[contrib] Add Ouro-1.4B NeuronX port

[dhwanw]

Description

NeuronX Distributed Inference port of ContextualAI/Ouro-1.4B, a 1.4B-parameter Universal Transformer. Ouro uses weight sharing across 4 UT steps over 24 layers, resulting in 96 unrolled physical layers for NXDI. The model features dual RMSNorm (pre+post sandwich) for both attention and MLP blocks, and intermediate norms at UT step boundaries.

Model Information

Model Name: Ouro-1.4B
Model Architecture: Decoder-only Universal Transformer -- 24 base layers x 4 UT steps = 96 unrolled physical layers, MHA (16 heads), RoPE, dual RMSNorm sandwich, SwiGLU MLP
Purpose: Text generation with Universal Transformer weight sharing

Checklist

Required Components

• Accuracy Test (test/integration/test_model.py)
  • Validates model generation and coherence
  • Performance benchmarks (TTFT, throughput)
  • Test can compile and run the model on Neuron
• README.md with the following sections:
  • Usage Example: Clear code example showing how to use the model
  • Compatibility Matrix: Table showing tested Neuron SDK versions and instance types
  • Example Checkpoints: Links to compatible model checkpoints
  • Testing Instructions: Command to run the test suite for the model
• Source Code (src/)
  • Modeling code following NxD Inference patterns

Optional Components

• Unit Tests (CPU or Neuron-based)

Folder Structure

/contrib/models/Ouro-1.4B/
  README.md
  /src
    modeling_ouro.py
  /test
    /integration
      test_model.py

Testing

Model was compiled and tested with TP=1, batch_size=1, seq_len=128, bfloat16 on trn1.32xlarge.

Test Results:

Test	Status	Result
Smoke Test	✅ PASS	Model loads successfully
Greedy Token Matching	✅ PASS	87.0% average (7/10 prompts at 100%)
Teacher-Forced Match	✅ PASS	98.0% average
Throughput	✅ PASS	23.3 tok/s

Compatibility

Tested with:

• Neuron SDK Version(s): 2.22
• Instance Type(s): trn1.32xlarge
• PyTorch Version: 2.9
• Python Version: 3.10
• Configuration: TP=1, batch_size=1, seq_len=128, bfloat16

Additional Information

• Universal Transformer loop: 4 UT steps over 24 layers with shared weights. Unrolled into 96 physical layers with duplicated weights so NXDI can iterate in a single pass.
• Dual layer norms: Each decoder layer applies pre-norm + post-norm sandwich for both attention and MLP blocks (4 RMSNorms per layer).
• Intermediate norm: Additional RMSNorm applied at UT step boundaries (every 24 layers).
• Separate KV cache per UT step: 96 total cache slots, one per unrolled layer.
• Weight conversion: HF's 24-layer weights are duplicated 4x. Intermediate norm weights come from the model's final norm.

Related Issues

N/A

vLLM Integration

• This model/feature is intended for use with vLLM
• Documentation includes vLLM registration instructions

---

By submitting this PR, I confirm that:

• I have read and followed the contributing guidelines
• This is a community contribution and may have limited testing compared to officially-supported models
• The code follows best practices and is well-documented
• All required components listed above are included