add TokenCounter and ask_with_images
This commit is contained in:
liangxinbing
parent
50ab26880e
commit
fb0d1c02a6
326
app/llm.py
326
app/llm.py
@ -1,3 +1,4 @@
|
||||
import math
|
||||
from typing import Dict, List, Optional, Union
|
||||
|
||||
import tiktoken
|
||||
@ -31,6 +32,142 @@ from app.schema import (
|
||||
REASONING_MODELS = ["o1", "o3-mini"]
|
||||
|
||||
|
||||
class TokenCounter:
|
||||
# Token constants
|
||||
BASE_MESSAGE_TOKENS = 4
|
||||
FORMAT_TOKENS = 2
|
||||
LOW_DETAIL_IMAGE_TOKENS = 85
|
||||
HIGH_DETAIL_TILE_TOKENS = 170
|
||||
|
||||
# Image processing constants
|
||||
MAX_SIZE = 2048
|
||||
HIGH_DETAIL_TARGET_SHORT_SIDE = 768
|
||||
TILE_SIZE = 512
|
||||
|
||||
def __init__(self, tokenizer):
|
||||
self.tokenizer = tokenizer
|
||||
|
||||
def count_text(self, text: str) -> int:
|
||||
"""Calculate tokens for a text string"""
|
||||
return 0 if not text else len(self.tokenizer.encode(text))
|
||||
|
||||
def count_image(self, image_item: dict) -> int:
|
||||
"""
|
||||
Calculate tokens for an image based on detail level and dimensions
|
||||
|
||||
For "low" detail: fixed 85 tokens
|
||||
For "high" detail:
|
||||
1. Scale to fit in 2048x2048 square
|
||||
2. Scale shortest side to 768px
|
||||
3. Count 512px tiles (170 tokens each)
|
||||
4. Add 85 tokens
|
||||
"""
|
||||
detail = image_item.get("detail", "medium")
|
||||
|
||||
# For low detail, always return fixed token count
|
||||
if detail == "low":
|
||||
return self.LOW_DETAIL_IMAGE_TOKENS
|
||||
|
||||
# For medium detail (default in OpenAI), use high detail calculation
|
||||
# OpenAI doesn't specify a separate calculation for medium
|
||||
|
||||
# For high detail, calculate based on dimensions if available
|
||||
if detail == "high" or detail == "medium":
|
||||
# If dimensions are provided in the image_item
|
||||
if "dimensions" in image_item:
|
||||
width, height = image_item["dimensions"]
|
||||
return self._calculate_high_detail_tokens(width, height)
|
||||
|
||||
# Default values when dimensions aren't available or detail level is unknown
|
||||
if detail == "high":
|
||||
# Default to a 1024x1024 image calculation for high detail
|
||||
return self._calculate_high_detail_tokens(1024, 1024) # 765 tokens
|
||||
elif detail == "medium":
|
||||
# Default to a medium-sized image for medium detail
|
||||
return 1024 # This matches the original default
|
||||
else:
|
||||
# For unknown detail levels, use medium as default
|
||||
return 1024
|
||||
|
||||
def _calculate_high_detail_tokens(self, width: int, height: int) -> int:
|
||||
"""Calculate tokens for high detail images based on dimensions"""
|
||||
# Step 1: Scale to fit in MAX_SIZE x MAX_SIZE square
|
||||
if width > self.MAX_SIZE or height > self.MAX_SIZE:
|
||||
scale = self.MAX_SIZE / max(width, height)
|
||||
width = int(width * scale)
|
||||
height = int(height * scale)
|
||||
|
||||
# Step 2: Scale so shortest side is HIGH_DETAIL_TARGET_SHORT_SIDE
|
||||
scale = self.HIGH_DETAIL_TARGET_SHORT_SIDE / min(width, height)
|
||||
scaled_width = int(width * scale)
|
||||
scaled_height = int(height * scale)
|
||||
|
||||
# Step 3: Count number of 512px tiles
|
||||
tiles_x = math.ceil(scaled_width / self.TILE_SIZE)
|
||||
tiles_y = math.ceil(scaled_height / self.TILE_SIZE)
|
||||
total_tiles = tiles_x * tiles_y
|
||||
|
||||
# Step 4: Calculate final token count
|
||||
return (
|
||||
total_tiles * self.HIGH_DETAIL_TILE_TOKENS
|
||||
) + self.LOW_DETAIL_IMAGE_TOKENS
|
||||
|
||||
def count_content(self, content: Union[str, List[Union[str, dict]]]) -> int:
|
||||
"""Calculate tokens for message content"""
|
||||
if not content:
|
||||
return 0
|
||||
|
||||
if isinstance(content, str):
|
||||
return self.count_text(content)
|
||||
|
||||
token_count = 0
|
||||
for item in content:
|
||||
if isinstance(item, str):
|
||||
token_count += self.count_text(item)
|
||||
elif isinstance(item, dict):
|
||||
if "text" in item:
|
||||
token_count += self.count_text(item["text"])
|
||||
elif "image_url" in item:
|
||||
token_count += self.count_image(item)
|
||||
return token_count
|
||||
|
||||
def count_tool_calls(self, tool_calls: List[dict]) -> int:
|
||||
