PR #44922: fix: pop output_* flags from kwargs in capture_outputs to prevent submodule leakage

• Repository: huggingface/transformers
• Author: s-zx
• State: closed | merged: False
• Link: https://github.com/huggingface/transformers/pull/44922

Description (problem / solution / changelog)

What does this PR do?

Fixes #44849.

When output_hidden_states=True (or output_attentions=True) is passed to model.generate(), the @capture_outputs decorator reads the flag value but leaves it in **kwargs. These flags then propagate through **kwargs chains deep into sub-models — specifically, into vision encoder blocks and attention functions that don't expect them.

For the Qwen3.5 (and Qwen VL family) this causes garbled generation when output_hidden_states=True is set: the flag reaches Qwen3_5VisionBlock.attn via Qwen3_5Model.get_image_features(**kwargs) → self.visual(**kwargs) → blk(**kwargs) → self.attn(**kwargs), corrupting intermediate attention tensors and causing the model to generate repetitive image-pad tokens instead of meaningful text.

Root cause

In capture_outputs (in output_capturing.py), the decorator uses kwargs.get(...) to read the output flags — but it does not remove them from kwargs. The underlying func(self, *args, **kwargs) call therefore still sees output_hidden_states=True, which then leaks into every submodule called with **kwargs.

Fix

After reading the values for all capturable flags, immediately pop them from kwargs:

for k in capturable_flags:
    kwargs.pop(f"output_{k}", None)
if "cross_attentions" in capturable_flags or "mask_decoder_attentions" in capturable_flags:
    kwargs.pop("output_attentions", None)

Since @capture_outputs already captures the requested outputs through forward hooks, the underlying forward function (and all modules it calls) does not need to receive these flags. This pop has no effect on output correctness but prevents any downstream damage.

The fix applies to all models using @capture_outputs, not just Qwen3.5.

Changed files

• src/transformers/utils/output_capturing.py (modified, +10/-0)