fix: clamp VQ indices to prevent CUDA out-of-bounds errors

[xander1421]

Summary

• Clamp codebook indices to valid range [0, codebook_size-1] before passing to ResidualVQ.get_output_from_indices()
• Prevents CUDA device-side assertion failures with newer versions of vector-quantize-pytorch

Problem

When using vector-quantize-pytorch >= 1.20 with PyTorch 2.10+ and Python 3.14, the codebook indices from HeartMuLa can occasionally exceed the valid codebook range, causing:

torch.AcceleratorError: CUDA error: device-side assert triggered

The error occurs in einx.get_at() when indexing into the codebook with out-of-bounds values.

Solution

Added bounds checking to clamp indices before the VQ lookup:

indices_input = codes_bestrq_emb.transpose(1, 2)
codebook_size = getattr(self.vq_embed, 'codebook_size', 8192)
indices_input = indices_input.clamp(0, codebook_size - 1)

Test Environment

• Python 3.14
• PyTorch 2.10.0+cu128
• vector-quantize-pytorch 1.27.15
• CUDA 12.8
• RTX 4000 series GPU

Test plan

• Verified fix resolves CUDA errors
• Generated audio output quality unchanged
• Tested with 30-second music generation

🤖 Generated with Claude Code