mirror of
https://github.com/BerriAI/litellm.git
synced 2026-11-08 08:08:37 +00:00
ab78459097
Response transformation crashed with TypeError when counting tokens for multimodal (dict-shaped) embedding inputs, since encoding.encode() was called directly on the dict instead of its text parts. Also removed a dead, equally-buggy duplicate of the same loop in _transform_response. Separately, the embeddings guardrail handler silently skipped multimodal dict inputs instead of scanning them, letting prohibited text/image content bypass the configured guardrail. It now extracts and scans the text parts while leaving image parts untouched.
243 lines
8.8 KiB
Python
243 lines
8.8 KiB
Python
import os
|
|
import sys
|
|
|
|
sys.path.insert(0, os.path.abspath("../.."))
|
|
import litellm
|
|
import requests
|
|
from bs4 import BeautifulSoup
|
|
|
|
# URL of the AWS Bedrock Pricing page
|
|
PRICING_URL = "https://aws.amazon.com/bedrock/pricing/"
|
|
|
|
# List of providers to extract pricing for
|
|
PROVIDERS = ["ai21", "anthropic", "meta", "cohere", "mistral", "stability", "amazon"]
|
|
|
|
|
|
def extract_amazon_pricing(section):
|
|
"""
|
|
Extracts pricing data for Amazon-specific models.
|
|
|
|
Args:
|
|
section (Tag): The BeautifulSoup Tag object for the Amazon section.
|
|
|
|
Returns:
|
|
dict: Pricing data for Amazon models.
|
|
"""
|
|
tabs = section.find_all("li", class_="lb-tabs-trigger")
|
|
panels = section.find_all("li", class_="lb-tabs-content-item")
|
|
|
|
amazon_pricing = {}
|
|
|
|
for tab, panel in zip(tabs, panels):
|
|
model_name = tab.get_text(strip=True)
|
|
table = panel.find("table")
|
|
if not table:
|
|
amazon_pricing[model_name] = "Pricing table not found"
|
|
continue
|
|
|
|
# Parse the table
|
|
rows = table.find_all("tr")
|
|
headers = [header.get_text(strip=True) for header in rows[0].find_all("td")]
|
|
model_pricing = {}
|
|
|
|
for row in rows[1:]:
|
|
cols = row.find_all("td")
|
|
if len(cols) < 3:
|
|
continue # Skip rows with insufficient data
|
|
|
|
feature_name = cols[0].get_text(strip=True)
|
|
input_price = cols[1].get_text(strip=True)
|
|
output_price = cols[2].get_text(strip=True)
|
|
model_pricing[feature_name] = {
|
|
headers[1]: input_price,
|
|
headers[2]: output_price,
|
|
}
|
|
|
|
amazon_pricing[model_name] = model_pricing
|
|
|
|
return amazon_pricing
|
|
|
|
|
|
def get_bedrock_pricing(url, providers):
|
|
"""
|
|
Fetches and parses AWS Bedrock pricing for specified providers.
|
|
|
|
Args:
|
|
url (str): URL of the AWS Bedrock pricing page.
|
|
providers (list): List of providers to extract pricing for.
|
|
|
|
Returns:
|
|
dict: A dictionary containing pricing data for the providers.
|
|
"""
|
|
response = requests.get(url)
|
|
response.raise_for_status()
|
|
soup = BeautifulSoup(response.text, "html.parser")
|
|
|
|
pricing_data = {}
|
|
|
|
for provider in providers:
|
|
if provider == "amazon":
|
|
section = soup.find(
|
|
"li",
|
|
class_="lb-tabs-accordion-trigger",
|
|
text=lambda t: t and "Amazon" in t,
|
|
)
|
|
if not section:
|
|
pricing_data[provider] = "Amazon section not found"
|
|
continue
|
|
|
|
amazon_section = section.find_next("li", class_="lb-tabs-content-item")
|
|
if not amazon_section:
|
|
pricing_data[provider] = "Amazon models section not found"
|
|
continue
|
|
|
|
pricing_data[provider] = extract_amazon_pricing(amazon_section)
|
|
else:
|
|
# General logic for other providers
|
|
section = soup.find(
|
|
"h2", text=lambda t: t and provider.lower() in t.lower()
|
|
)
|
|
if not section:
|
|
pricing_data[provider] = "Provider section not found"
|
|
continue
|
|
|
|
table = section.find_next("table")
|
|
if not table:
|
|
pricing_data[provider] = "Pricing table not found"
|
|
continue
|
|
|
|
rows = table.find_all("tr")
|
|
headers = [header.get_text(strip=True) for header in rows[0].find_all("td")]
|
|
provider_pricing = {}
|
|
|
|
for row in rows[1:]:
|
|
cols = row.find_all("td")
|
|
if len(cols) < 3:
|
|
continue
|
|
|
|
model_name = cols[0].get_text(strip=True)
|
|
input_price = cols[1].get_text(strip=True)
|
|
output_price = cols[2].get_text(strip=True)
|
|
provider_pricing[model_name] = {
|
|
"Price per 1,000 input tokens": input_price,
|
|
"Price per 1,000 output tokens": output_price,
|
|
}
|
|
|
|
pricing_data[provider] = provider_pricing
