Qwen Image 2.1 — measurement details

Back to comparison tables. Scenario links in those tables open the sections below; no local report download is required to read the results.

Labels and units: read the labels and memory glossary before comparing runtime names, phase timings, tracked totals or memory peaks.

Workload and measurement scope

Base BF16, 512×512, 20 steps, seed 42, capybara prompt, CFG 1 and KV cache. Shared official leaf text encoder; resident untiled VAE on Windows/Linux, untiled non-streamed leaf VAE on WSL.

Checkpoint: Qwen/Qwen-Image-2.1, snapshot d26bb61231c349cf6b7896fa83353113880e1ba3. Official Diffusers 0.41.0 in the primary platform comparisons. The prompt is "A capybara wearing a wizard hat, reading by candlelight, oil painting". CFG 1, KV cache enabled. No compute compilation or local VAE math changes in the primary 512-square cases. The separate 1024-square cases retain their explicit local VAE/cleanup settings.

Text encoder load, attachment, encode and cleanup are separate monitored phases. The text-encode number excludes the neighboring load/cleanup rows. Denoise includes the runner's completion/CPU latent readback boundary. Transformer attachment is separate from denoise. VAE loading is separate from the decode phase.

All primary controls are staged modular runners. The official leaf control is not the older monolithic pipeline whose encoder/denoise/decode could not be separated. Baseline-relative console VRAM must not replace absolute driver peaks.

All displayed phase rows are existing observations, not new runs or inferred phase estimates. Memory comes from each recorded phase's peak_vram_absolute_gb and peak_ram_gb with ram_metric=rss. Times use elapsed_sec; global peaks and tracked totals come from the same report as that row. Values are rounded to two decimals for display only. Missing fields remain unavailable. Phase tables show selected boundaries, not an additive decomposition; do not add aggregates to their child events or sum memory peaks. Host callbacks do not establish completed warm GPU timing. See metric definitions.

Platform conditions

Windows comparisons recorded successful standby purge before the transformer; phase-local purge can also be included in tracked totals. Linux and WSL do not execute the Windows purge. WSL auxiliary leaf hooks are non-streamed and native registration is bounded; those routes are not equivalent to Windows/Linux streamed controls. Do not treat different platform observations as causal A/Bs.

Linux protected NVMe observations dated 2026-10-13 use driver 595.99.02 and a fresh post-reboot/cache session, unlike the earlier route trios. This prevents attributing their timing difference solely to the allocator. Primary cases are single observations, not repeated statistical estimates or low-RAM certification. The compared completed outputs matched within the recorded route sets; this is not a universal quality or cross-platform byte-identity claim.

BF16 scenarios

Windows / Official leaf

The diffusers_leaf_offload_compat control executes official Diffusers leaf hooks, not DDO dynamic denoising.

Windows / one_shot_fast

--ddo-backend linear --ddo-preset one_shot_fast executes linear_runtime.

Windows / low_ram_safe

--ddo-backend linear --ddo-preset low_ram_safe executes linear_runtime.

Windows / vmm_fast (reused observation)

--ddo-backend vmm --ddo-preset auto resolves to vmm_fast, executing vmm_fast_runtime.

Fixed transformer capacity: 4 GiB active weights; 640 MiB total staging. Retained source / registration ceilings: 16 / 16 GiB. Native allocation is process-wide: encoder/VAE driver baselines can differ even when their component offload policies are unchanged. This is the reused official-package native observation, not a repeat alongside the Windows leaf/linear runs. Matching workload fields and output do not prove identical source-checkout or driver/cache state.

Windows / nvme_baseline

--ddo-backend vmm --ddo-preset nvme_baseline executes protected vmm_runtime.

Fixed transformer capacity: 4 GiB active weights; two 640 MiB staging slots. Retained source / registration ceilings: 0 / 0 GiB. Normal PyTorch allocation plus native weight arena, not the process-wide native Torch allocator. Zero persistent source cache is not zero OS file-cache usage.

Linux / Official leaf

The diffusers_leaf_offload_compat control executes official Diffusers leaf hooks, not DDO dynamic denoising.

Linux / one_shot_fast

--ddo-backend linear --ddo-preset one_shot_fast executes linear_runtime.

Linux / vmm_fast

--ddo-backend vmm --ddo-preset auto resolves to vmm_fast, executing vmm_fast_runtime.

