[contrib] Add HunyuanVideo-1.5 NeuronX port

[lutfanm-aws]

Summary

• Adds NeuronX Distributed Inference implementation of HunyuanVideo-1.5 (8.3B text-to-video diffusion model)
• Multi-component pipeline: Qwen2.5-VL-7B encoder, 54-block MMDiT backbone (TP=8), 3D VAE decoder
• All components run on NeuronCores — no CPU fallback
• 33× speedup on transformer backbone vs CPU, ~3.4 min total for 480p video on trn2.48xlarge

Model Details

• Architecture: MMDiT (54 blocks) + DCAE VAE
• Parameters: ~16B total (8.3B backbone + 7B text encoder + 871M VAE)
• Output: 33 frames at 480×640 (~1.1s at 30fps)
• Precision: BF16
• Source: hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v

Performance (trn2.48xlarge, TP=8)

Component	CPU (est.)	Neuron	Speedup
Transformer (100 passes)	~75 min	2.3 min	33×
Text encoding	~2 min	34s	3.5×
VAE decode	99s	29s	3.4×
Total	~79 min	3.4 min	~23×

Files

• contrib/models/HunyuanVideo-1.5/src/ — Model implementations, compile & inference scripts
• contrib/models/HunyuanVideo-1.5/test/ — Integration tests
• contrib/models/HunyuanVideo-1.5/README.md — Documentation with setup, usage, and benchmarks