RTX 5070 (Blackwell) not supported — CUDA kernel error on Docker backend

Hi, the Docker backend fails to start on RTX 5000 series GPUs (Blackwell architecture).

GPU: NVIDIA GeForce RTX 5070 Laptop GPU Compute Capability: 12.0 (Blackwell) Driver: 592.15 OS: Windows 11

Error from duix-avatar-tts container:

RuntimeError: CUDA error: no kernel image is available for execution on the device

Full traceback points to fish_speech/models/text2semantic/llama.py during setup_caches → KVCache initialization.

Root cause: The Docker images are compiled with a PyTorch version that doesn't include CUDA kernels for compute capability 12.0. PyTorch 2.7+ added Blackwell support.

Request: Please rebuild and push updated Docker images (guiji2025/fish-speech-ziming, guiji2025/fun-asr, guiji2025/duix.avatar) compiled against PyTorch 2.7+ with sm_120 support.

RTX 5070/5080/5090 are now widely in use — happy to help test updated images.