Qwen Image 2.1

Model and requirements

The runners use Qwen Image 2.1 custom blocks with official Diffusers. Complete installation first.

Available runners

Inspect available options:

python -B scripts/qwen_image_21/run_dynamic_modular.py --help

Run the model

Linear DDO: one_shot_fast

This base recipe fixes the 512-square comparison workload:

python -B scripts/qwen_image_21/run_dynamic_modular.py --variant base --width 512 --height 512 --steps 20 --seed 42 --ddo-preset one_shot_fast --text-encoder-preset diffusers_leaf_offload_compat --vae-resident --purge-before-transformer --memory-monitor rss --show-metrics --save-metrics

It uses the default capybara prompt. Run elevated on Windows for standby purge; outside Windows purge is a no-op. Exact platform comparisons and native recipes are in Qwen benchmarks.

Native VMM: vmm/auto on Windows

After installing the native allocator through the setup guide, run:

python -B scripts/qwen_image_21/run_dynamic_modular_vmm.py `
  --width 512 --height 512 --steps 20 --seed 42 `
  --ddo-vmm-resident-weight-budget-gb 4 `
  --ddo-vmm-total-staging-gb 0.625 `
  --ddo-vmm-source-cache-gb 16 `
  --ddo-vmm-source-registered-gb 16 `
  --ddo-vmm-execution-backing-reclaim `
  --vae-resident --purge-before-transformer `
  --memory-monitor rss --show-profile --show-metrics --save-metrics

This usage recipe explicitly enables execution backing reclaim. That opt-in does not change the recorded conditions or results of historical benchmark rows.

The runner defaults to --ddo-backend vmm --ddo-preset auto --variant base. This selects vmm_fast_runtime, using the same default capybara prompt, BF16, CFG 1 and enabled KV cache. The runner stages the text encoder with official Diffusers leaf offload; the command explicitly selects a resident, untiled VAE. The capacities match the recorded Windows native workload: 4 GiB of active weights, 640 MiB of total host staging and separate 16 GiB source-storage and CUDA-registration ceilings. These are separate budgets, not a total VRAM limit.

Platform-specific conditions and the protected vmm/nvme_baseline comparison remain documented in measurement details. The SDNQ runner is a separate route and does not use this native recipe.

The public image VMM runners share one DDO parameter vocabulary and defaults table. Specify total staging capacity in GiB. Native execution investigations use the separate diagnostic entry.

Turbo uses its saved sampling schedule, not an arbitrary base step count. Do not substitute Turbo in a base benchmark.

Runtime and preset selection

See presets. The encoder and VAE can use different strategies from the transformer. --vae-resident matches a resident-VAE control; --vae-leaf-offload favors reduced weight residency. If VAE activations still exceed memory, try --vae-tiling.

Optional weight formats

Use the separate SDNQ INT4 guide for saved quantized weights. This is not a native VMM or general FP8 recipe.

Metrics and profile

Use --show-profile --show-metrics --save-metrics where supported. Reports and media are saved under --output-dir; the runner prints their paths. Select --memory-monitor rss to match the RSS benchmark tables.

Runner-specific parameters

Linear capacity profile

Create a profile for the BF16 base workload below. This performs 20 real denoise steps and stops before VAE loading, decode and image export. Run elevated on Windows for the requested purge.

python -B scripts/qwen_image_21/run_dynamic_modular.py `
  --ddo-backend linear --ddo-preset one_shot_fast `
  --variant base --width 1024 --height 1024 --steps 20 --seed 42 `
  --ddo-profile qwen21-base-bf16-1024 --build-ddo-profile `
  --purge-before-transformer --memory-monitor rss --save-metrics `
  --output-dir outputs/qwen_linear_profile

For normal generation, repeat the same recipe without --build-ddo-profile and use a separate output directory. Turbo needs its own profile name and retains its saved checkpoint schedule; --steps does not replace Turbo sigmas. See profile storage and scope.

See Qwen parameters and shared controls.