text_to_dollars / app.py
m-ric's picture
m-ric HF staff
Update app.py
08b68f4 verified
raw
history blame
10 kB
import gradio as gr
import pandas as pd
import requests
import json
import tiktoken
import matplotlib.pyplot as plt
PRICES_URL = "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json"
# Ensure TOKEN_COSTS is up to date when the module is loaded
try:
response = requests.get(PRICES_URL)
if response.status_code == 200:
TOKEN_COSTS = response.json()
else:
raise Exception(f"Failed to fetch token costs, status code: {response.status_code}")
except Exception as e:
print(f'Failed to update token costs with error: {e}. Using static costs.')
with open("model_prices.json", "r") as f:
TOKEN_COSTS = json.load(f)
TOKEN_COSTS = pd.DataFrame.from_dict(TOKEN_COSTS, orient='index').reset_index()
TOKEN_COSTS.columns = ['model'] + list(TOKEN_COSTS.columns[1:])
TOKEN_COSTS = TOKEN_COSTS.loc[
(~TOKEN_COSTS["model"].str.contains("sample_spec"))
& (~TOKEN_COSTS["input_cost_per_token"].isnull())
& (~TOKEN_COSTS["output_cost_per_token"].isnull())
& (TOKEN_COSTS["input_cost_per_token"] > 0)
& (TOKEN_COSTS["output_cost_per_token"] > 0)
]
TOKEN_COSTS["supports_vision"] = TOKEN_COSTS["supports_vision"].fillna(False)
def clean_names(s):
s = s.replace("_", " ").replace("ai", "AI")
return s[0].upper() + s[1:]
TOKEN_COSTS["litellm_provider"] = TOKEN_COSTS["litellm_provider"].apply(clean_names)
cmap = plt.get_cmap('RdYlGn_r') # Red-Yellow-Green colormap, reversed
def count_string_tokens(string: str, model: str) -> int:
try:
encoding = tiktoken.encoding_for_model(model.split('/')[-1])
except:
if len(model.split('/')) > 1:
try:
encoding = tiktoken.encoding_for_model(model.split('/')[-2] + '/' + model.split('/')[-1])
except KeyError:
print(f"Model {model} not found. Using cl100k_base encoding.")
encoding = tiktoken.get_encoding("cl100k_base")
else:
print(f"Model {model} not found. Using cl100k_base encoding.")
encoding = tiktoken.get_encoding("cl100k_base")
return len(encoding.encode(string))
def calculate_total_cost(prompt_tokens: int, completion_tokens: int, model: str) -> float:
model_data = TOKEN_COSTS[TOKEN_COSTS['model'] == model].iloc[0]
prompt_cost = prompt_tokens * model_data['input_cost_per_token']
completion_cost = completion_tokens * model_data['output_cost_per_token']
return prompt_cost, completion_cost
def update_model_list(function_calling, litellm_provider, max_price, supports_vision, supports_max_input_tokens):
filtered_models = TOKEN_COSTS.loc[TOKEN_COSTS["max_input_tokens"] >= supports_max_input_tokens*1000]
if litellm_provider != "Any":
filtered_models = filtered_models[filtered_models['litellm_provider'] == litellm_provider]
if supports_vision:
filtered_models = filtered_models[filtered_models['supports_vision']]
list_models = filtered_models['model'].tolist()
return gr.Dropdown(choices=list_models, value=list_models[0] if list_models else "No model found for this combination!")
