#49325 fix(modeling_utils): propagate attention implementation to nested sub-configs (#49137)

1. Adjusted attention implementation is not propagated

The resolver maps sub-configs to implementations:

target_implementations = self._get_subconfig_target_implementations(attn_implementation)
...
sub_implementation = target_implementations.get(id(submodule.config), requested_implementation)

But attn_implementation is the caller's raw value. The adjusted value is produced earlier:

requested_implementation = self._check_and_adjust_attn_implementation(...)

Base handed that adjusted value straight to submodules:

sub_implementation = requested_implementation

Because the mapping wins for every config reachable through sub_configs, submodules receive the unadjusted string. That loses the sdpa→eager downgrade, the paged| strip and the _compatible_flash_implementations correction. For flash_attention_2 with the package absent, the root becomes FLASH_ATTN_KERNEL_FALLBACK[...], but a submodule re-resolving raw flash_attention_2 takes the get_correct_attn_implementation path and raises ImportError from _flash_attn_can_dispatch. The shipped test tests/utils/test_modeling_utils.py:3333-3337 pins that after set_attn_implementation("flash_attention_2") the value the model ends on is the KERNEL FALLBACK, not the raw requested name; that is now true of the root only. Pass requested_implementation into _get_subconfig_target_implementations so the mapping values are the adjusted implementation.

2. Nested "" keys are ignored

The resolver only reads a dict's own "" key at the root:

if is_root:
    target = spec.get("", current_attn)
else:
    target = inherited_impl

For {"vision_config": {"": "eager"}}, the nested dict's "" key is never read, so that branch stays at inherited_impl or its current value. The construction-time setter this PR uses as its reference does value.get("", current_attn) at every level and re-enters itself via subconfig._attn_implementation = sub_implementation (configuration_utils.py:462-478), so the same input would set the nested config eager at construction. Runtime and construction therefore disagree. Read spec.get("", ...) at every level, matching the construction-time setter.

3. Flat keys now bind at any depth

The resolver passes the same root dict down every level:

elif child_dict is not None:
    _resolve(subconfig, child_dict, child_inherited, is_root=False)

Base matched a flat key only against direct children:

for subconfig_key in self.config.sub_configs:
    ...
    getattr(self.config, subconfig_key) is submodule.config

Now a flat key naming a deeper sub-config is honored wherever it appears. In the shipped qwen3_omni_moe tree this is not hypothetical: Qwen3OmniMoeThinkerConfig.sub_configs declares text_config, and Qwen3OmniMoeTalkerConfig.sub_configs declares text_config again, one level deeper than the root. A single flat text_config key now applies to both branches; under the old one-level match it could only name a direct child of the root. Decide the intended depth rule before shipping; if flat keys should stay direct-child-only, restore the base matching, otherwise document and test the new behavior.

4. set_experts_implementation remains one level deep

The sibling setter still uses the old one-level match:

for subconfig_key in self.config.sub_configs:
    ...
    getattr(self.config, subconfig_key) is submodule.config

The config-side _experts_implementation setter is already recursive:

value.get(subconfig_key, current_subconfig_moe)
...
subconfig._experts_implementation = sub_implementation

So construction-time dispatch reaches nested experts configs while the runtime setter does not. get_experts_implementation is documented as the counterpart of set_experts_implementation, to snapshot and restore (modeling_utils.py:1761-1780), and that restore is what needs nested awareness. Make set_experts_implementation nested-aware like attention.

5. Minor: docs, annotation, and test coverage

• modeling_utils.py:1964: the annotation says dict[int, str], but current_attn = getattr(cfg, "_attn_implementation", None) is the fallback target, and mapping[id(cfg)] = target stores it unchecked. A nested sub-config built with no implementation yields None. get_correct_attn_implementation accepts str | None and maps None to "sdpa", so the None is absorbed downstream rather than impossible. Make it dict[int, str | None].
• modeling_utils.py:2008: the docstring still says dict keys are "the sub_configs name". The resolver now accepts nested dicts and binds a flat key at any depth (1973-2003), so the documented contract is wrong. Update it to describe the nested form and the depth rule.
• tests/models/deepseek_ocr2/test_modeling_deepseek_ocr2.py:154-180: the four parts cover a branch-named string, a nested dict, a flat key and flag cleanup. The nested "" form and a name reused at two depths (qwen3_omni_moe's text_config) are absent, so the two cases the resolver disagrees on are untested. Add them.