Metadata-Version: 2.4
Name: flashdreams
Version: 0.1.0rc1
Summary: High-performance streaming video diffusion framework with pluggable model backends
License-Expression: Apache-2.0
Project-URL: Homepage, https://github.com/NVIDIA/flashdreams
Project-URL: Repository, https://github.com/NVIDIA/flashdreams
Project-URL: Bug Tracker, https://github.com/NVIDIA/flashdreams/issues
Requires-Python: >=3.10
Description-Content-Type: text/markdown
Requires-Dist: boto3>=1.35
Requires-Dist: botocore>=1.35
Requires-Dist: einops>=0.7
Requires-Dist: filelock>=3
Requires-Dist: ftfy>=6.0
Requires-Dist: huggingface-hub>=0.33
Requires-Dist: loguru>=0.7
Requires-Dist: numpy>=1.24
Requires-Dist: nvidia-ml-py>=12.0
Requires-Dist: safetensors>=0.4
Requires-Dist: tqdm>=4.60
Requires-Dist: transformers<6,>=5.0
Requires-Dist: urllib3>=2.7.0
Requires-Dist: tyro>=1.0
Requires-Dist: torch>=2.9
Requires-Dist: triton-windows>=3.5.0; sys_platform == "win32"
Provides-Extra: dev
Requires-Dist: pytest>=8.0; extra == "dev"
Requires-Dist: mediapy>=1.1; extra == "dev"
Requires-Dist: pytest-manual-marker>=2.0; extra == "dev"
Requires-Dist: tomli>=2.0; extra == "dev"
Requires-Dist: transformer-engine[core-cu13,pytorch]>=2.12; sys_platform != "win32" and extra == "dev"
Provides-Extra: examples
Requires-Dist: mediapy>=1.1; extra == "examples"
Requires-Dist: opencv-python-headless>=4.5; extra == "examples"
Requires-Dist: scipy>=1.11; extra == "examples"
Provides-Extra: runners
Requires-Dist: mediapy>=1.1; extra == "runners"
Requires-Dist: opencv-python-headless>=4.5; extra == "runners"
Requires-Dist: scipy>=1.11; extra == "runners"
Provides-Extra: serving
Requires-Dist: aiohttp>=3.9; extra == "serving"
Requires-Dist: aiortc>=1.9; extra == "serving"

<!--
SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
SPDX-License-Identifier: Apache-2.0
-->

# flashdreams

A high-performance streaming inference framework for video diffusion models
with a plugin architecture for model backends.

## Features

- **Streaming Inference** -- Autoregressive chunk-wise video generation with
  per-rollout cache state for bounded VRAM and arbitrarily long rollouts
- **Plugin Architecture** -- Entry-point-based model discovery; third-party
  packages register runner configs that appear automatically in the CLI
- **Multi-GPU** -- Context parallelism via torchrun with automatic sharding
  across ranks
- **Performance** -- torch.compile support with CUDA graph capture and replay
- **Serving** -- WebRTC and gRPC integrations for real-time interactive
  applications

## Supported Models

Wan 2.1/2.2, Cosmos Predict2, and more via first-party integration packages.

## Installation

```bash
pip install flashdreams
```

## Documentation

<https://github.com/NVIDIA/flashdreams>
