fix: namespace package conflict of optimum and optimum-quanto

[ParagEkbote]

Description

As described in the issue, I have removed the namespace conflict by only using the optimum-quanto package and removing optimum from deps. Could you please review?

cc: @davidberenstein1957

Related Issue

Fixes #296

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• This change requires a documentation update

How Has This Been Tested?

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

Additional Notes

[johannaSommer]

Hey @ParagEkbote thanks a lot for your effort on fixing this! The changes look good to me content wise, did you get the chance to test this locally for any tiny model?
Additionally, it would be important to keep the "import_algorithm_packages" function - it helps us isolate imports for better overall handling of all of our different packages. Otherwise almost ready to merge, thanks again 😊🫡

[ParagEkbote]

I've added back the function as requested but resolving the namespace conflict remains a bit more difficult. By using pipdeptree, I was able to find out that whisper-s2t also depends on optimum.

Warning!!! Possibly conflicting dependencies found:
* whisper-s2t==1.3.1
 - optimum [required: Any, installed: ?]
------------------------------------------------------------------------
├── optimum-quanto [required: >=0.2.5, installed: 0.2.7]
│   ├── torch [required: >=2.6.0, installed: 2.7.0]
│   │   ├── filelock [required: Any, installed: 3.19.1]
│   │   ├── typing_extensions [required: >=4.10.0, installed: 4.14.1]
│   │   ├── setuptools [required: Any, installed: 80.9.0]
│   │   ├── sympy [required: >=1.13.3, installed: 1.14.0]
--
    ├── optimum [required: Any, installed: ?]
    ├── transformers [required: Any, installed: 4.52.4]
    │   ├── filelock [required: Any, installed: 3.19.1]
    │   ├── huggingface-hub [required: >=0.30.0,<1.0, installed: 0.34.4]
    │   │   ├── filelock [required: Any, installed: 3.19.1]
    │   │   ├── fsspec [required: >=2023.5.0, installed: 2024.12.0]

But for now, by perfoming my changes, the optimum-quanto package is selected only. The test code is as follows:

import numpy as np
import torch
import torchvision
from torchvision import transforms

from pruna import SmashConfig, smash


# 1. Load a small Torch model (just for testing)
model = torchvision.models.resnet18(weights=None)  
model.eval()


# Initialize SmashConfig
smash_config = SmashConfig()
smash_config["quantizer"] = "quanto"

# Smash the model
smashed_model = smash(
    model=model,
    smash_config=smash_config,
)

image = np.random.randint(0, 256, size=(224, 224, 3), dtype=np.float32)
input_tensor = transforms.ToTensor()(image).unsqueeze(0)

# 5. Run inference
with torch.no_grad():
    output = smashed_model(input_tensor)

print("Output shape:", output.shape)

WDYT?

[johannaSommer]

I think my initial comment wasn't super clear - sorry about that. 😕 We definitely have to keep the logic of the imported modules and accessing these items through the dict (e.g. no top level optimum import.

After reviewing it, I think the change we need is only to remove it from the pyproject.toml and adjust the import as follows:

    def import_algorithm_packages(self) -> Dict[str, Any]:
        """
        Provide a algorithm packages for the algorithm.

        Returns
        -------
        Dict[str, Any]
            The algorithm packages.
        """
        import optimum.quanto as quanto
        from optimum.quanto import Calibration, freeze, quantize

        return dict(Calibration=Calibration, freeze=freeze, quantize=quantize, quanto=quanto)

And then adjust the getattr to

weights = getattr(imported_modules["quanto"], smash_config["weight_bits"])
activations = getattr(imported_modules["quanto"], smash_config["act_bits"]) if smash_config["act_bits"] is not None else None

[ParagEkbote]

Note that this is only to keep mypy happy, this will not execute at runtime.

[ParagEkbote]

I've added back the function as requested but resolving the namespace conflict remains a bit more difficult. By using pipdeptree, I was able to find out that whisper-s2t also depends on optimum.

Warning!!! Possibly conflicting dependencies found:
* whisper-s2t==1.3.1
 - optimum [required: Any, installed: ?]
------------------------------------------------------------------------
├── optimum-quanto [required: >=0.2.5, installed: 0.2.7]
│   ├── torch [required: >=2.6.0, installed: 2.7.0]
│   │   ├── filelock [required: Any, installed: 3.19.1]
│   │   ├── typing_extensions [required: >=4.10.0, installed: 4.14.1]
│   │   ├── setuptools [required: Any, installed: 80.9.0]
│   │   ├── sympy [required: >=1.13.3, installed: 1.14.0]
--
    ├── optimum [required: Any, installed: ?]
    ├── transformers [required: Any, installed: 4.52.4]
    │   ├── filelock [required: Any, installed: 3.19.1]
    │   ├── huggingface-hub [required: >=0.30.0,<1.0, installed: 0.34.4]
    │   │   ├── filelock [required: Any, installed: 3.19.1]
    │   │   ├── fsspec [required: >=2023.5.0, installed: 2024.12.0]

But for now, by perfoming my changes, the optimum-quanto package is selected only. The test code is as follows:

import numpy as np
import torch
import torchvision
from torchvision import transforms

from pruna import SmashConfig, smash


# 1. Load a small Torch model (just for testing)
model = torchvision.models.resnet18(weights=None)  
model.eval()


# Initialize SmashConfig
smash_config = SmashConfig()
smash_config["quantizer"] = "quanto"

# Smash the model
smashed_model = smash(
    model=model,
    smash_config=smash_config,
)

image = np.random.randint(0, 256, size=(224, 224, 3), dtype=np.float32)
input_tensor = transforms.ToTensor()(image).unsqueeze(0)

# 5. Run inference
with torch.no_grad():
    output = smashed_model(input_tensor)

print("Output shape:", output.shape)

WDYT?

Is the whisper-s2t dep going to be updated in future versions?

[simlang]

Thank you for looking at this! I left some comments, which are probably some leftovers from adding imported_modules again.
What do you think?

[simlang]

Looks good! LGTM! 🚀
Thanks!

[johannaSommer]

Thanks a lot for your patience and for tackling this!