DPMSolverMultistepScheduler out-of-bounds error when using beta_schedule="squaredcos_cap_v2" + use_karras_sigmas=True

Describe the bug

Hi, I might be misunderstanding something here — I’m not very experienced with submitting issues to Diffusers yet, so please feel free to correct me if I’ve overlooked anything.

When using DPMSolverMultistepScheduler with the combination:

DPMSolverMultistepScheduler(
    beta_schedule="squaredcos_cap_v2",
    use_karras_sigmas=True,
)

calling set_timesteps(20) produces duplicate timesteps, which causes the multistep solver’s internal index tracking to drift, eventually leading to an index out of bounds error during sampling.

Reproduction

import diffusers
import torch
from diffusers import DPMSolverMultistepScheduler

print("Diffusers version:", diffusers.__version__)
print("Torch version:", torch.__version__)

scheduler = DPMSolverMultistepScheduler(
    beta_schedule="squaredcos_cap_v2",
    use_karras_sigmas=True,
)

scheduler.set_timesteps(20)

print("Timesteps:", scheduler.timesteps)

latents = torch.randn(1, 4, 64, 64)
noise_pred = torch.randn_like(latents)

for i, t in enumerate(scheduler.timesteps):
    print(f"Step {i}: t={t.item()}")
    try:
        result = scheduler.step(noise_pred, t, latents)
        latents = result.prev_sample
    except Exception as e:
        print(f"ERROR: {e}")
        break

Logs

Diffusers version: 0.35.2
Torch version: 2.9.0+cu128
Timesteps: tensor([998, 998, 998, 998, 998, 998, 998, 998, 998, 997, 996, 994, 989, 978,
        949, 867, 623, 221,  33,   0])
Step 0: t=998
Step 1: t=998
Step 2: t=998
Step 3: t=998
Step 4: t=998
Step 5: t=998
Step 6: t=998
Step 7: t=998
Step 8: t=998
Step 9: t=997
Step 10: t=996
Step 11: t=994
Step 12: t=989
Step 13: t=978
Step 14: t=949
Step 15: t=867
Step 16: t=623
Step 17: t=221
Step 18: t=33
Step 19: t=0
ERROR: index 21 is out of bounds for dimension 0 with size 21

System Info

🤗 Diffusers version: 0.35.2
Platform: Linux-6.8.0-88-generic-x86_64-with-glibc2.39
Running on Google Colab?: No
Python version: 3.12.3
PyTorch version (GPU?): 2.9.0+cu128 (True)
Flax version (CPU?/GPU?/TPU?): not installed (NA)
Jax version: not installed
JaxLib version: not installed
Huggingface_hub version: 0.35.3
Transformers version: 4.57.1
Accelerate version: 1.10.1
PEFT version: not installed
Bitsandbytes version: not installed
Safetensors version: 0.6.2
xFormers version: not installed
Accelerator: NVIDIA GeForce RTX 4070 Ti SUPER, 16376 MiB
Using GPU in script?:
Using distributed or parallel set-up in script?:

Who can help?

No response

Fix scheduler index out of bounds error with Karras sigmas

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.

hey @scp3125

Do you have an actual use case for squaredcos_cap_v2 + use_karras_sigmas? I'm not entirely sure why squaredcos_cap_v2 was excluded from rounding in the scheduler code. I can dig into the git commits history, but any change here would require extensive testing from our end, so I want to make sure there's a real use case first.
Thanks!

hey @scp3125

Do you have an actual use case for squaredcos_cap_v2 + use_karras_sigmas? I'm not entirely sure why squaredcos_cap_v2 was excluded from rounding in the scheduler code. I can dig into the git commits history, but any change here would require extensive testing from our end, so I want to make sure there's a real use case first. Thanks!

Yes, this is a real inference use case.

import diffusers
import torch
from diffusers import DPMSolverMultistepScheduler, StableDiffusionPipeline

print("Diffusers:", diffusers.__version__)
print("Torch:", torch.__version__)

device = "cuda" if torch.cuda.is_available() else "cpu"

pipe = StableDiffusionPipeline.from_pretrained(
    "stable-diffusion-v1-5/stable-diffusion-v1-5",
    requires_safety_checker=False,
).to(device)
pipe.set_progress_bar_config(disable=True)

pipe.scheduler = DPMSolverMultistepScheduler.from_config(
    pipe.scheduler.config,
    beta_schedule="squaredcos_cap_v2",
    use_karras_sigmas=True,
)

pipe.scheduler.set_timesteps(20, device=device)
print("Timesteps:", pipe.scheduler.timesteps)

out = pipe(
    prompt=["an astronaut riding a horse on mars"],
    num_inference_steps=20,
)
print("OK, got", len(out.images), "images")

