from gradio.components import Component
import gradio as gr
import pandas as pd
from abc import ABC, abstractclassmethod
import inspect
class BaseTCOModel(ABC):
# TO DO: Find way to specify which component should be used for computing cost
def __setattr__(self, name, value):
if isinstance(value, Component):
self._components.append(value)
self.__dict__[name] = value
def __init__(self):
super(BaseTCOModel, self).__setattr__("_components", [])
self.use_case = None
self.num_users = None
self.input_tokens = None
self.output_tokens = None
def get_components(self) -> list[Component]:
return self._components
def get_components_for_cost_computing(self):
return self.components_for_cost_computing
def get_name(self):
return self.name
def register_components_for_cost_computing(self):
args = inspect.getfullargspec(self.compute_cost_per_token)[0][1:]
self.components_for_cost_computing = [self.__getattribute__(arg) for arg in args]
@abstractclassmethod
def compute_cost_per_token(self):
pass
@abstractclassmethod
def render(self):
pass
def set_name(self, name):
self.name = name
def set_formula(self, formula):
self.formula = formula
def get_formula(self):
return self.formula
def set_latency(self, latency):
self.latency = latency
def get_latency(self):
return self.latency
class OpenAIModel(BaseTCOModel):
def __init__(self):
self.set_name("(SaaS) OpenAI")
self.set_formula(r"""$CR = \frac{CIT\_1K \times IT + COT\_1K \times OT}{1000}$
with:
CR = Cost per Request
CIT_1K = Cost per 1000 Input Tokens (from OpenAI's pricing web page)
COT_1K = Cost per 1000 Output Tokens
IT = Input Tokens
OT = Output Tokens
""")
self.latency = "15s" #Default value for GPT4
super().__init__()
def render(self):
def on_model_change(model):
if model == "GPT-4":
self.latency = "15s"
return gr.Dropdown.update(choices=["8K", "32K"])
else:
self.latency = "5s"
return gr.Dropdown.update(choices=["4K", "16K"], value="4K")
self.model = gr.Dropdown(["GPT-4", "GPT-3.5 Turbo"], value="GPT-4",
label="OpenAI models",
interactive=True, visible=False)
self.context_length = gr.Dropdown(["8K", "32K"], value="8K", interactive=True,
label="Context size",
visible=False, info="Number of tokens the model considers when processing text")
self.model.change(on_model_change, inputs=self.model, outputs=self.context_length)
def compute_cost_per_token(self, model, context_length):
"""Cost per token = """
if model == "GPT-4" and context_length == "8K":
cost_per_1k_input_tokens = 0.03
cost_per_1k_output_tokens = 0.06
elif model == "GPT-4" and context_length == "32K":
cost_per_1k_input_tokens = 0.06
cost_per_1k_output_tokens = 0.12
elif model == "GPT-3.5" and context_length == "4K":
cost_per_1k_input_tokens = 0.0015
cost_per_1k_output_tokens = 0.002
else:
cost_per_1k_input_tokens = 0.003
cost_per_1k_output_tokens = 0.004
cost_per_input_token = (cost_per_1k_input_tokens / 1000)
cost_per_output_token = (cost_per_1k_output_tokens / 1000)
return cost_per_input_token, cost_per_output_token
class OpenSourceLlama2Model(BaseTCOModel):
def __init__(self):
self.set_name("(Open source) Llama 2")
self.set_formula(r"""$CT = \frac{VM\_CH \times 100}{3600 \times U} \times (\frac{IT}{ITS} + \frac{OT}{OTS})$
with:
CT = Cost per Token
VM_CH = VM Cost per Hour
ITS = Input Tokens per Second
OTS = Output Tokens per Second
U = Used
IT = Input Tokens
OT = Output Tokens
""")
self.set_latency("27s")
super().__init__()
def render(self):
vm_choices = ["1x Nvidia A100 (Azure NC24ads A100 v4)",
"2x Nvidia A100 (Azure NC24ads A100 v4)",
"2x Nvidia A100 (Azure ND96amsr A100 v4)"]
def on_model_change(model):
if model == "Llama 2 7B":
return [gr.Dropdown.update(choices=vm_choices),
gr.Markdown.update(value="To see the benchmark results use for the Llama2-7B model, [click here](https://example.com/script)"),
gr.Number.update(value=3.6730),
gr.Number.update(value=694.38),
gr.Number.update(value=694.38),
]
else:
not_supported_vm = ["1x Nvidia A100 (Azure NC24ads A100 v4)", "2x Nvidia A100 (Azure NC24ads A100 v4)"]
