Presets And Platforms

DDO presets are route policies, not model-specific hacks. The same preset can apply different routes to different component names.

Presets

Preset	Recommended for	Behavior
auto	Default user-facing preset	Resolves to the current recommended policy. Today that is one_shot_fast plus platform and backend safety checks.
one_shot_fast	BF16/FP16 dense transformer weights on low-VRAM Windows or comparable systems	Uses DDO dynamic offload for standard dense transformer linears, selected resident modules, and RAM-aware pinned CPU weights.
low_ram_safe	Lower accelerator pressure	Uses a smaller resident-module budget and avoids expensive RAM assumptions. Slower but safer.
wsl_compat	WSL stacks where pinned memory or stream behavior is unstable	Disables pinned CPU memory and risky stream behavior.
warm_process	Server-style repeated generation	Keeps process-lifetime caches where possible. Compare separately from cold one-shot runs.
diffusers_offload_compat	Official compatibility baseline	Routes components to Diffusers block-level group offload.
diffusers_leaf_offload_compat	Quantized backends and fine-grained official offload	Routes components to Diffusers leaf-level group offload.
off	Manual control	DDO does not apply offload hooks.

Component Policy

The table above and default component policy below describe the linear/compat family. Explicit --ddo-backend vmm --ddo-preset auto selects vmm_fast; vmm/nvme_baseline selects the protected legacy VMM route. Selected native encoder/transformer staged workloads passed on Windows and native Linux, but whole-component encoder ownership remains opt-in. See native VMM usage.

Typical policy shape:

Component	Dense BF16 default	Quantized/backend-specific default
transformer / unet	DDO dynamic offload when beneficial	Diffusers group offload compatibility route
text_encoder	Diffusers group offload	Diffusers group offload
vae	Diffusers group offload or manual CUDA decode stage	Diffusers group offload or manual CUDA decode stage
small connectors / adapters	Diffusers group offload or caller-managed CUDA move	Diffusers group offload or caller-managed CUDA move

Platform Notes

The VAE now defaults to no-stream leaf for one_shot_fast/auto, low_ram_safe and resolved vmm_fast settings. Other selected presets preserve historical resident staged-host behavior. Explicit per-component configs/routes remain authoritative. See component strategies for the exact table, convolution opt-in and acceptance scope.

Windows:

• auto / one_shot_fast is the main fast path from the current Windows BF16 benchmark results.
• Windows standby-list purge can help between phase-staged execution phases.
• The planner pins CPU weights only when available system RAM leaves configured headroom.

WSL:

• Prefer wsl_compat first.
• The measured WSL stack benefited from DDO dynamic offload with pinning disabled.
• Official Diffusers leaf group offload was unstable in one measured WSL run.

Native Linux:

• Compare auto with diffusers_leaf_offload_compat.
• DDO can improve denoise throughput, but official Diffusers leaf group offload may win cold one-shot totals when setup is much lower.
• Warm/server runs can change this tradeoff because DDO setup caches amortize across requests.

Quantized backends:

• Prefer diffusers_leaf_offload_compat or diffusers_offload_compat.
• DDO detects SDNQ-style packed modules and preserves their forward.
• Dense-linear DDO acceleration should only patch standard 2D nn.Linear weights.