Add GLM-4-9B-0414 contrib model port

[dhwanw]

Description

Adds NeuronX port of GLM-4-9B-0414 (model_type=glm4) to the contrib models collection.

Model Information

Field	Value
Model	zai-org/GLM-4-9B-0414
Architecture	Glm4ForCausalLM (decoder-only, 4 RMSNorm layers per block)
Parameters	9B
TP Degree	2
Precision	BF16

Checklist

• Model compiles successfully on Neuron
• Token matching validated (69.38% greedy, 98.12% teacher-forced)
• Performance profiled (12.3 tok/s)
• README with architecture details, usage, validation results
• Integration tests included

Folder Structure

contrib/models/GLM-4-9B-0414/
├── README.md
├── src/
│   ├── __init__.py
│   └── modeling_glm4.py
└── test/
    └── integration/
        └── test_model.py

Testing

• Token Match (greedy): 69.38% (10 prompts, 32 tokens each)
• Token Match (teacher-forced): 98.12%
• Throughput: 12.3 tok/s (TP=2, BS=1, seq_len=128)

Note: Uses custom Glm4NeuronConfig (fused_qkv=True, attn_cls=NeuronGlm4Attention). This is a different model_type ("glm4") from glm-4-9b-chat-hf ("glm") with 4 RMSNorm layers per block instead of 2.

Compatibility

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

🤖 Generated with Claude Code