choices = [x for x in vm_choices if x not in not_supported_vm]
return [gr.Dropdown.update(choices=choices, value="2x Nvidia A100 (Azure ND96amsr A100 v4)"),
gr.Markdown.update(value="To see the benchmark results used for the Llama2-70B model, [click here](https://www.cursor.so/blog/llama-inference#user-content-fn-llama-paper)"),
gr.Number.update(value=2*37.186),
gr.Number.update(value=2860),
gr.Number.update(value=18.545),
]
def on_vm_change(model, vm):
# TO DO: load info from CSV
if model == "Llama 2 7B" and vm == "1x Nvidia A100 (Azure NC24ads A100 v4)":
return [gr.Number.update(value=4.777), gr.Number.update(value=694.38), gr.Number.update(value=694.38)]
elif model == "Llama 2 7B" and vm == "2x Nvidia A100 (Azure NC24ads A100 v4)":
return [gr.Number.update(value=2*4.777), gr.Number.update(value=1388.76), gr.Number.update(value=1388.76)]
elif model == "Llama 2 7B" and vm == "2x Nvidia A100 (Azure ND96amsr A100 v4)":
return [gr.Number.update(value=2*37.186), gr.Number.update(value=2777.52), gr.Number.update(value=2777.52)]
elif model == "Llama 2 70B" and vm == "2x Nvidia A100 (Azure ND96amsr A100 v4)":
return [gr.Number.update(value=2*37.186), gr.Number.update(value=2860), gr.Number.update(value=18.449)]
self.model = gr.Dropdown(["Llama 2 7B", "Llama 2 70B"], value="Llama 2 70B", label="OpenSource models", visible=False)
self.vm = gr.Dropdown(choices=["2x Nvidia A100 (Azure ND96amsr A100 v4)"],
value="2x Nvidia A100 (Azure ND96amsr A100 v4)",
visible=False,
label="Instance of VM with GPU",
info="Your options for this choice depend on the model you previously chose"
)
self.vm_cost_per_hour = gr.Number(2*37.186, label="VM instance cost per hour",
interactive=False, visible=False)
self.input_tokens_per_second = gr.Number(2860, visible=False,
label="Number of output tokens per second for this specific model and VM instance",
interactive=False
)
self.output_tokens_per_second = gr.Number(18.449, visible=False,
label="Number of output tokens per second for this specific model and VM instance",
interactive=False
)
self.info = gr.Markdown("To see the script used to benchmark the Llama2-70B model, [click here](https://www.cursor.so/blog/llama-inference#user-content-fn-llama-paper)", interactive=False, visible=False)
self.model.change(on_model_change, inputs=self.model, outputs=[self.vm, self.info, self.vm_cost_per_hour, self.input_tokens_per_second, self.output_tokens_per_second])
self.vm.change(on_vm_change, inputs=[self.model, self.vm], outputs=[self.vm_cost_per_hour, self.input_tokens_per_second, self.output_tokens_per_second])
self.used = gr.Slider(minimum=0.01, value=30., step=0.01, label="% used",
info="Percentage of time the GPU is used",
interactive=True,
visible=False)
def compute_cost_per_token(self, vm_cost_per_hour, input_tokens_per_second, output_tokens_per_second, used):
cost_per_input_token = vm_cost_per_hour * 100 / (3600 * used * input_tokens_per_second)
cost_per_output_token = vm_cost_per_hour * 100 / (3600 * used * output_tokens_per_second)
return cost_per_input_token, cost_per_output_token
class OpenSourceDIY(BaseTCOModel):
def __init__(self):
self.set_name("(Open source) DIY")
self.set_formula(r"""$CT = \frac{VM\_CH \times 100}{3600 \times U} \times (\frac{IT}{ITS} + \frac{OT}{OTS})$
with:
CT = Cost per Token
VM_CH = VM Cost per Hour
ITS = Input Tokens per Second
OTS = Output Tokens per Second
U = Used
IT = Input Tokens
OT = Output Tokens
""")
self.set_latency("The latency can't be estimated in the DIY scenario for the model isn't defined")
super().__init__()
def render(self):
self.info = gr.Markdown("Compute the cost/token based on our formula below, using your own parameters", visible=False)
self.display_formula = gr.Markdown(r"""$CT = \frac{VM\_CH \times 100}{3600 \times U} \times (\frac{IT}{ITS} + \frac{OT}{OTS})$
with:
CT = Cost per Token
VM_CH = VM Cost per Hour
ITS = Input Tokens per Second
OTS = Output Tokens per Second
