Update sd_samplers_cfg_denoiser.py

[Kittensx]

This commit introduces a safeguard to ensure that schedulers and samplers strictly adhere to the requested step count during iterative processing. The change prevents the system from exceeding the total number of steps, avoiding potential errors, inconsistencies, or unexpected behavior caused by unintended step overflows.

A check is added to verify that the current step (self.step) does not exceed or equal the total allowed steps (self.total_steps). If the condition is met, an InterruptedException is raised, gracefully halting further processing.

Description

• a simple description of what you're trying to accomplish
• a summary of changes in code
• which issues it fixes, if any

Screenshots/videos:

Checklist:

• I have read contributing wiki page
• I have performed a self-review of my own code
• My code follows the style guidelines
• My code passes tests

[Kittensx]

I found this error while working with an advanced scheduler - though I fixed the code in the scheduler, I thought this would be easily fixed with a safeguard in this file, which is the next file up from parse_prompter.py. During the loop, if a scheduler or sampler modifies the steps (accidentally or on purpose), instead of erroring out it simply drops any steps that exceed the requested user steps and finishes the image.

[w-e-w]

this change directly breaks Sampling method: DPM adaptive
which is by design to have a non-deterministic amount of steps

---

lets suppose this doesn't break stuff like DPM adaptive, I still don't really see a reason to add the safeguard
as this really is not a common issue, it's only happens if you're working with some advanced things
and the other side of "safeguarding" is "limiting possibilities"
it is totally possible that some extension may wish to developerly modify the steps was not updating the total steps, sp this safeguard will break those extensions

in my opinion, fixing the thing that is actually causing issue is better than adding unnecessary safeguards that potentially cause more issues of its own

[Kittensx]

this change directly breaks Sampling method: DPM adaptive which is by design to have a non-deterministic amount of steps

lets suppose this doesn't break stuff like DPM adaptive, I still don't really see a reason to add the safeguard as this really is not a common issue, it's only happens if you're working with some advanced things and the other side of "safeguarding" is "limiting possibilities" it is totally possible that some extension may wish to developerly modify the steps was not updating the total steps, sp this safeguard will break those extensions

in my opinion, fixing the thing that is actually causing issue is better than adding unnecessary safeguards that potentially cause more issues of its own

It's true that advanced things like custom schedulers may modify steps without actually using that terminology "self.step" or self.steps" by manipulating the sigmas_karras in a loop without knowingly doing so. If DPM adaptive works by manipulating the sigmas instead of manipulating the steps, which would make the most sense, then it wouldn't effect it. For example, in my scheduler, I create a loop and inside it I manipulate things,

` sigmas_karras = get_sigmas_karras(n=n, sigma_min=sigma_min, sigma_max=sigma_max, device=device)
...

 target_length = min(len(sigmas_karras), len(sigmas_exponential))  
sigmas_karras = sigmas_karras[:target_length]
sigmas_exponential = sigmas_exponential[:target_length]

...

for i in range(len(sigmas_karras)): .....which in effect is equal to the number of steps the user wants....
nowhere in the code above (used in my scheduler) do i reference steps...the original code worked until I created a class and then it needed n, and then I needed to find how to reference n in your code, and I figured it out but at first setting n to an arbitrary number to make the code work when using the scheduler, it would force exactly those numbers of steps...versus what the user specified...

Hence....safety.

You don't want to ever introduce the ability for new code to contradict what the user specified in the gui. I would think.

You mentioned DPM adaptive not working when introducing this code. I would argue in that case, that DPM adaptive could be structured in a way which doesn't interfere with self.step or self.steps.

If code is used to introduce a new sampler or scheduler which forces a particular number of steps, let's say 30 steps, and the user specifies 50 steps, then you'll get a bad image, because the noise generated from 50 steps is different than 30 steps and the program thinks it has time to get there...but if the program says that n=30 insteaad of the 50 that the person requested then they could think the program is bad....or if they set the number of steps to 1 or basically anything less than 20 for most things.... then no matter what the person puts 35, 40, 45 etc. you'll only get 1 step completed because you let a program modify the steps directly in code.

A safeguard stops such programs in their tracks. You could even add a helpful message like:
if self.step >= self.total_steps: print(f"Warning: Step {self.steps} exceeds requested steps ({self.total_steps}). Halting further processing.") raise sd_samplers_common.InterruptedException

And when you say it breaks DPM Adaptive, are you referencing this code: https://github.com/CompVis/stable-diffusion/blob/main/ldm/models/diffusion/ddpm.py

Where they do modify N internally - which wouldn't effect this safeguard, since self.step and self.steps changes via code. This safeguard would protect in the instance given above as an example.

