[Qwen3]: If qwen3 is used along with peft config, peft adds opcl obj no… (#926)

…t injested further

## Summary
Fixes : #925
Fix TypeError: liger_fused_linear_cross_entropy() got an unexpected
keyword
argument 'return_dict' that occurs when using Liger Kernel with PEFT and
   transformers Trainer.

The return_dict parameter is a standard transformers parameter that
controls
output format (ModelOutput vs tuple). When using PEFT with Liger Kernel
models,
   this parameter is passed through **kwargs all the way to
liger_fused_linear_cross_entropy() which doesn't accept it, causing
training to
   crash.

This PR adds kwargs.pop("return_dict", None) in all affected model files
to
remove the parameter before it reaches the loss calculation functions.

<!---
## Details
 Root Cause:

     - transformers Trainer passes return_dict in model inputs
     - PEFT wrapper forwards all kwargs to base model
- Liger Kernel model implementations pass **kwargs to
LigerForCausalLMLoss()
- This propagates to liger_fused_linear_cross_entropy() which doesn't
accept return_dict
--->

## Testing Done
- Verified the fix resolves the TypeError with Qwen3 + PEFT +
transformers Trainer
     - Tested training runs successfully complete without the error

<!-- 
Replace BLANK with your device type. For example, A100-80G-PCIe

Complete the following tasks before sending your PR, and replace `[ ]`
with
`[x]` to indicate you have done them. 
-->

- Hardware Type: H100 * 8 GPUs
- []  run make test to ensure correctness
- []  run make checkstyle to ensure code style
- []  run make test-convergence to ensure convergence

Author: yeshsurya

src/liger_kernel/transformers/model/qwen3.py

@@ -83,6 +83,8 @@ def lce_forward(
     kept_hidden_states = hidden_states[:, slice_indices, :]
 
     shift_labels = kwargs.pop("shift_labels", None)
+    # Remove output-control parameters that shouldn't be passed to loss functions
+    kwargs.pop("return_dict", None)
     logits = None
     loss = None
     token_accuracy = None