U = Used
IT = Input Tokens
OT = Output Tokens
""", visible=False)
self.vm_cost_per_hour = gr.Number(3.5, label="VM instance cost per hour",
interactive=True, visible=False)
self.input_tokens_per_second = gr.Number(300, visible=False,
label="Number of input tokens per second processed for this specific model and VM instance",
interactive=True
)
self.output_tokens_per_second = gr.Number(300, visible=False,
label="Number of output tokens per second processed for this specific model and VM instance",
interactive=True
)
self.used = gr.Slider(minimum=0.01, value=50., step=0.01, label="% used",
info="Percentage of time the GPU is used",
interactive=True,
visible=False)
def compute_cost_per_token(self, vm_cost_per_hour, input_tokens_per_second, output_tokens_per_second, used):
cost_per_input_token = vm_cost_per_hour * 100 / (3600 * used * input_tokens_per_second)
cost_per_output_token = vm_cost_per_hour * 100 / (3600 * used * output_tokens_per_second)
return cost_per_input_token, cost_per_output_token
class CohereModel(BaseTCOModel):
def __init__(self):
self.set_name("(SaaS) Cohere")
self.set_formula(r"""$CR = \frac{CT\_1M \times (IT + OT)}{1000000}$
with:
CR = Cost per Request
CT_1M = Cost per one million Tokens (from Cohere's pricing web page)
IT = Input Tokens
OT = Output Tokens
""")
self.set_latency("")
super().__init__()
def render(self):
self.model = gr.Dropdown(["Default", "Custom"], value="Default",
label="Model",
interactive=True, visible=False)
if self.use_case == "Summarize":
self.model: gr.Dropdown.update(choices=["Default"])
elif self.use_case == "Question-answering":
self.model: gr.Dropdown.update(choices=["Default", "Custom"])
else:
self.model: gr.Dropdown.update(choices=["Default", "Custom"])
def compute_cost_per_token(self, model):
"""Cost per token = """
use_case = self.use_case
if use_case == "Generate":
if model == "Default":
cost_per_1M_tokens = 15
else:
cost_per_1M_tokens = 30
elif use_case == "Summarize":
cost_per_1M_tokens = 15
else:
cost_per_1M_tokens = 200
cost_per_input_token = cost_per_1M_tokens / 1000000
cost_per_output_token = cost_per_1M_tokens / 1000000
return cost_per_input_token, cost_per_output_token
class ModelPage:
def __init__(self, Models: BaseTCOModel):
self.models: list[BaseTCOModel] = []
for Model in Models:
model = Model()
self.models.append(model)
def render(self):
for model in self.models:
model.render()
model.register_components_for_cost_computing()
def get_all_components(self) -> list[Component]:
output = []
for model in self.models:
output += model.get_components()
return output
def get_all_components_for_cost_computing(self) -> list[Component]:
output = []
for model in self.models:
output += model.get_components_for_cost_computing()
return output
def make_model_visible(self, name:str, use_case: gr.Dropdown, num_users: gr.Number, input_tokens: gr.Slider, output_tokens: gr.Slider):
# First decide which indexes
output = []
for model in self.models:
if model.get_name() == name:
output+= [gr.update(visible=True)] * len(model.get_components())
# Set use_case and num_users values in the model
model.use_case = use_case
model.num_users = num_users
model.input_tokens = input_tokens
model.output_tokens = output_tokens
else:
output+= [gr.update(visible=False)] * len(model.get_components())
return output
def compute_cost_per_token(self, *args):
begin=0
current_model = args[-1]
for model in self.models:
model_n_args = len(model.get_components_for_cost_computing())
if current_model == model.get_name():
model_args = args[begin:begin+model_n_args]
cost_per_input_token, cost_per_output_token = model.compute_cost_per_token(*model_args)
model_tco = cost_per_input_token * model.input_tokens + cost_per_output_token * model.output_tokens
formula = model.get_formula()
latency = model.get_latency()
return f"Model {current_model} has a cost/request of: ${model_tco}", model_tco, formula, f"The average latency of this model is {latency}"
begin = begin+model_n_args