跳到主要内容

Nvidia NIM

https://docs.api.nvidia.com/nim/reference/

提示

我们支持所有 Nvidia NIM 模型，只需在发送 litellm 请求时，将 model=nvidia_nim/<any-model-on-nvidia_nim> 设置为前缀即可

属性详情
描述Nvidia NIM 是一个提供简单 API 用于部署和使用 AI 模型的平台。LiteLLM 支持来自 Nvidia NIM 的所有模型
LiteLLM 上的提供商路由nvidia_nim/
提供商文档Nvidia NIM 文档 ↗
提供商 API 端点https://integrate.api.nvidia.com/v1/
支持的 OpenAI 端点/chat/completions, /completions, /responses, /embeddings

API 密钥​

# env variable
os.environ['NVIDIA_NIM_API_KEY'] = ""
os.environ['NVIDIA_NIM_API_BASE'] = "" # [OPTIONAL] - default is https://integrate.api.nvidia.com/v1/

使用示例​

from litellm import completion
import os

os.environ['NVIDIA_NIM_API_KEY'] = ""
response = completion(
    model="nvidia_nim/meta/llama3-70b-instruct",
    messages=[
        {
            "role": "user",
            "content": "What's the weather like in Boston today in Fahrenheit?",
        }
    ],
    temperature=0.2,        # optional
    top_p=0.9,              # optional
    frequency_penalty=0.1,  # optional
    presence_penalty=0.1,   # optional
    max_tokens=10,          # optional
    stop=["\n\n"],          # optional
)
print(response)

使用示例 - 流式​

from litellm import completion
import os

os.environ['NVIDIA_NIM_API_KEY'] = ""
response = completion(
    model="nvidia_nim/meta/llama3-70b-instruct",
    messages=[
        {
            "role": "user",
            "content": "What's the weather like in Boston today in Fahrenheit?",
        }
    ],
    stream=True,
    temperature=0.2,        # optional
    top_p=0.9,              # optional
    frequency_penalty=0.1,  # optional
    presence_penalty=0.1,   # optional
    max_tokens=10,          # optional
    stop=["\n\n"],          # optional
)

for chunk in response:
    print(chunk)

用法 - embedding​

import litellm
import os

response = litellm.embedding(
    model="nvidia_nim/nvidia/nv-embedqa-e5-v5",               # add `nvidia_nim/` prefix to model so litellm knows to route to Nvidia NIM
    input=["good morning from litellm"],
    encoding_format = "float", 
    user_id = "user-1234",

    # Nvidia NIM Specific Parameters
    input_type = "passage", # Optional
    truncate = "NONE" # Optional
)
print(response)

用法 - LiteLLM 代理服务器​

以下是使用 LiteLLM 代理服务器调用 Nvidia NIM 端点的方法

  1. 修改 config.yaml

    model_list:
      - model_name: my-model
        litellm_params:
          model: nvidia_nim/<your-model-name>  # add nvidia_nim/ prefix to route as Nvidia NIM provider
          api_key: api-key                 # api key to send your model
         # api_base: "" # [OPTIONAL] - default is https://integrate.api.nvidia.com/v1/

  1. 启动代理

    $ litellm --config /path/to/config.yaml

  2. 向 LiteLLM 代理服务器发送请求

    import openai
    client = openai.OpenAI(
        api_key="sk-1234",             # pass litellm proxy key, if you're using virtual keys
        base_url="http://0.0.0.0:4000" # litellm-proxy-base url
    )

    response = client.chat.completions.create(
        model="my-model",
        messages = [
            {
                "role": "user",
                "content": "what llm are you"
            }
        ],
    )

    print(response)

支持的模型 - 💥 支持所有 Nvidia NIM 模型！​

我们支持所有 nvidia_nim 模型，只需在发送补全请求时，将 nvidia_nim/ 设置为前缀即可

模型名称函数调用
nvidia/nemotron-4-340b-rewardcompletion(model="nvidia_nim/nvidia/nemotron-4-340b-reward", messages)
01-ai/yi-largecompletion(model="nvidia_nim/01-ai/yi-large", messages)
aisingapore/sea-lion-7b-instructcompletion(model="nvidia_nim/aisingapore/sea-lion-7b-instruct", messages)
databricks/dbrx-instructcompletion(model="nvidia_nim/databricks/dbrx-instruct", messages)
google/gemma-7bcompletion(model="nvidia_nim/google/gemma-7b", messages)
google/gemma-2bcompletion(model="nvidia_nim/google/gemma-2b", messages)
google/codegemma-1.1-7bcompletion(model="nvidia_nim/google/codegemma-1.1-7b", messages)
google/codegemma-7bcompletion(model="nvidia_nim/google/codegemma-7b", messages)
google/recurrentgemma-2bcompletion(model="nvidia_nim/google/recurrentgemma-2b", messages)
ibm/granite-34b-code-instructcompletion(model="nvidia_nim/ibm/granite-34b-code-instruct", messages)
ibm/granite-8b-code-instructcompletion(model="nvidia_nim/ibm/granite-8b-code-instruct", messages)
mediatek/breeze-7b-instructcompletion(model="nvidia_nim/mediatek/breeze-7b-instruct", messages)
meta/codellama-70bcompletion(model="nvidia_nim/meta/codellama-70b", messages)
meta/llama2-70bcompletion(model="nvidia_nim/meta/llama2-70b", messages)
meta/llama3-8bcompletion(model="nvidia_nim/meta/llama3-8b", messages)
meta/llama3-70bcompletion(model="nvidia_nim/meta/llama3-70b", messages)
microsoft/phi-3-medium-4k-instructcompletion(model="nvidia_nim/microsoft/phi-3-medium-4k-instruct", messages)
microsoft/phi-3-mini-128k-instructcompletion(model="nvidia_nim/microsoft/phi-3-mini-128k-instruct", messages)
microsoft/phi-3-mini-4k-instructcompletion(model="nvidia_nim/microsoft/phi-3-mini-4k-instruct", messages)
microsoft/phi-3-small-128k-instructcompletion(model="nvidia_nim/microsoft/phi-3-small-128k-instruct", messages)
microsoft/phi-3-small-8k-instructcompletion(model="nvidia_nim/microsoft/phi-3-small-8k-instruct", messages)
mistralai/codestral-22b-instruct-v0.1completion(model="nvidia_nim/mistralai/codestral-22b-instruct-v0.1", messages)
mistralai/mistral-7b-instructcompletion(model="nvidia_nim/mistralai/mistral-7b-instruct", messages)
mistralai/mistral-7b-instruct-v0.3completion(model="nvidia_nim/mistralai/mistral-7b-instruct-v0.3", messages)
mistralai/mixtral-8x7b-instructcompletion(model="nvidia_nim/mistralai/mixtral-8x7b-instruct", messages)
mistralai/mixtral-8x22b-instructcompletion(model="nvidia_nim/mistralai/mixtral-8x22b-instruct", messages)
mistralai/mistral-largecompletion(model="nvidia_nim/mistralai/mistral-large", messages)
nvidia/nemotron-4-340b-instructcompletion(model="nvidia_nim/nvidia/nemotron-4-340b-instruct", messages)
seallms/seallm-7b-v2.5completion(model="nvidia_nim/seallms/seallm-7b-v2.5", messages)
snowflake/arcticcompletion(model="nvidia_nim/snowflake/arctic", messages)
upstage/solar-10.7b-instructcompletion(model="nvidia_nim/upstage/solar-10.7b-instruct", messages)