RTX50系显卡上如何部署？

CPU：Intel Xeon w7-3565X (64) @ 4.80 GHz
GPU：NVIDIA GeForce RTX 5090 32G
内存：DDR5-5600 ECC Registered 64G × 16 = 1024G
系统：Ubuntu 22.04.4 LTS x86_64
KTranformers 版本：v0.3.2
CUDA 版本：12.9
运行 DeepSeek-V2-Lite-Chat 模型：
```bash
python -m ktransformers.local_chat --model_path deepseek-ai/DeepSeek-V2-Lite-Chat --gguf_path ./DeepSeek-V2-Lite-Chat-GGUF
```

Hugging Face transformers 的不同版本似乎都无法支持 Kt 运行：
* transformers = 4.43.2
```bash
ImportError: cannot import name 'FlashAttentionKwargs' from 'transformers.modeling_flash_attention_utils' (/home/xiejianan/miniconda3/envs/ktransformers/lib/python3.11/site-packages/transformers/modeling_flash_attention_utils.py)
```
* transformers = 4.56.1
```bash
ValueError: You should provide exactly one of layers or layer_class_to_replicate to initialize a Cache.
```

求问在 RTX50 系显卡上（CUDA版本大于12.8）如何部署 Ktransformers？

Start with the `python -m ktransformers.local_chat` entry point and reproduce the reported DeepSeek-V2-Lite-Chat failures using KTransformers v0.3.2, CUDA 12.9, and the two listed transformers versions. Compare the resulting requirements and errors to determine whether RTX 5090 deployment is supported; done means a verified deployment path or clearly documented incompatibility.

Written by the indexing model from the issue text.