Output:

Diffusers: 0.35.2
Torch: 2.8.0
Loading pipeline components...: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 7/7 [00:00<00:00, 29.48it/s]
Timesteps: tensor([998, 998, 998, 998, 998, 998, 998, 998, 998, 997, 996, 994, 989, 978,
        949, 867, 623, 221,  33,   0])
Traceback (most recent call last):
  File "/Users/py-test/./src/demo/main.py", line 25, in <module>
    out = pipe(
          ^^^^^
  File "/Users/py-test/.venv/lib/python3.12/site-packages/torch/utils/_contextlib.py", line 120, in decorate_context
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/Users/py-test/.venv/lib/python3.12/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py", line 1061, in __call__
    latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/py-test/.venv/lib/python3.12/site-packages/diffusers/schedulers/scheduling_dpmsolver_multistep.py", line 1112, in step
    prev_sample = self.multistep_dpm_solver_second_order_update(self.model_outputs, sample=sample, noise=noise)
                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/py-test/.venv/lib/python3.12/site-packages/diffusers/schedulers/scheduling_dpmsolver_multistep.py", line 835, in multistep_dpm_solver_second_order_update
    self.sigmas[self.step_index + 1],
    ~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^
IndexError: index 21 is out of bounds for dimension 0 with size 21

Version:

- 🤗 Diffusers version: 0.35.2
- Platform: macOS-26.2-arm64-arm-64bit
- Running on Google Colab?: No
- Python version: 3.12.11
- PyTorch version (GPU?): 2.8.0 (False)
- Flax version (CPU?/GPU?/TPU?): not installed (NA)
- Jax version: not installed
- JaxLib version: not installed
- Huggingface_hub version: 0.35.3
- Transformers version: 4.57.3
- Accelerate version: 1.10.1
- PEFT version: 0.18.0
- Bitsandbytes version: not installed
- Safetensors version: 0.6.2
- xFormers version: not installed
- Accelerator: Apple M4 Pro
- Using GPU in script?: <fill in>
- Using distributed or parallel set-up in script?: <fill in>

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.

fix: deduplicate timesteps in DPMSolverMultistepScheduler to prevent …

fix: deduplicate timesteps in DPMSolverMultistep for squaredcos_cap_v…

@yiyixuxu

@yiyixuxu on the use-case question: kandinsky-community/kandinsky-3 ships DDPMScheduler with beta_schedule="squaredcos_cap_v2", so swapping in a faster solver the usual way, X.from_config(pipe.scheduler.config, use_karras_sigmas=True), hits this.

It isn't limited to DPMSolverMultistepScheduler. With that config and 25 steps on current main:

Scheduler	Duplicated timesteps	Result
DPMSolverMultistep	11	`IndexError: index 26 is out of bounds`
UniPCMultistep	10	`AssertionError`
DEISMultistep	10	`IndexError: index 26 is out of bounds`
DPMSolverSinglestep	10	`IndexError: list index out of range`
SASolver	10	`AssertionError`

In each case the Karras sigmas map to near-identical timesteps at the top of the cosine schedule, and the step index drifts past the end. #13240 only touches DPMSolverMultistepScheduler.

Would you take a fix that covers all five, with a test running each of them on the cosine + Karras config? Happy to put it together if that's the direction you want.

Found while property-testing the schedulers; prepared with AI assistance, reproduced and checked by me.

Navigation Menu

Repository navigation

DPMSolverMultistepScheduler out-of-bounds error when using beta_schedule="squaredcos_cap_v2" + use_karras_sigmas=True #12771

Describe the bug

Reproduction

Logs

System Info

Who can help?

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

DPMSolverMultistepScheduler out-of-bounds error when using beta_schedule="squaredcos_cap_v2" + use_karras_sigmas=True #12771

Description

Describe the bug

Reproduction

Logs

System Info

Who can help?

Activity

github-actions commented on Jan 9, 2026

yiyixuxu commented on Jan 10, 2026

dylanzhao25 commented on Jan 11, 2026

github-actions commented on Feb 4, 2026

xThreeh commented on Oct 7, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions