[raimbekovm]

Fixes #43186

What does this PR do?

Fixes floating-point precision issue in JanusImageProcessor.resize() where int() truncation caused incorrect output dimensions.

Problem:

height = 2522, size = 384
delta = 384 / 2522 = 0.15225218081681999...
height * delta = 383.9999999999...
int(height * delta) = 383  # Wrong!
round(height * delta) = 384  # Correct!

Solution:
Replace int() with round() for proper rounding of scaled dimensions.

Why edit modular_janus.py?

The file image_processing_deepseek_vl.py (mentioned in the issue) is auto-generated from modular_janus.py. The header states:

🚨 This file was automatically generated from src/transformers/models/janus/modular_janus.py.
   Do NOT edit this file manually...

DeepseekVL imports JanusImageProcessor, so fixing the source fixes all dependent models:

• Janus
• DeepseekVL
• DeepseekVL Hybrid

[Rocketknight1]

Bug seems legit but I'd like someone with more image processor experience to check that we're still matching the behaviour that these models expect, because sometimes models are trained with slightly buggy preprocessing code and it becomes important to match that behaviour later. cc @molbap @yonigozlan @zucchini-nlp

[zucchini-nlp]

Makes sense to me, we need the sides to be exactly 'size' length. Can you run make fix-repo to convert modular files?

[raimbekovm]

Done. Ran the modular converter and pushed the regenerated file.

Note: The same int() truncation issue exists in two other places:

• image_processing_deepseek_vl.py (generated from modular_deepseek_vl.py which inherits from JanusImageProcessor)
• image_processing_janus_fast.py (not a modular file)

Should I include these in this PR or handle them separately?

[zucchini-nlp]

Should I include these in this PR or handle them separately?

Running make fix-repo should have auto-fixed those, no? I see the CI is still complaining about these models

[raimbekovm]

Fixed. Added the missing changes for deepseek_vl and janus_fast.

[zucchini-nlp]

@bot /repo

[github-actions]

Repo. Consistency fix runs successfully without any file modified.

[yonigozlan]

run-slow: deepseek_vl, janus, deepseek_vl_hybrid

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[zucchini-nlp]

run-slow: deepseek_vl, janus, deepseek_vl_hybrid

[zucchini-nlp]

slow tests cannot be triggered 😢

[molbap]

run-slow: deepseek_vl, deepseek_vl_hybrid, janus

[github-actions]

This comment contains run-slow, running the specified jobs:

models: ["models/deepseek_vl", "models/deepseek_vl_hybrid", "models/janus"]
quantizations: []

[github-actions]

CI Results

Workflow Run ⚙️

Model CI Report

❌ Failed tests

• deepseek_vl_hybrid:
  tests/models/deepseek_vl_hybrid/test_modeling_deepseek_vl_hybrid.py::DeepseekVLHybridIntegrationTest::test_model_text_generation

• janus:
  tests/models/janus/test_modeling_janus.py::JanusIntegrationTest::test_model_text_generation

[yonigozlan]

I got similar results locally with slow tests now failing due to this changes. However it seems to me that the processing now works as it's supposed to , so ok to change the expected values (and add a 🚨 as it breaks BC a bit)? @molbap @zucchini-nlp

[zucchini-nlp]

Yeah, sounds good to me. Let's just update test values with the new ones and merge as breaking

[yonigozlan]

run-slow: deepseek_vl, deepseek_vl_hybrid, janus

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: deepseek_vl, deepseek_vl_hybrid, janus