fix: support RTX 5070 Ti (Blackwell/SM_100) + fix web UI voice upload bug

[newlix]

Summary

• RTX 5070 Ti (Blackwell) support: upgrade CUDA 12.4 → 12.8, PyTorch 2.3.1 → 2.7.0, TensorRT ≥10.8.0, onnxruntime-gpu 1.18 → 1.21
• Fix ruamel.yaml incompatibility: pin ruamel.yaml==0.17.40 to fix AttributeError: 'Loader' object has no attribute 'max_depth' with HyperPyYAML 1.2.2
• Fix web UI voice upload bug: frontend was sending prompt_audio but backend expected prompt_wav, causing all direct audio uploads to silently fail

Changes

Dockerfile & docker/Dockerfile

• Base image: nvidia/cuda:12.4.1-cudnn-devel-ubuntu22.04 → nvidia/cuda:12.8.1-cudnn-devel-ubuntu22.04

requirements.txt

• PyTorch index: cu121 → cu128
• torch 2.3.1 → 2.7.0
• torchaudio 2.3.1 → 2.7.0
• onnxruntime-gpu 1.18.0 → 1.21.0
• tensorrt-cu12 pinned ==10.13.3.9 → >=10.8.0 (supports Blackwell)
• Add ruamel.yaml==0.17.40 (fixes startup crash with HyperPyYAML)
• Remove legacy onnxruntime-cuda-12 private index (not needed for cu128)

app.py

• Fix FormData field name: prompt_audio → prompt_wav in web UI generate() function

Test plan

• Build Docker image on CUDA 12.8 host with RTX 5070 Ti
• Verify service starts without ruamel.yaml error
• Verify web UI audio file upload works end-to-end
• Verify voice cloning output quality unchanged

🤖 Generated with Claude Code