feat: Implement CUDA BF16 error handling with automatic fallback to CPU for model inference and generation.

fix: Configure CUDA precision to disable BF16 on older GPUs.

feat: Automatically set generation config's `pad_token_id` and `eos_token_id` from the tokenizer and suppress Hugging Face logging warnings.

fix: Ensure proper `pad_token_id` configuration and `attention_mask` generation for DeepSeek OCR model.

feat: Improve Hugging Face cache management and enable mixed-precision inference for GPU models.

fix: Add a fallback definition for `is_torch_fx_available` in `transformers.utils.import_utils`.

feat: Add LlamaFlashAttention2 compatibility alias and eager attention implementation for model loading.

refactor: Reorder imports and the `spaces` dummy decorator definition to the top of the file.

docs: update README.md

refactor: Simplify library imports and remove verbose version checks in `app_hf.py`.

refactor: Enhance dependency import robustness with try-except blocks and add version logging for Gradio and HuggingFace Hub.

chore: Update project dependencies.

fix: Address MPS compatibility issues, ensure explicit model dtype, and improve Gradio file input handling.

Fix Gradio CSS argument placement for ZeroGPU

Implement warning suppression, ensure pad token ID for generation, enable deterministic sampling, refine Gradio UI CSS and clear functionality, and add `.env` to .gitignore."

Clean up README_HF.md

Update README with HF metadata and add remote

Initial commit: DeepSeek-OCR-2 & MedGemma-1.5 multimodal analysis app with ZeroGPU support