Add Aria text decoder (MoE) contrib model port

[dhwanw]

Description

Adds NeuronX port of the Aria text decoder (MoE) to the contrib models collection.

Model Information

Field	Value
Model	rhymes-ai/Aria
Architecture	MoE Decoder-only transformer (LLaMA-based)
Parameters	~3.9B total (text decoder), 64 routed experts, top-6 routing
TP Degree	8
Precision	BF16

Checklist

• Model compiles successfully on Neuron
• Token matching validated (72.81% greedy, 98.12% teacher-forced)
• Performance profiled (4.2 tok/s)
• README with architecture details, usage, validation results
• Integration tests included

Folder Structure

contrib/models/Aria/
├── README.md
├── src/
│   ├── __init__.py
│   └── modeling_aria_text.py
└── test/
    └── integration/
        └── test_model.py

Testing

• Token Match (greedy): 72.81% (10 prompts, 32 tokens each)
• Token Match (teacher-forced): 98.12%
• Throughput: 4.2 tok/s (TP=8, BS=1, seq_len=512)

Note: This port implements only the text decoder. Greedy divergence is expected for MoE models in bf16 due to cascading expert routing differences.

Compatibility

• Neuron SDK: 2.22+
• Instance: trn1.32xlarge

🤖 Generated with Claude Code