Max length check for Chipper

[ajjimeno]

Chipper uses a decoder with a max_length variable used during the generation step to limit the maximum number of tokens generated by Chipper. It might raise an error or produce non-sensical output after going over that limit.

Unfortunately, the max_length variable was not being used even if it was set in the configuration of each chipper model. This PR solved that problem.

You can check running the code below if it works. Without the proposed change, the document will be generated as usual. With the proposed change, up to 200 characters will be generated. The limit of 200 tokens is achieved by modifying the chipper model max_length parameter from 1536 to 200 to ensure that there is a limit easy to test.

from unstructured_inference.inference.layout import DocumentLayout
from unstructured_inference.models.base import get_model
from unstructured_inference.models.chipper import MODEL_TYPES


MODEL_TYPES['chipperv3']["max_length"] = 200

model = get_model("chipper")

doc = DocumentLayout.from_file(
    "sample-docs/layout-parser-paper.pdf",
    element_extraction_model=model,
    pdf_image_dpi=300,
)

print(doc)

[awalker4]

This could really help in the api!