What's the expected WAN 2.2 I2V SDNQ performance on 4090?

[SAC020]

I am running a I2V on WAN 2.2 14B SDNQ, 416p, 193 frames, 24fps, 50 steps / frame, DPM++ 2M SDE Flowmatch

I am getting 36s/it with plenty of VRAM buffer (no swapping). Feels extremely slow, for example compared with framepack

Is this normal / expected performance?

(I'm on windows, recent dev build, can provide details about the exact version if needed)

08:20:35-825145 DEBUG    Image resize: source=4096:4096 target=416:416 mode="Crop" upscaler="None" type=pil time=0.09
                         fn=f:generate
08:20:35-826926 DEBUG    Video: op=I2V init=<PIL.Image.Image image mode=RGB size=4096x4096 at 0x227336524D0>
                         resized=<PIL.Image.Image image mode=RGB size=416x416 at 0x226A5829750>
08:20:36-574860 DEBUG    GC: current={'gpu': 11.39, 'ram': 33.69, 'oom': 0} prev={'gpu': 12.75, 'ram': 33.21}
                         load={'gpu': 47, 'ram': 53} gc={'gpu': 1.36, 'py': 1447} fn=f:generate why=video time=0.75
08:20:36-583553 DEBUG    Video: cls=WanImageToVideoPipeline width=416 height=416 frames=193 steps=50
08:20:36-588249 DEBUG    Sampler: "DPM2++ 2M SDE FlowMatch" class=FlowMatchDPMSolverMultistepScheduler
                         config={'num_train_timesteps': 1000, 'beta_start': 0.00085, 'beta_end': 0.012, 'beta_schedule':
                         'linear', 'solver_order': 2, 'algorithm_type': 'dpmsolver++2Msde', 'solver_type': 'midpoint',
                         'shift': 5, 'midpoint_ratio': 0.5, 's_noise': 1.0, 'use_noise_sampler': True, 'base_shift':
                         0.5, 'max_shift': 1.15, 'base_image_seq_len': 256, 'max_image_seq_len': 4096}
08:20:36-590526 INFO     Processing modifiers: apply
08:20:36-591423 INFO     Apply Cache-DiT: config="DBCACHE_F8B0_W8M0MC0_R0.08" calibrator="None"
08:20:36-606867 INFO     Base: pipeline=WanImageToVideoPipeline task=IMAGE_2_IMAGE batch=1/1x1 set={'prompt': 212,
                         'negative_prompt': 0, 'guidance_scale': 5.0, 'generator': 'cuda:[2339060185]',
                         'num_inference_steps': 50, 'num_frames': 193, 'output_type': 'pil', 'image': <PIL.Image.Image
                         image mode=RGB size=416x416 at 0x226A5829750>, 'width': 416, 'height': 416, 'parser': 'fixed'}
08:20:36-620402 DEBUG    Encode: prompt="prompt" hijack=True
08:20:57-385254 DEBUG    Encode: vae=AutoencoderKLWan slicing=True tiling=True latents=[1, 3, 193, 416,
                         416]:cuda:0:torch.bfloat16 time=10.099
Progress 36.36s/it █████████████████████████████████ 100% 50/50 30:18 00:00 Base
08:51:28-817627 DEBUG    Decode: vae=AutoencoderKLWan slicing=True tiling=True latents=[1, 16, 49, 52, 52]:cuda:0
                         dtype=torch.bfloat16 time=13.344
08:51:29-602282 DEBUG    Generated: frames=193
08:51:29-603262 INFO     Processing modifiers: unapply
08:51:33-733574 INFO     Processed: images=193 its=5.20 ops=['video']
08:51:33-734851 DEBUG    Processed: timers={'total': 1917.89, 'pipeline': 1853.0, 'vae': 23.45, 'onload': 14.47,
                         'offload': 11.92, 'te': 10.62, 'post': 4.13, 'gc': 0.29}
08:51:33-737315 DEBUG    Processed: memory={'ram': {'total': 63.92, 'rss': 24.83, 'used': 37.79, 'free': 26.13, 'avail':
                         26.13, 'buffers': 0, 'cached': 0}, 'gpu': {'used': 6.24, 'total': 23.99, 'active': 1.44,
                         'peak': 14.52, 'retries': 0, 'oom': 0, 'swap': 0}, 'job': ''}
08:51:33-738647 INFO     Video: name="WAN 2.2 A14B SDNQ I2V" cls=WanImageToVideoPipeline frames=193 time=1857.15
08:51:34-014032 DEBUG    Video: video=True export=False safetensors=False interpolate=0
08:51:34-014732 DEBUG    Video: encode=193 raw=400797696 latent=torch.Size([1, 3, 193, 416, 416]) fps=24 codec=libx264
                         ext=mp4 options="crf:16"
08:51:34-270013 INFO     Video: file="outputs\video\2025-11-07-Wan2.2-I2V-A14B-SDNQ-uint4-svd-r32-20250833.mp4"
                         codec=libx264 frames=193 width=416 height=416 fps=24 options={'crf': '16'}

[vladmandic]

that's 193/36 ~ 5.4 frames-iterations-per-second, not that bad, especially for i2v which is slower than 2tv.