add support for gemma4 model

[n1ck-guo]

Description

Please briefly describe your main changes, the motivation.

Type of Change

• Bug fix
• New feature
• Documentation update
• Performance improvement
• Code refactoring
• Other (please specify):

Related Issues

Fixes or relates to #

Checklist Before Submitting

• My code has been tested locally.
• Documentation has been updated as needed.
• New or updated tests are included where applicable.

[Copilot]

Pull request overview

Adds runtime handling for the gemma4 model type by patching Gemma4 decoder layers to avoid shape mismatches during auto-round block-wise quantization.

Changes:

• Added gemma4 to the special model list and introduced a Gemma4-specific patch routine.
• Hooked the patch into _handle_special_model when model.config.model_type == "gemma4".
• Removed a couple of stray whitespace-only lines near ignore-layer registrations.

[n1ck-guo]

/azp run Unit-Test-CUDA-AutoRound

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[JGSphaela]

Hi @n1ck-guo , I encountered a crash while testing this PR with google/gemma-4-E2B-it using iters 0 (RTN).

The Issue:
The quantization crashes with TypeError: 'NoneType' object does not support item assignment at transformers/models/gemma4/modeling_gemma4.py:1226. This happens because shared_kv_states is passed as None to the attention module.

Root Cause:
The patched_layer_forward signature in auto_round/special_model_handler.py is missing the shared_kv_states argument, which is the 3rd positional argument in the latest Gemma4DecoderLayer.forward implementation.

For Gemma-4 E2B/E4B, this dictionary is required to manage the shared KV cache between anchor and sharer layers. When auto-round patches the layer, it accidentally drops this argument.

Suggested Fix:
Update the signature in _patch_gemma4_model to include shared_kv_states and ensure it is propagated to orig_fwd. Additionally, since auto-round processes blocks individually, we should initialize a shared dictionary in the patching closure to maintain the state across layers:

# In _patch_gemma4_model
shared_kv_states_global = {}

# In patched_layer_forward signature
def patched_layer_forward(self, hidden_states, per_layer_input=None, shared_kv_states=None, ...):
    if shared_kv_states is None:
        shared_kv_states = shared_kv_states_global
    ...
    return orig_fwd(..., shared_kv_states=shared_kv_states, ...)

This fix resolved the issue in my local tests and allowed quantization to complete.