)
if not (self.bot.translate is None): if not (self.bot.translate is None):

Write Preview
Loading…
Cancel
Save