Hugging Face's logo

bharatgenai
/
patram-7b-instruct

Image-Text-to-Text
Transformers
Safetensors
English
patram
text-generation
visual-document-understanding
visual-question-answering
indian-documents
conversational
custom_code
Model card Files
xet
 Community
7

Unable to determine the version compatibilities

#6
by AlfredManan - opened
Discussion
AlfredManan

Hi BharatGen Team,

I am trying to run inference using patram-7b-instruct locally, but I am running into multiple errors due to library version incompatibilities. The model card does not specify exact pinned versions for the required libraries, which makes it very difficult to set up a working environment.

Errors encountered so far:

  1. ImportError: cannot import name 'TextKwargs' from 'transformers.processing_utils'
    → Caused by older versions of transformers that do not have TextKwargs

  2. AttributeError: 'NoneType' object has no attribute 'size' at modeling_patram.py:2320
    → past_key_values is None on the first forward pass, caused by newer transformers versions (≥ 4.47) that changed how KV cache is initialized

  3. AssertionError at modeling_patram.py:2669 — assert generation_config.use_cache
    → The model hardcodes use_cache=True, but newer transformers versions conflict with this

My setup:

  • GPU: 22 GB VRAM
  • Python: 3.10
  • OS: Linux

Request:
Could you please share a complete requirements.txt with exact pinned versions for all dependencies that are confirmed to work with this model? Specifically:

  • transformers==?
  • torch==?
  • accelerate==?
  • pydantic==?
  • Any other pinned versions

This will help the community avoid spending hours debugging environment issues and focus on actually using the model.

Thank you!
hrithiksagar-bgen
 BharatGen AI org
•
edited 13 days ago

Hi, thanks for sharing the detailed error trace and environment details.

For patram-7b-instruct, the dependency versions we currently recommend for local inference are:

torch==2.7.0
torchvision==0.22.0
transformers==4.50.3
accelerate==0.26.0
einops==0.8.1

The errors you ran into are generally due to version mismatches across the transformers stack, especially around cache handling and generation behavior across releases.

Using a clean environment with the pinned versions above should resolve these issues.

We’ve also added a requirements.txt file to the repository files section so the recommended setup is easier to reproduce.

Thanks for reporting this.

Sign up or log in to comment