[contrib] Add CodeGen-350M-mono NeuronX port

[dhwanw]

Description

NeuronX Distributed Inference port of Salesforce/codegen-350M-mono, a 350M-parameter decoder-only transformer for code generation. CodeGen uses GPT-J-style architecture with partial RoPE (32/64 dims with interleaved rotation), parallel residual connections (attn + mlp + residual), and fused QKV with (Q, V, K) interleaved order requiring special weight decomposition.

Model Information

Model Name: CodeGen-350M-mono
Model Architecture: Decoder-only transformer (GPT-J variant) with partial RoPE, parallel residual connections, fused QKV, GELU-new activation, LayerNorm
Purpose: Code generation (monolingual Python)

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/codegen-350M-mono/
  README.md
  /src
    modeling_codegen.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	100% match on 14/30 prompts (54.48% avg across all prompts)
Teacher-Forced Match	✅ PASS	97.03% average
Throughput	✅ PASS	187.4 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

• Partial RoPE: Only 32 of 64 head dimensions use rotary embeddings, following GPT-J's interleaved rotation convention (not the standard LLaMA half-split).
• Fused QKV: HuggingFace checkpoint uses fused qkv_proj with mp_num=4 and (Q, V, K) interleaved order -- the weight converter handles this decomposition.
• Parallel residual: Attention and MLP both operate on the layer-normed input, and their outputs are summed with the residual.

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