|
|
|
|
return pricing_data
|
|
|
|
|
|
model_substring_map = {
|
|
"ai21": {"jurassic-2": "j2"},
|
|
"anthropic": {"claude-2-1": "claude-v2:1", "claude-2-0": "claude-v2"},
|
|
"meta": {"llama-2-chat-(13b)": "llama2-13b-chat"},
|
|
"cohere": {
|
|
"r+": "r-plus",
|
|
"embed-3-english": "embed-english-v3",
|
|
"embed-3-multilingual": "embed-multilingual-v3",
|
|
},
|
|
} # aliases used by bedrock in their real model name vs. pricing page
|
|
|
|
|
|
def _handle_meta_model_name(model_name: str) -> str:
|
|
# Check if it's a Llama 2 chat model
|
|
if "llama-2-chat-" in model_name.lower():
|
|
# Extract the size (e.g., 13b, 70b) using string manipulation
|
|
# Look for pattern between "chat-(" and ")"
|
|
import re
|
|
|
|
if match := re.search(r"chat-\((\d+b)\)", model_name.lower()):
|
|
size = match.group(1)
|
|
return f"meta.llama2-{size}-chat"
|
|
return model_name
|
|
|
|
|
|
def _handle_cohere_model_name(model_name: str) -> str:
|
|
if model_name.endswith("command-r"):
|
|
return "cohere.command-r-v1"
|
|
return model_name
|
|
|
|
|
|
def _create_bedrock_model_name(provider: str, model_name: str):
|
|
complete_model_name = f"{provider.lower()}.{model_name.replace(' ', '-').replace('.', '-').replace('*', '').lower()}"
|
|
for provider_key, map in model_substring_map.items():
|
|
if provider_key == provider:
|
|
for model_substring, replacement in map.items():
|
|
print(
|
|
f"model_substring: {model_substring}, replacement: {replacement}, received model_name: {model_name}"
|
|
)
|
|
if model_substring in complete_model_name:
|
|
print(f"model_name: {complete_model_name}")
|
|
complete_model_name = complete_model_name.replace(
|
|
model_substring, replacement
|
|
)
|
|
print(f"model_name: {complete_model_name}")
|
|
if provider == "meta":
|
|
complete_model_name = _handle_meta_model_name(complete_model_name)
|
|
if provider == "cohere":
|
|
complete_model_name = _handle_cohere_model_name(complete_model_name)
|
|
return complete_model_name
|
|
|
|
|
|
def _convert_str_to_float(price_str: str) -> float:
|
|
if "$" not in price_str:
|
|
return 0.0
|
|
return float(price_str.replace("$", ""))
|
|
|
|
|
|
def _check_if_model_name_in_pricing(
|
|
bedrock_model_name: str,
|
|
input_cost_per_1k_tokens: str,
|
|
output_cost_per_1k_tokens: str,
|
|
):
|
|
os.environ["LITELLM_LOCAL_MODEL_COST_MAP"] = "True"
|
|
litellm.model_cost = litellm.get_model_cost_map(url="")
|
|
|
|
for model, value in litellm.model_cost.items():
|
|
if model.startswith(bedrock_model_name):
|
|
input_cost_per_token = (
|
|
_convert_str_to_float(input_cost_per_1k_tokens) / 1000
|
|
)
|
|
output_cost_per_token = (
|
|
_convert_str_to_float(output_cost_per_1k_tokens) / 1000
|
|
)
|
|
assert round(value["input_cost_per_token"], 10) == round(
|
|
input_cost_per_token, 10
|
|
), f"Invalid input cost per token for {model} \n Bedrock pricing page name={bedrock_model_name} \n Got={value['input_cost_per_token']}, Expected={input_cost_per_token}"
|
|
assert round(value["output_cost_per_token"], 10) == round(
|
|
output_cost_per_token, 10
|
|
), f"Invalid output cost per token for {model} \n Bedrock pricing page name={bedrock_model_name} \n Got={value['output_cost_per_token']}, Expected={output_cost_per_token}"
|
|
return True
|
|
return False
|
|
|
|
|
|
if __name__ == "__main__":
|
|
try:
|
|
pricing = get_bedrock_pricing(PRICING_URL, PROVIDERS)
|
|
print("AWS Bedrock On-Demand Pricing:")
|
|
for provider, data in pricing.items():
|
|
print(f"\n{provider.capitalize()}:")
|
|
if isinstance(data, dict):
|
|
for model, details in data.items():
|
|
complete_model_name = _create_bedrock_model_name(provider, model)
|
|
print(f"details: {details}")
|
|
assert _check_if_model_name_in_pricing(
|
|
bedrock_model_name=complete_model_name,
|
|
input_cost_per_1k_tokens=details[
|
|
"Price per 1,000 input tokens"
|
|
],
|
|
output_cost_per_1k_tokens=details[
|
|
"Price per 1,000 output tokens"
|
|
],
|
|
), f"Model {complete_model_name} not found in litellm.model_cost"
|
|
print(f" {complete_model_name}:")
|
|
if isinstance(details, dict):
|
|
for detail, value in details.items():
|
|
print(f" {detail}: {value}")
|
|
else:
|
|
print(f" {details}")
|
|
else:
|
|
print(f" {data}")
|
|
except requests.RequestException as e:
|
|
print(f"Error fetching pricing data: {e}")
|