"""Calculate tokens for tool calls"""
|
||||
token_count = 0
|
||||
for tool_call in tool_calls:
|
||||
if "function" in tool_call:
|
||||
function = tool_call["function"]
|
||||
token_count += self.count_text(function.get("name", ""))
|
||||
token_count += self.count_text(function.get("arguments", ""))
|
||||
return token_count
|
||||
|
||||
def count_message_tokens(self, messages: List[dict]) -> int:
|
||||
"""Calculate the total number of tokens in a message list"""
|
||||
total_tokens = self.FORMAT_TOKENS # Base format tokens
|
||||
|
||||
for message in messages:
|
||||
tokens = self.BASE_MESSAGE_TOKENS # Base tokens per message
|
||||
|
||||
# Add role tokens
|
||||
tokens += self.count_text(message.get("role", ""))
|
||||
|
||||
# Add content tokens
|
||||
if "content" in message:
|
||||
tokens += self.count_content(message["content"])
|
||||
|
||||
# Add tool calls tokens
|
||||
if "tool_calls" in message:
|
||||
tokens += self.count_tool_calls(message["tool_calls"])
|
||||
|
||||
# Add name and tool_call_id tokens
|
||||
tokens += self.count_text(message.get("name", ""))
|
||||
tokens += self.count_text(message.get("tool_call_id", ""))
|
||||
|
||||
total_tokens += tokens
|
||||
|
||||
return total_tokens
|
||||
|
||||
|
||||
class LLM:
|
||||
_instances: Dict[str, "LLM"] = {}
|
||||
|
||||
@ -81,6 +218,8 @@ class LLM:
|
||||
else:
|
||||
self.client = AsyncOpenAI(api_key=self.api_key, base_url=self.base_url)
|
||||
|
||||
self.token_counter = TokenCounter(self.tokenizer)
|
||||
|
||||
def count_tokens(self, text: str) -> int:
|
||||
"""Calculate the number of tokens in a text"""
|
||||
if not text:
|
||||
@ -88,46 +227,7 @@ class LLM:
|
||||
return len(self.tokenizer.encode(text))
|
||||
|
||||
def count_message_tokens(self, messages: List[dict]) -> int:
|
||||
"""Calculate the number of tokens in a message list"""
|
||||
token_count = 0
|
||||
for message in messages:
|
||||
# Base token count for each message (according to OpenAI's calculation method)
|
||||
token_count += 4 # Base token count for each message
|
||||
|
||||
# Calculate tokens for the role
|
||||
if "role" in message:
|
||||
token_count += self.count_tokens(message["role"])
|
||||
|
||||
# Calculate tokens for the content
|
||||
if "content" in message and message["content"]:
|
||||
token_count += self.count_tokens(message["content"])
|
||||
|
||||
# Calculate tokens for tool calls
|
||||
if "tool_calls" in message and message["tool_calls"]:
|
||||
for tool_call in message["tool_calls"]:
|
||||
if "function" in tool_call:
|
||||
# Function name
|
||||
if "name" in tool_call["function"]:
|
||||
token_count += self.count_tokens(
|
||||
tool_call["function"]["name"]
|
||||
)
|
||||
# Function arguments
|
||||
if "arguments" in tool_call["function"]:
|
||||
token_count += self.count_tokens(
|
||||
tool_call["function"]["arguments"]
|
||||
)
|
||||
|
||||
# Calculate tokens for tool responses
|
||||
if "name" in message and message["name"]:
|
||||
token_count += self.count_tokens(message["name"])
|
||||
|
||||
if "tool_call_id" in message and message["tool_call_id"]:
|
||||
token_count += self.count_tokens(message["tool_call_id"])
|
||||
|
||||
# Add extra tokens for message format
|
||||
token_count += 2 # Extra tokens for message format
|
||||
|
||||
return token_count
|
||||
return self.token_counter.count_message_tokens(messages)
|
||||
|
||||
def update_token_count(self, input_tokens: int) -> None:
|
||||
"""Update token counts"""
|
||||
@ -313,6 +413,152 @@ class LLM:
|
||||
logger.error(f"Unexpected error in ask: {e}")
|
||||
raise
|
||||
|
||||
@retry(
|
||||
wait=wait_random_exponential(min=1, max=60),
|
||||
stop=stop_after_attempt(6),
|
||||
retry=retry_if_exception_type(
|
||||
(OpenAIError, Exception, ValueError)
|
||||
), # Don't retry TokenLimitExceeded
|
||||
)
|
||||
async def ask_with_images(
|
||||
self,
|
||||
messages: List[Union[dict, Message]],
|
||||
images: List[Union[str, dict]],
|
||||
system_msgs: Optional[List[Union[dict, Message]]] = None,
|
||||
stream: bool = False,
|
||||
temperature: Optional[float] = None,
|
||||
) -> str:
|
||||
"""
|
||||
Send a prompt with images to the LLM and get the response.
