transformers - 💡(How to fix) Fix CLIPTextModel and CLIPTextModelWithProjection have inconsistent structure since v5.6 (.text_model removed from one but not the other)

System Info

• transformers version: reproduced on 5.6.0 and 5.9.0 (unaffected on 5.5.4)
• Platform: Windows-11-10.0.26200-SP0 (also reproducible cross-platform; the bug is in pure model construction)
• Python version: 3.12.9
• PyTorch version: 2.7.1+cu128
• Accelerate version: 1.8.1

This is a pure-Python structural issue in model __init__; it does not depend on weights, device, or platform.

Who can help?

Leaving for triage — this concerns the CLIP text model implementation in models/clip/modeling_clip.py.

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

Since v5.6.0, CLIPTextModel was flattened — its submodules (embeddings, encoder, final_layer_norm) are now assigned directly in __init__ and the previous self.text_model (CLIPTextTransformer) wrapper was removed. However, the closely related CLIPTextModelWithProjection was not changed: its __init__ still does self.text_model = CLIPTextModel._from_config(config), so it retains a .text_model submodule.

The result is that two sibling classes now have inconsistent structure, and CLIPTextModel lost its long-standing .text_model attribute with no deprecation path.

from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTextConfig

cfg = CLIPTextConfig()

print("CLIPTextModel has .text_model:              ", hasattr(CLIPTextModel(cfg), "text_model"))
print("CLIPTextModelWithProjection has .text_model:", hasattr(CLIPTextModelWithProjection(cfg), "text_model"))

Output:

Confirmed in source:

• v5.5.x — CLIPTextModel.__init__: self.text_model = CLIPTextTransformer(config)
• v5.6.0+ — CLIPTextModel.__init__: self.embeddings = ...; self.encoder = ...; self.final_layer_norm = ... (no text_model)
• v5.6.0+ — CLIPTextModelWithProjection.__init__: still self.text_model = CLIPTextModel._from_config(config)

Any code that reads model.text_model.<...> on a CLIPTextModel (a very common, long-stable pattern) now raises:

AttributeError: 'CLIPTextModel' object has no attribute 'text_model'

while the identical access on CLIPTextModelWithProjection keeps working — so the regression appears for only some model variants and is easy to misattribute to downstream code.

Expected behavior

One of the following, rather than an undocumented, asymmetric removal:

1. Consistency — flatten both classes the same way (or neither), so siblings share structure; or
2. Backward-compatible accessor — keep a .text_model property on CLIPTextModel that returns the flattened module (so existing model.text_model.embeddings... access continues to resolve); or
3. Documentation — if the flattening is intentional and permanent, call it out as a breaking change in the v5 migration guide, including the .text_model removal and the resulting asymmetry with CLIPTextModelWithProjection.

The core concern is that CLIPTextModel and CLIPTextModelWithProjection are now structurally inconsistent, and a widely-used attribute was removed silently.