[w-e-w]

If you actually test Sampling method: DPM adaptive you will quickly see the issue
left current (expeted), right your PR

"DPM adaptive ignores the user setp input"

if your PR is going to be merged, then the things it breaks needs to be fixed in your PR

As demonstrated you with DPM adaptive, yours change may have the potential to break existing things
so I can only assume that it may other things as well, things that we are not even aware of
such as extensions maintained by others developers, assuming that they are still maintaining it

when possible breaking change should be avoided
unless there's a very good reason to introduce to do so

---

You don't want to ever introduce the ability for new code to contradict what the user specified in the gui. I would think.

if you're saying this because your other PR was closed
I believe I've stated very clearly, your PR was closed because it is "dead code" https://en.wikipedia.org/wiki/Dead_code
adding instructions on how people to use it dose not make it a valid PR

[Kittensx]

As I understand it, DPM adaptive modifies the sigmas to determine the appropriate count of steps. The issue I pointed out was the ability for a scheduler to change the number of steps and you have no safeguards against it. I suggested the code modification to protect the system, and DPM adaptive would not be effected because of the way it works. DPM adaptive does not attempt to change the steps that a uuser inputs, they instead adjust sigma / noise levels dynamically based on the noise schedule.

As Chat GPT Summarizes the issue:

DPM Adaptive (Denoising Diffusion Probabilistic Model Adaptive) works differently from traditional samplers because it dynamically adjusts the number of steps based on the target noise schedule and tolerance settings rather than strictly following the user-specified step count.

How DPM Adaptive Works:
Adaptive Step Selection – Unlike standard samplers that use a fixed number of steps, DPM Adaptive chooses step intervals dynamically based on the noise schedule. It aims to minimize errors while optimizing efficiency.
Sigma-Based Control – Instead of relying on an explicit step count (self.step), DPM Adaptive modifies sigmas (noise values) to determine the appropriate number of steps. This means the number of steps taken can vary unpredictably, and it may use fewer or more steps than the user-specified count.
Convergence-Driven Stopping – The sampler stops early if it reaches a convergence threshold, meaning the actual number of steps depends on how quickly it denoises the image.
Why the Safeguard Proposal Conflicts with DPM Adaptive
The proposed safeguard (if self.step >= self.total_steps: raise Exception) assumes that all samplers must strictly follow the user-specified step count. However, DPM Adaptive ignores the step count entirely in favor of an adaptive approach. If this safeguard is added:

DPM Adaptive would fail because it doesn’t track self.step in a predictable manner.
Other custom samplers might break if they also use a dynamic step approach.
Extensions modifying steps dynamically would be impacted.
The Key Disagreement
The author of the PR believes samplers should never contradict the GUI step count.
w-e-w (the reviewer) argues that the safeguard would limit flexibility and break samplers like DPM Adaptive, which are designed to ignore the step count.
Conclusion
DPM Adaptive does not "ignore" user steps arbitrarily; rather, it uses a different method (sigma-based control) to determine step count dynamically. Any hard-coded step limit will break its intended functionality. Instead of enforcing a step-based safeguard, a more nuanced approach (like checking if a sampler explicitly requests an adaptive step count) might be a better solution.

[Kittensx]

did you use chat gpt to send this? also that version of the request I am no longer concerned about. I made a valid suggestion and was ignored. So I've learned not to try to update on this project because my suggestions are not regarded as being helpful. So i'll take my suggestions somewhere else and just work on my own projects.

…

On Wed, Jun 25, 2025 at 11:53 AM dcq01 ***@***.***> wrote: ***@***.**** commented on this pull request. Hi! I'm a grad student working on a research project about using large language models to automate code review. Based on your commit 41ba1be <41ba1be> and the changes in modules/sd_samplers_cfg_denoiser.py, my tool generated this comment: 1. *Exception Handling*: Ensure that sd_samplers_common.InterruptedException is properly handled in the calling context to avoid unhandled exceptions. 2. *Logical Flow*: The new check for self.step >= self.total_steps is a good addition to prevent further processing when the maximum number of steps has been reached. 3. *Testing*: Add unit tests that cover scenarios where self.step is equal to or exceeds self.total_steps to verify that the exception is raised as expected. 4. *Boundary Testing*: Test the condition where self.step is exactly equal to self.total_steps to confirm that the exception is raised correctly. 5. *Indentation Error*: The line raise sd_samplers_common.InterruptedException is incorrectly indented. It should align with the if statement above it. 6. *Comment Clarity*: The comment added before the if statement could be improved by explaining the purpose of self.step and self.total_steps . 7. *Comment Style Consistency*: The comment style should be consistent throughout the code. The new comment should follow the same format as other comments in the file. 8. *Logging*: Log the occurrence of the InterruptedException for debugging and monitoring purposes to help identify potential misuse or unexpected behavior in production environments. 9. *Documentation*: Update any relevant documentation to reflect the new exception and the conditions under which it is raised. As part of my research, I'm trying to understand how useful these comments are in real-world development. If you have a moment, I'd be super grateful if you could quickly reply to these two yes/no questions: 1. Does this comment provide suggestions from a dimension you hadn’t considered? 2. 2. Do you find this comment helpful? Thanks a lot for your time and feedback! And sorry again if this message is a bother. — Reply to this email directly, view it on GitHub <#16814 (review)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BLYPP3SUGJY2HZIBYVERV2D3FLAXHAVCNFSM6AAAAABV7ID4XWVHI2DSMVQWIX3LMV43YUDVNRWFEZLROVSXG5CSMV3GSZLXHMZDSNJYG42TSMRUGQ> . You are receiving this because you authored the thread.Message ID: ***@***.*** com>