|
||||
|
||||
Args:
|
||||
messages: List of conversation messages
|
||||
images: List of image URLs or image data dictionaries
|
||||
system_msgs: Optional system messages to prepend
|
||||
stream (bool): Whether to stream the response
|
||||
temperature (float): Sampling temperature for the response
|
||||
|
||||
Returns:
|
||||
str: The generated response
|
||||
|
||||
Raises:
|
||||
TokenLimitExceeded: If token limits are exceeded
|
||||
ValueError: If messages are invalid or response is empty
|
||||
OpenAIError: If API call fails after retries
|
||||
Exception: For unexpected errors
|
||||
"""
|
||||
try:
|
||||
# Format messages
|
||||
formatted_messages = self.format_messages(messages)
|
||||
|
||||
# Ensure the last message is from the user to attach images
|
||||
if not formatted_messages or formatted_messages[-1]["role"] != "user":
|
||||
raise ValueError(
|
||||
"The last message must be from the user to attach images"
|
||||
)
|
||||
|
||||
# Process the last user message to include images
|
||||
last_message = formatted_messages[-1]
|
||||
|
||||
# Convert content to multimodal format if needed
|
||||
content = last_message["content"]
|
||||
multimodal_content = (
|
||||
[{"type": "text", "text": content}]
|
||||
if isinstance(content, str)
|
||||
else content
|
||||
if isinstance(content, list)
|
||||
else []
|
||||
)
|
||||
|
||||
# Add images to content
|
||||
for image in images:
|
||||
if isinstance(image, str):
|
||||
multimodal_content.append(
|
||||
{"type": "image_url", "image_url": {"url": image}}
|
||||
)
|
||||
elif isinstance(image, dict) and "url" in image:
|
||||
multimodal_content.append({"type": "image_url", "image_url": image})
|
||||
elif isinstance(image, dict) and "image_url" in image:
|
||||
multimodal_content.append(image)
|
||||
else:
|
||||
raise ValueError(f"Unsupported image format: {image}")
|
||||
|
||||
# Update the message with multimodal content
|
||||
last_message["content"] = multimodal_content
|
||||
|
||||
# Add system messages if provided
|
||||
if system_msgs:
|
||||
all_messages = self.format_messages(system_msgs) + formatted_messages
|
||||
else:
|
||||
all_messages = formatted_messages
|
||||
|
||||
# Calculate tokens and check limits
|
||||
input_tokens = self.count_message_tokens(all_messages)
|
||||
if not self.check_token_limit(input_tokens):
|
||||
raise TokenLimitExceeded(self.get_limit_error_message(input_tokens))
|
||||
|
||||
# Set up API parameters
|
||||
params = {
|
||||
"model": self.model,
|
||||
"messages": all_messages,
|
||||
"stream": stream,
|
||||
}
|
||||
|
||||
# Add model-specific parameters
|
||||
if self.model in REASONING_MODELS:
|
||||
params["max_completion_tokens"] = self.max_tokens
|
||||
else:
|
||||
params["max_tokens"] = self.max_tokens
|
||||
params["temperature"] = (
|
||||
temperature if temperature is not None else self.temperature
|
||||
)
|
||||
|
||||
# Handle non-streaming request
|
||||
if not stream:
|
||||
response = await self.client.chat.completions.create(**params)
|
||||
|
||||
if not response.choices or not response.choices[0].message.content:
|
||||
raise ValueError("Empty or invalid response from LLM")
|
||||
|
||||
self.update_token_count(response.usage.prompt_tokens)
|
||||
return response.choices[0].message.content
|
||||
|
||||
# Handle streaming request
|
||||
self.update_token_count(input_tokens)
|
||||
response = await self.client.chat.completions.create(**params)
|
||||
|
||||
collected_messages = []
|
||||
async for chunk in response:
|
||||
chunk_message = chunk.choices[0].delta.content or ""
|
||||
collected_messages.append(chunk_message)
|
||||
print(chunk_message, end="", flush=True)
|
||||
|
||||
print() # Newline after streaming
|
||||
full_response = "".join(collected_messages).strip()
|
||||
|
||||
if not full_response:
|
||||
raise ValueError("Empty response from streaming LLM")
|
||||
|
||||
return full_response
|
||||
|
||||
except TokenLimitExceeded:
|
||||
raise
|
||||
except ValueError as ve:
|
||||
logger.error(f"Validation error in ask_with_images: {ve}")
|
||||
raise
|
||||
except OpenAIError as oe:
|
||||
logger.error(f"OpenAI API error: {oe}")
|
||||
if isinstance(oe, AuthenticationError):
|
||||
logger.error("Authentication failed. Check API key.")
|
||||
elif isinstance(oe, RateLimitError):
|
||||
logger.error("Rate limit exceeded. Consider increasing retry attempts.")
|
||||
elif isinstance(oe, APIError):
|
||||
logger.error(f"API error: {oe}")
|
||||
raise
|
||||
except Exception as e:
|
||||
logger.error(f"Unexpected error in ask_with_images: {e}")
|
||||
raise
|
||||
|
||||
@retry(
|
||||
wait=wait_random_exponential(min=1, max=60),
|
||||
stop=stop_after_attempt(6),
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user