Incompatible transformers version requirements: mlx-vlm vs docling-ibm-models on macOS with docling v2.31

[ddoron9]

When using docling v2.31 on macOS and attempting to install mlx-vlm for Apple Silicon support, there is a dependency conflict with docling-ibm-models due to incompatible transformers version requirements.

mlx-vlm requires transformers >=4.51.3
docling-ibm-models requires transformers >=4.42.0,<4.43.0

This makes it impossible to install both mlx-vlm and docling-ibm-models in the same environment, as their required transformers versions do not overlap. This is a significant issue for users who want to leverage both IBM models and Apple Silicon acceleration in the same project.

Even if I attempt to downgrade mlx-vlm to match the transformers version required by docling-ibm-models, it is not feasible because other dependencies are too old and cannot be downgraded to such versions.

Please consider updating the transformers version constraint in docling-ibm-models to allow compatibility with mlx-vlm and recent versions of transformers, or provide guidance on how to use both packages together in a single environment.

Thank you!

[cau-git]

Is this seen on an x86-based macOS? Because the transformers requirement <4.43.0 is only for x86 architecture.

[ddoron9]

When using docling v2.31 on macOS and attempting to install mlx-vlm for Apple Silicon support, there is a dependency conflict with docling-ibm-models due to incompatible transformers version requirements.

• mlx-vlm requires transformers >=4.51.3
• docling-ibm-models requires transformers >=4.42.0,<4.43.0
  This makes it impossible to install both mlx-vlm and docling-ibm-models in the same environment, as their required transformers versions do not overlap. This is a significant issue for users who want to leverage both IBM models and Apple Silicon acceleration in the same project.
  Even attempting to downgrade mlx-vlm is not feasible because older versions depend on outdated libraries that are incompatible with the rest of the ecosystem.
  👉 Important clarification:
This issue only occurs when using poetry to install dependencies from the official docling Git repository (via poetry install, then poetry add mlx-vlm).
When using pyenv with pip, the dependencies resolve without conflict, and both docling-ibm-models and mlx-vlm can coexist.