Add YuE Music Generation contrib model (full-song from lyrics)

[jimburtoft]

Summary

• Adds YuE (M-A-P/HKUST) full-song music generation from lyrics to contrib
• Two-stage NxDI pipeline: S1 (7B LLaMA, TP=2) generates coarse codec tokens with classifier-free guidance, S2 (1B LLaMA, TP=1) refines via teacher-forcing
• Includes NKI MLP kernel optimization (20% speedup) and S2 batching

Model Details

• Source: YuE (M-A-P/HKUST, Apache 2.0)
• Architecture: S1 = 7B LLaMA (vocab ~84K), S2 = 1B LLaMA (vocab ~84K), xcodec_mini audio decoder
• Instance: trn2.3xlarge

Performance

Config	Pipeline Total	RTF	vs GPU (L4)
Baseline	514s	17.6x	2.8x faster
+ KV-cache + NKI + bs=2	~440s	~15x	3.3x faster

S1: 34.5 tok/s (3.5x faster than L4 GPU). S2: 4.3x faster than L4 GPU with KV-cache.

Key Technical Contributions

• Subprocess architecture: S1 and S2 run in separate processes (NxDI models with different TP cannot coexist)
• Custom CFG generation loop: Maintains synchronized KV caches for conditional/unconditional rows
• NKI MLP TKG kernel: Fuses RMSNorm + Gate/Up + SiLU + Down, eliminates 5-6 HBM round-trips per layer
• S2 KV-cache optimization: 3.3x speedup over naive teacher-forcing loop

Testing

• 10 files: README.md, genre.txt, lyrics.txt, setup.sh, src/{init,nki_mlp_patch,yue_e2e_neuron,yue_stage1_worker,yue_stage2_worker}.py, test/{init,integration/{init,test_model},unit/init}.py
• Integration tests: lyrics parsing, file format validation, E2E pipeline (with and without NKI kernels)
• Validated on trn2.3xlarge with SDK 2.28