[Benchmark] Inference Performance on RTX 5090 with Wan2.1-1.3B (PyTorch 2.8.0)

Hi authors,

Thanks for your great work! I reproduce this on RTX 5090, this is my script:

python turbodiffusion/inference/wan2.1_t2v_infer.py     --model Wan2.1-1.3B     --dit_path checkpoints/TurboWan2.1-T2V-1.3B-480P-quant.pth     --resolution 480p     --prompt "A stylish woman walks down a Tokyo street filled with warm glowing neon and animated city signage. She wears a black leather jacket, a long red dress, and black boots, and carries a black purse. She wears sunglasses and red lipstick. She walks confidently and casually. The street is damp and reflective, creating a mirror effect of the colorful lights. Many pedestrians walk about."     --num_samples 1     --num_steps 3     --quant_linear     --attention_type sagesla     --sla_topk 0.1

and the output:

[01-19 08:01:52|INFO|turbodiffusion/inference/wan2.1_t2v_infer.py:88:<module>] Generating with prompt: A stylish woman walks down a Tokyo street filled with warm glowing neon and animated city signage. She wears a black leather jacket, a long red dress, and black boots, and carries a black purse. She wears sunglasses and red lipstick. She walks confidently and casually. The street is damp and reflective, creating a mirror effect of the colorful lights. Many pedestrians walk about.
Sampling: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:02<00:00,  1.25it/s]
[01-19 08:01:54|INFO|turbodiffusion/inference/wan2.1_t2v_infer.py:142:<module>] Sampling completed in 2.40 seconds.

And here is how I implemented the timing measurement:

    t1 = time.time()
    for i, (t_cur, t_next) in enumerate(tqdm(list(zip(t_steps[:-1], t_steps[1:])), desc="Sampling", total=total_steps)):
        with torch.no_grad():
            v_pred = net(x_B_C_T_H_W=x.to(**tensor_kwargs), timesteps_B_T=(t_cur.float() * ones * 1000).to(**tensor_kwargs), **condition).to(
                torch.float64
            )
            x = (1 - t_next) * (x - t_cur * v_pred) + t_next * torch.randn(
                *x.shape,
                dtype=torch.float32,
                device=tensor_kwargs["device"],
                generator=generator,
            )
    t2 = time.time()
    log.info(f"Sampling completed in {t2 - t1:.2f} seconds.")

Did I perform the test correctly? I only measured the time for the DiT part. My PyTorch version is 2.8.0+cu129. The 2.4s is not the same as 1.9s as reported.

Start with turbodiffusion/inference/wan2.1_t2v_infer.py and reproduce the supplied command on an RTX 5090 with PyTorch 2.8.0+cu129. Inspect the timing block around the sampling loop and compare its 2.40-second result with the reported 1.9-second benchmark; done means determining whether the measurement procedure or environment explains the difference.