def compute_all(input_type, prompt_text, completion_text, base_prompt_tokens, base_completion_tokens, models):
results = []
for model in models:
if input_type == "Text Input":
prompt_tokens = count_string_tokens(prompt_text, model)
completion_tokens = count_string_tokens(completion_text, model)
else: # Token Count Input
prompt_tokens = int(base_prompt_tokens)
completion_tokens = int(base_completion_tokens)
model_data = TOKEN_COSTS[TOKEN_COSTS['model'] == model].iloc[0]
prompt_cost, completion_cost = calculate_total_cost(prompt_tokens, completion_tokens, model)
total_cost = prompt_cost + completion_cost
results.append({
"Model": model,
"Provider": model_data['litellm_provider'],
"Input Cost / M tokens": model_data['input_cost_per_token']*1e6,
"Output Cost / M tokens": model_data['output_cost_per_token']*1e6,
"Total Cost": round(total_cost, 6),
})
df = pd.DataFrame(results)
if len(df) > 1:
norm = plt.Normalize(df['Total Cost'].min(), df['Total Cost'].max())
def apply_color(val):
color = cmap(norm(val))
rgba = tuple(int(x * 255) for x in color[:3]) + (0.3,) # 0.5 for 50% opacity
return f'background-color: rgba{rgba}'
else:
def apply_color(val):
return "background-color: var(--input-background-fill)"
# Apply colors and formatting
def style_cell(val, column):
style = ''
if column == 'Total Cost':
style += 'font-weight: bold; '
style += apply_color(val)
if column in ['Total Cost']:
val = f'${val:.6f}'
if column in ["Input Cost / M tokens", "Output Cost / M tokens"]:
val = f'${val}'
return f'<td style="{style}">{val}</td>'
html_table = '<table class="styled-table">'
html_table += '<thead><tr>'
for col in df.columns:
html_table += f'<th>{col}</th>'
html_table += '</tr></thead><tbody>'
for _, row in df.iterrows():
html_table += '<tr>'
for col in df.columns:
html_table += style_cell(row[col], col)
html_table += '</tr>'
html_table += '</tbody></table>'
return html_table
def toggle_input_visibility(choice):
return (
gr.Group(visible=(choice == "Text Input")),
gr.Group(visible=(choice == "Token Count Input"))
)
with gr.Blocks(css="""
.styled-table {
border-collapse: collapse;
margin: 25px 0;
font-family: Arial, sans-serif;
width: 100%;
}
.styled-table th, .styled-table td {
padding: 12px 15px;
text-align: left;
vertical-align: middle;
}
.styled-table tbody tr {
border-bottom: 1px solid #dddddd;
}
.styled-table tbody tr:nth-of-type(even) {
background-color: var(--input-background-fill);
}
.styled-table tbody tr:nth-of-type(odd) {
background-color: var(--block-background-fill);
}
""", theme=gr.themes.Soft(primary_hue=gr.themes.colors.yellow, secondary_hue=gr.themes.colors.orange)) as demo:
gr.Markdown("""
# Text-to-Dollars: Get the price of your LLM API calls!
Based on prices data from [BerriAI's litellm](https://github.com/BerriAI/litellm/blob/main/model_prices_and_context_window.json).
""")
with gr.Row():
with gr.Column():
gr.Markdown("## Input type:")
input_type = gr.Radio(["Text Input", "Token Count Input"], label="Input Type", value="Text Input")
with gr.Group() as text_input_group:
prompt_text = gr.Textbox(label="Prompt", value="Tell me a joke about AI.", lines=3)
completion_text = gr.Textbox(label="Completion", value="Certainly: Why did the neural network go to therapy? It had too many deep issues!", lines=3)
with gr.Group(visible=False) as token_input_group:
prompt_tokens_input = gr.Number(label="Prompt Tokens", value=1500)
completion_tokens_input = gr.Number(label="Completion Tokens", value=2000)
with gr.Column():
gr.Markdown("## Model choice:")
with gr.Row():
with gr.Column():
function_calling = gr.Checkbox(label="Supports Tool Calling", value=False)
supports_vision = gr.Checkbox(label="Supports Vision", value=False)
with gr.Column():
supports_max_input_tokens = gr.Slider(label="Min Supported Input Length (thousands tokens)", minimum=2, maximum=256, step=2, value=2)
max_price = gr.Slider(label="Max Price per Input Token", minimum=0, maximum=0.001, step=0.00001, value=0.001, visible=False, interactive=False)
litellm_provider = gr.Dropdown(label="Inference Provider", choices=["Any"] + TOKEN_COSTS['litellm_provider'].unique().tolist(), value="Any")
model = gr.Dropdown(label="Models (at least 1)", choices=TOKEN_COSTS['model'].tolist(), value=["anyscale/meta-llama/Meta-Llama-3-8B-Instruct", "gpt-4o", "claude-3-sonnet-20240229"], multiselect=True)
gr.Markdown("## Resulting Costs 👇")
with gr.Row():
results_table = gr.HTML()
input_type.change(
toggle_input_visibility,
inputs=[input_type],
outputs=[text_input_group, token_input_group]
)
gr.on(
triggers=[function_calling.change, litellm_provider.change, max_price.change, supports_vision.change, supports_max_input_tokens.change],
fn=update_model_list,
inputs=[function_calling, litellm_provider, max_price, supports_vision, supports_max_input_tokens],
outputs=model,
)
gr.on(
triggers=[
input_type.change,
prompt_text.change,
completion_text.change,
prompt_tokens_input.change,
completion_tokens_input.change,
function_calling.change,
litellm_provider.change,
# max_price.change,
supports_vision.change,
supports_max_input_tokens.change,
model.change
],
fn=compute_all,
inputs=[
input_type,
prompt_text,
completion_text,
prompt_tokens_input,
completion_tokens_input,
model
],
outputs=results_table
)
# Load results on page load
demo.load(
fn=compute_all,
inputs=[
input_type,
prompt_text,
completion_text,
prompt_tokens_input,
completion_tokens_input,
model
],
outputs=results_table
)
if __name__ == "__main__":
demo.launch()