Fixed transformer capacity: 4 GiB active weights; 640 MiB total staging. Retained source / registration ceilings: 16 / 16 GiB. Native allocation is process-wide: encoder/VAE driver baselines can differ even when their component offload policies are unchanged.

Linux / nvme_baseline

--ddo-backend vmm --ddo-preset nvme_baseline executes protected vmm_runtime.

Fixed transformer capacity: 4 GiB active weights; two 448 MiB staging slots. Retained source / registration ceilings: 0 / 0 GiB. Normal PyTorch allocation plus native weight arena, not the process-wide native Torch allocator. Zero persistent source cache is not zero OS file-cache usage.

WSL / Official leaf (no stream)

The diffusers_leaf_offload_compat control executes official Diffusers leaf hooks, not DDO dynamic denoising.

WSL / wsl_compat (3 GiB override)

--ddo-backend linear --ddo-preset wsl_compat executes linear_runtime.

The resident-module budget is explicitly 3 GiB, CPU pinning is disabled and automatic budgeting is off. This validates the override, not preset-default capacity; a configured ceiling is not measured residency.

WSL / vmm_fast (bounded)

--ddo-backend vmm --ddo-preset auto resolves to vmm_fast, executing vmm_fast_runtime.

Fixed transformer capacity: 4 GiB active weights; 384 MiB total staging. Retained source / registration ceilings: 16 / 0.5 GiB. Native allocation is process-wide: encoder/VAE driver baselines can differ even when their component offload policies are unchanged.

WSL / nvme_baseline

--ddo-backend vmm --ddo-preset nvme_baseline executes protected vmm_runtime.

Fixed transformer capacity: 4 GiB active weights; two 448 MiB staging slots. Retained source / registration ceilings: 0 / 0 GiB. Normal PyTorch allocation plus native weight arena, not the process-wide native Torch allocator. Zero persistent source cache is not zero OS file-cache usage.

Windows 1024×1024 — separate VAE settings

Base BF16, 20 steps, seed 42; untiled leaf VAE. Native uses explicit cache trimming; this is not an allocator-only A/B.

Windows / one_shot_fast / leaf VAE

Same base model, prompt, 20 steps and seed 42 as the 512-square workload, but resolution is 1024×1024. Leaf untiled VAE uses explicitly accepted local options; do not attribute their effects to transformer offload alone.

Windows / vmm_fast / leaf VAE + trim

Same base model, prompt, 20 steps and seed 42 as the 512-square workload, but resolution is 1024×1024. Leaf untiled VAE uses explicitly accepted local options; do not attribute their effects to transformer offload alone. This case explicitly trims native idle cache before VAE. Native trim and cuBLAS workspace clearing are not implied by selecting a VAE preset.

Windows SDNQ INT4

Base 512×512, 20 steps, seed 42; BF16 auxiliaries, resident stock VAE. Output parity is within the INT4 pair, not a BF16 quality-equivalence claim.

Windows / SDNQ / one_shot_fast

Saved checkpoint OzzyGT/Qwen_Image_2_1_sdnq_dynamic_4bit, snapshot 59d4c3610379d91fb023c4d0b6527c76755491c4, SDNQ 0.2.7 and official Diffusers 0.41.0. BF16 auxiliaries and stock resident VAE remain unchanged. one_shot_fast executes packed sdnq_runtime, preserving SDNQ math. Normal PyTorch allocation, not native VMM. Outputs match within the INT4 pair; BF16-versus-INT4 quality equivalence was not established.

Windows / SDNQ / official leaf

Saved checkpoint OzzyGT/Qwen_Image_2_1_sdnq_dynamic_4bit, snapshot 59d4c3610379d91fb023c4d0b6527c76755491c4, SDNQ 0.2.7 and official Diffusers 0.41.0. BF16 auxiliaries and stock resident VAE remain unchanged. diffusers_leaf_offload_compat executes official streamed leaf with SDNQ math. Normal PyTorch allocation, not native VMM. Outputs match within the INT4 pair; BF16-versus-INT4 quality equivalence was not established.

Reproduce this workload

Use the runner guide and comparison controls, retaining the model, schedule, inputs, component routes and the scenario-specific capacities above. Runner defaults or a different custom-block revision do not guarantee equivalent conditions. Review trusted custom-block code and keep platform environments separate. Package/build validation and raw engineering investigations are developer concerns, not additional timing rows or prerequisites for interpreting these tables.