vision-agents-plugins-nvidia 0.3.2

NVIDIA VLM plugin for Vision Agents

NVIDIA Plugin for Vision Agents

NVIDIA VLM integration for Vision Agents. Supports vision language models through NVIDIA's Chat Completions API with NVCF asset management.

Features

• Video understanding: Automatically buffers and forwards video frames to NVIDIA VLM models
• Streaming responses: Supports streaming text responses with real-time chunk events
• Asset management: Automatically uploads frames as assets and cleans them up after use
• Configurable frame rate and buffer duration for optimal performance

Installation

uv add vision-agents[nvidia]

Configuration

Set your NVIDIA API key:

export NVIDIA_API_KEY=your_nvidia_api_key

Usage

Basic VLM Usage

from vision_agents.plugins import nvidia

vlm = nvidia.VLM(
    model="nvidia/cosmos-reason2-8b",
    fps=1,
    frame_buffer_seconds=10,
)

# VLM automatically buffers video frames when used with an Agent
response = await vlm.simple_response("What do you see?")
print(response.text)

With Custom Configuration

from vision_agents.plugins import nvidia

vlm = nvidia.VLM(
    model="nvidia/cosmos-reason2-8b",
    api_key="your-api-key",
    fps=2,
    frame_buffer_seconds=15,
    frame_width=1280,
    frame_height=720,
    max_tokens=2048,
    temperature=0.3,
    top_p=0.8,
)

Configuration

Parameter	Description	Default	Type
model	NVIDIA model ID	"nvidia/cosmos-reason2-8b"	str
api_key	NVIDIA API token (or use NVIDIA_API_KEY env var)	None	Optional[str]
fps	Frames per second to buffer	1	int
frame_buffer_seconds	Number of seconds to buffer	10	int
frame_width	Width of video frames to send	800	int
frame_height	Height of video frames to send	600	int
max_tokens	Maximum response tokens	1024	int
temperature	Temperature for sampling	0.2	float
top_p	Top-p sampling parameter	0.7	float
frames_per_second	Frames per second for video models	8	int

Dependencies

• vision-agents: Core Vision Agents framework
• aiohttp>=3.9.0: Async HTTP client for API requests