Fix MPS out of memory error by disabling PyTorch memory watermark

[Copilot]

PyTorch MPS backend hits a hard memory limit during generation on macOS, failing with "MPS allocated: 27.06 GB, max allowed: 27.20 GB". The error message suggests the fix.

Changes

• Set PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.0 in launcher.py at startup
• Set PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.0 in music_service.py before generation

This disables PyTorch's MPS memory watermark, allowing full use of unified memory on Apple Silicon/Metal-enabled Macs.

# launcher.py
os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"
os.environ["PYTORCH_MPS_HIGH_WATERMARK_RATIO"] = "0.0"

# music_service.py - in _run_pipeline()
os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"
os.environ["PYTORCH_MPS_HIGH_WATERMARK_RATIO"] = "0.0"

MPS-only change. No impact on CUDA/CPU paths.

Original prompt

---

This section details on the original issue you should resolve

<issue_title>Generation Failed</issue_title>
<issue_description>The app starts and loads models but generation fails with:
Generation Failed: MPS backend out of memory (MPS allocated: 27.06 GB, other allocations: 512.00 KB, max allowed: 27.20 GB). Tried to allocate 192.00 MB on private pool. Use PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.0 to disable upper limit for memory allocations (may cause system failure).</issue_description>

Comments on the Issue (you are @copilot in this section)

• Fixes Generation Failed #9

---

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.