- Removed the direct call to `_load_target_weights` in the constructor of `Qwen3DFlashForCausalLM`.
- Added a class method `from_pretrained` to handle loading weights and configurations more effectively.
- Updated weight handling to ensure compatibility with the target data type.
- Modified the `extract_dflash_debug_bundle.py` script to use `dtype` instead of `torch_dtype` and added `attn_implementation` parameter for draft model loading.

These changes improve the model's initialization process and enhance the flexibility of loading configurations.

…dels

- Introduced functions to check and annotate hidden states in models during export.
- Enhanced configuration to include hidden state outputs for models with multiple hidden layers.
- Implemented a test suite to validate hidden state annotations in exported OpenVINO models.

These changes improve the model export process by allowing the inclusion of hidden states, which is essential for certain text generation tasks.

- Implemented helper functions to find and add model outputs based on tensor names.
- Added a new test case to validate that annotated hidden state outputs match those from PyTorch for the GPT-2 model.
- Enhanced the export process to include hidden state outputs, ensuring compatibility with text generation tasks.

These changes improve the testing framework for OpenVINO model exports, specifically focusing on hidden state annotations.

- Added support for overriding the DFlash block size via the environment variable `DFLASH_BLOCK_SIZE_OVERRIDE`.
- Included error handling to ensure the block size is an integer greater than 1.
- This enhancement allows for more flexible configuration of DFlash model exports, improving usability and performance.

These changes contribute to the ongoing improvements in the OpenVINO export process for DFlash models.

- Added support for committed prefix cache policy in DFlash models by updating runtime information.
- Modified `DFlashDummyInputGenerator` to use "hidden_states" instead of "target_hidden" for input names.
- Updated Qwen3DFlash model to handle hidden states and past key values more effectively during inference.
- Introduced a new script to compare DFlash cache semantics between original and patched models.
- Enhanced tests to validate the integration of hidden states and ensure consistency in outputs.

These changes improve the functionality and testing of DFlash models within the OpenVINO framework, ensuring better performance and reliability.