pipeline/models.py

import os
from dataclasses import dataclass


@dataclass(unsafe_hash=True)
class Model(object):
    model_display_name: str
    model_name: str
    api_url: str
    provider: str
    hourly_cost: int = None
    cost_description: str = None
    supports_functions: str = False
    size_billion_parameters: int = None  # in billion paramters
    cost_per_million_tokens: int = None
    cost_per_million_input_tokens: int = None
    cost_per_million_output_tokens: int = None
    input_size: int = None

    def __post_init__(self):
        self.cost_per_million_input_tokens = self.cost_per_million_input_tokens or self.cost_per_million_tokens
        self.cost_per_million_output_tokens = self.cost_per_million_output_tokens or self.cost_per_million_tokens

    @property
    def cost(self):
        if self.cost_description:
            return self.cost_description
        if self.hourly_cost:
            return f"${self.hourly_cost:.2g} / hour"
        if self.cost_per_million_tokens:
            return f"${self.cost_per_million_tokens:.2g} / 1M tokens"
        elif self.cost_per_million_input_tokens and self.cost_per_million_output_tokens:
            return f"${self.cost_per_million_input_tokens:.2g} / 1M input tokens, ${self.cost_per_million_output_tokens:.2g} / 1M output tokens"


env = os.environ.get

MODELS = [
    # source: https://openai.com/pricing
    # converted costs from dollar/1K tokens to dollar/1M for readability and together_ai comparability
    Model(
        "gpt-3.5-turbo",
        "gpt-3.5-turbo",
        None,
        "OpenAI",
        supports_functions=True,
        cost_per_million_input_tokens=1,
        cost_per_million_output_tokens=2,
        # https://learn.microsoft.com/en-us/answers/questions/1356487/what-is-the-exact-maximum-input-tokens-of-azure-gp
        input_size=4096
    ),
    Model(
        "gpt-4-turbo",
        "gpt-4-1106-preview",
        None,
        "OpenAI",
        supports_functions=True,
        cost_per_million_input_tokens=10,
        cost_per_million_output_tokens=30,
        # https://writesonic.com/blog/gpt-4-turbo-vs-gpt-4
        input_size=128_000,
    ),
    Model(
        "gpt-4",
        "gpt-4",
        None,
        "OpenAI",
        supports_functions=True,
        cost_per_million_input_tokens=30,
        cost_per_million_output_tokens=60,
        input_size=32_000,
    ),
    # source: https://www.together.ai/pricing
    Model(
        "llama-2-70b-chat",
        "together_ai/togethercomputer/llama-2-70b-chat",
        None,
        "Together AI",
        cost_per_million_tokens=0.9,
        size_billion_parameters=70,
        # https://github.com/facebookresearch/llama/issues/148
        input_size=2048,
    ),
    Model(
        "Mixtral-8x7B-Instruct-v0.1",
        "together_ai/mistralai/Mixtral-8x7B-Instruct-v0.1",
        None,
        "Together AI",
        size_billion_parameters=8 * 7,
        cost_per_million_tokens=0.9,
    ),
    # taken from endpoint pages
    Model(
        "zephyr-7b-beta",
        "huggingface/HuggingFaceH4/zephyr-7b-beta",
        env("ZEPHYR_7B_BETA_URL"),
        "Hugging Face Inference Endpoint",
        hourly_cost=1.30,
        size_billion_parameters=7,
    ),
    Model(
        "Mistral-7B-Instruct-v0.2",
        "huggingface/mistralai/Mistral-7B-Instruct-v0.2",
        env("MISTRAL_7B_BETA_URL"),
        "Hugging Face Inference Endpoint",
        hourly_cost=1.30,
        size_billion_parameters=7,
    ),
    Model(
        "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
        "huggingface/TinyLlama/TinyLlama-1.1B-Chat-v1.0",
        env("TINY_LLAMA_URL"),
        "Hugging Face Inference Endpoint",
        hourly_cost=0.60,
        size_billion_parameters=1.1,
    ),
    Model(
        "gemini-pro",
        "gemini-pro",
        None,
        "Google VertexAI",
        # https://ai.google.dev/pricing
        cost_description="$0.25 / 1M input characters, $0.5 / 1M output characters (60 queries per minute are free)",
        cost_per_million_input_tokens=0.25,
        cost_per_million_output_tokens=0.5,
    ),
    Model(
        "chat-bison (PaLM 2)",
        "chat-bison",
        None,
        "Google VertexAI",
        # https://cloud.google.com/vertex-ai/docs/generative-ai/pricing
        cost_per_million_input_tokens=0.25,
        cost_per_million_output_tokens=0.5,
        # https://ai.google.dev/models/palm
        input_size=8196,
    ),
    Model(
        "chat-bison-32k (PaLM 2 32K)",
        "chat-bison-32k",
        None,
        "Google VertexAI",
        # https://cloud.google.com/vertex-ai/docs/generative-ai/pricing
        cost_per_million_input_tokens=0.25,
        cost_per_million_output_tokens=0.5,
    ),
]

MODELS = [model for model in MODELS 
          if model.model_name=="together_ai/mistralai/Mixtral-8x7B-Instruct-v0.1"
          or model.model_name=="huggingface/HuggingFaceH4/zephyr-7b-beta"]