PNDMScheduler Inference Timesteps Bug => Repeated Inference Timestep

[liming-ai]

Describe the bug

from diffusers import PNDMScheduler

scheduler = PNDMScheduler.from_pretrained('runwayml/stable-diffusion-v1-5', subfolder="scheduler")
scheduler.set_timesteps(20)
print(scheduler.timesteps)

# there are 21 timesteps with repeated step
# tensor([951, 901, 901, 851, 801, 751, 701, 651, 601, 551, 501, 451, 401, 351,
#         301, 251, 201, 151, 101,  51,   1])

Reproduction

from diffusers import PNDMScheduler

scheduler = PNDMScheduler.from_pretrained('runwayml/stable-diffusion-v1-5', subfolder="scheduler")
scheduler.set_timesteps(20)
print(scheduler.timesteps)

# there are 21 timesteps with repeated step
# tensor([951, 901, 901, 851, 801, 751, 701, 651, 601, 551, 501, 451, 401, 351,
#         301, 251, 201, 151, 101,  51,   1])

Logs

No response

System Info

• 🤗 Diffusers version: 0.30.0.dev0
• Platform: Linux-5.15.120.bsk.6-amd64-x86_64-with-glibc2.31
• Running on Google Colab?: No
• Python version: 3.9.2
• PyTorch version (GPU?): 2.1.0+cu122 (True)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Huggingface_hub version: 0.24.2
• Transformers version: 4.31.0
• Accelerate version: 0.30.1
• PEFT version: 0.11.1
• Bitsandbytes version: not installed
• Safetensors version: 0.4.3
• xFormers version: not installed

Who can help?

No response

[a-r-r-o-w]

Could you provide more context as to why you think this is incorrect? If possible, could you compare against the original implementation and verify if the Diffusers behaviour is indeed different from the original formulation? If yes, feel free to PR and tag @yiyixuxu