PR #45379: fix(config): add deepstack_visual_indexes to Qwen3_5MoeVisionConfig

• Repository: huggingface/transformers
• Author: hijingsong
• State: open | merged: False
• Link: https://github.com/huggingface/transformers/pull/45379

Description (problem / solution / changelog)

Problem

The @strict decorator on Qwen3_5MoeVisionConfig silently drops the deepstack_visual_indexes field during config loading because it is not declared as a class attribute. Every Qwen3.5 MoE model on HuggingFace ships with this field in its config.json (e.g. Qwen/Qwen3.5-35B-A3B-Base).

Root Cause

Qwen3_5MoeVisionConfig inherits from Qwen3_5VisionConfig, which overrides deepstack_visual_indexes with an AttributeError() sentinel (inherited from the parent chain). The generated configuration_qwen3_5_moe.py does not include the field at all, so @strict rejects it when deserializing from config.json.

Fix

Override deepstack_visual_indexes as a properly typed class attribute on Qwen3_5MoeVisionConfig with default (), matching the type annotation used in Qwen3VLVisionConfig. This ensures:

• The field is accepted when loading model configs
• config.vision_config.deepstack_visual_indexes returns the value from config.json
• No silent data loss during deserialization

Verification

from transformers import AutoConfig
config = AutoConfig.from_pretrained("Qwen/Qwen3.5-35B-A3B-Base", trust_remote_code=True)
print(config.vision_config.deepstack_visual_indexes)  # [] — no longer raises AttributeError

Fixes #45375

Changes

• modular_qwen3_5_moe.py: Add deepstack_visual_indexes field override with docstring
• configuration_qwen3_5_moe.py: Add deepstack_visual_indexes field with docstring (generated)

Changed files

• src/transformers/models/qwen3_5_moe/configuration_qwen3_5_moe.py (modified, +3/-0)
• src/transformers/models/qwen3_5_moe/modular_qwen3_5_moe.py (modified, +6/-1)

PR #45380: fix Qwen3_5MoeVisionConfig deepstack_visual_indexes silently dropped by @strict (Issue: https://github.com/huggingface/transformers/issues/45375)

• Repository: huggingface/transformers
• Author: Brianzhengca
• State: closed | merged: False
• Link: https://github.com/huggingface/transformers/pull/45380

Description (problem / solution / changelog)

What does this PR do?

Fixes #45375 (issue)

Root Cause

In modular_qwen3_5.py, Qwen3_5VisionConfig was marked @strict but also had deepstack_visual_indexes = AttributeError(). That meant the field was effectively undeclared for strict loading, so when a real checkpoint config included "deepstack_visual_indexes": [], Transformers silently dropped it during config construction. Since Qwen3_5MoeVisionConfig inherits that shared Qwen3.5 vision config shape, the bug affected both plain Qwen3.5 and Qwen3.5-MoE configs.

Fix

I restored deepstack_visual_indexes as a real config attribute in modular_qwen3_5.py, with an empty default so Qwen3.5 keeps deepstack disabled by default.

Tests

from transformers import Qwen3_5Config, Qwen3_5MoeConfig

for cls in [Qwen3_5Config, Qwen3_5MoeConfig]:
    config = cls(vision_config={"deepstack_visual_indexes": [1, 2]})
    print(
        cls.__name__,
        type(config.vision_config.deepstack_visual_indexes).__name__,
        config.vision_config.deepstack_visual_indexes,
    )

Output:

Qwen3_5MoeConfig list [1, 2]

Changed files

• src/transformers/models/qwen3_5/configuration_qwen3_5.py (modified, +4/-0)
• src/transformers/models/qwen3_5/modular_qwen3_5.py (modified, +3/-1)
• src/transformers/models/qwen3_5_moe/configuration_qwen3_5_moe.py (modified, +4/-0)