DDO examples
Model-specific Modular Diffusers runners and measured comparisons using DDO, covering LTX 2.3 image, LTX 2.3 video/audio and Qwen Image 2.1.
DDO manages weight loading and offloading while Diffusers/backend code computes the model. The linear route uses PyTorch-backed offloading; the optional native VMM route uses CUDA Virtual Memory Management to keep pointer addresses stable while reusing bounded physical GPU backing safely. VMM is not activation compression or a guarantee of local-VRAM placement. See the library overview.
Use the examples
Read the documentation website, including setup and benchmarks.
Start with the user guide to install packages and choose a
model runner. This path does not require a DDO source checkout or native compiler.
DDO 0.1.0rc4 and its optional VMM allocator 0.1.0rc3 are available on PyPI.
Follow the package setup guide to install the packages and runner dependencies
in one environment.
Compare performance and memory
Start with this existing Windows LTX 2.3 video comparison: BF16, 1280×704, 5 seconds/121 frames at 24 fps, 8 steps, seed 43, RTX 3060 Ti 8 GiB and 64 GiB system RAM. All three use a shared Diffusers leaf text encoder and resident temporal VAE. The method is the transformer route. Times are seconds; memory peaks are sampled absolute driver GPU usage and process RSS in GiB.
Before comparing rows, read the labels and memory glossary for runtime labels, latency scope and peak-memory definitions.
All rows use default PyTorch SDPA for video self-attention. The internal SDPA kernel was not recorded.
| Runtime method | Denoise (s) | Video VAE decode (s) | Tracked total (s) | Peak GPU (absolute, GiB) | Peak RAM (RSS, GiB) |
|---|---|---|---|---|---|
| Diffusers streamed leaf | 392.91 | 22.52 | 452.33 | 4.72 | 41.34 |
DDO linear (one_shot_fast) |
262.22 | 23.47 | 336.44 | 4.72 | 47.87 |
DDO native VMM (vmm_fast) |
207.40 | 24.00 | 277.09 | 5.21 | 32.65 |
Video VAE decode excludes decoder loading, audio decode and vocoder work. The tracked total already includes these phases; it is not denoise plus decode.
Tracked total excludes command startup and downloads. These single observations compare configured pipelines, not the isolated effect of the allocator.
| Workload | What to compare | Full tables and conditions |
|---|---|---|
| LTX video | Denoise, decoder latency and memory; native is fastest in this primary five-second set. | Video benchmarks |
| LTX image | Linear has shorter Windows denoise; native has shorter tracked total. | Image benchmarks |
| Qwen Image 2.1 | Compare resolution, encoder/VAE routes and observation sessions before ranking methods. | Qwen benchmarks |
Each model page separates phase tables and finishes each primary platform set with a pipeline-total table. Phases can use different offload methods; Diffusers leaf VAE may reduce decode GPU peaks at extra latency. No single method wins every phase or memory metric. Read the summary, hardware and metric definitions. Native Linux and WSL results stay in separate groups. These are experimental prereleases; further model architectures and hardware need validation.
On the website, top menus select pages; the left contents list selects sections of the current page.
Develop or contribute
Use the developer guide for editable source setup, architecture, validation, documentation preview and release procedures. Generic allocation/runtime policy stay in DDO and its allocator; model math belongs to independent custom-block repositories.