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

contrib/models/GLM-4-9B-0414/README.md

@@ -0,0 +1,94 @@
+# Contrib Model: GLM-4-9B-0414
+
+NeuronX Distributed Inference implementation of GLM-4-9B-0414 (model_type="glm4").
+
+## Model Information
+
+- **HuggingFace ID:** `zai-org/GLM-4-9B-0414`
+- **Architecture:** Glm4ForCausalLM (decoder-only, 4 RMSNorm layers per block)
+- **Parameters:** 9B
+- **License:** Check HuggingFace model card
+
+## Key Differences from glm-4-9b-chat-hf
+
+| Feature | GLM-4-9B-0414 | glm-4-9b-chat-hf |
+|---------|---------------|-------------------|
+| model_type | glm4 | glm |
+| Layer norms per block | 4 | 2 |
+| rms_norm_eps | 1e-05 | 1.5625e-07 |
+| max_position_embeddings | 32768 | 131072 |
+
+## Validation Results
+
+**Validated:** 2026-03-18
+**Configuration:** batch_size=1, seq_len=128, tp_degree=2, bf16
+
+| Metric | Value |
+|--------|-------|
+| Teacher-forced accuracy | **98.12%** (PASSED) |
+| Greedy accuracy | 69.38% |
+| Throughput (tp=2, bs=1) | 2.4 tokens/sec |
+| Instance | trn1.32xlarge |
+
+### Token Match Details (10 prompts, 32 tokens each)
+
+| Prompt | Greedy | Teacher-Forced |
+|--------|--------|----------------|
+| "The theory of general relativity..." | 100.0% | 100.0% |
+| "The French Revolution began in..." | 100.0% | 100.0% |
+| "To solve a quadratic equation..." | 100.0% | 100.0% |
+| "Once upon a time in a distant galaxy..." | 96.9% | 96.9% |
+| "def fibonacci(n):..." | 18.8% | 96.9% |
+| "The Amazon River flows through..." | 46.9% | 96.9% |
+| "The concept of free will..." | 31.2% | 96.9% |
+| "To make a cup of coffee, first..." | 100.0% | 100.0% |
+| "List three benefits of regular exercise..." | 9.4% | 96.9% |
+| "If all roses are flowers..." | 90.6% | 96.9% |
+
+## Important: Use Glm4NeuronConfig
+
+The compile script must use `Glm4NeuronConfig` (not base `NeuronConfig`) to set
+`fused_qkv=True` and `attn_cls="NeuronGlm4Attention"`. Using base `NeuronConfig`
+will cause a `KeyError` during weight sharding because the preshard hook expects
+separate QKV keys while `convert_hf_to_neuron_state_dict` produces fused `Wqkv`.
+
+## Compilation
+
+- Compiler: neuronx-cc with flags `--enable-saturate-infinity --enable-mixed-precision-accumulation --auto-cast=none --model-type transformer -O1`
+- NEFFs: context_encoding + token_generation
+
+## Usage
+
+```python
+from modeling_glm4 import NeuronGlm4ForCausalLM, Glm4InferenceConfig, Glm4NeuronConfig
+
+neuron_config = Glm4NeuronConfig(
+    tp_degree=2,
+    batch_size=1,
+    seq_len=128,
+    torch_dtype=torch.bfloat16,
+    save_sharded_checkpoint=True,
+)
+config = Glm4InferenceConfig.from_pretrained(model_path, neuron_config=neuron_config)
+model = NeuronGlm4ForCausalLM(model_path, config)
+model.compile(compiled_model_path)
+model.load(compiled_model_path)
+```
+
+## Performance
+
+Profiled on trn1.32xlarge (single NeuronCore utilization):
+
+| Metric | Context Encoding | Token Generation |
+|--------|-----------------|------------------|
+| Throughput | - | 12.3 tok/s |
+| MBU (Memory) | 4.4% | 9.1% |
+| MFU (Compute) | 2.3% | 0.0% |
+
+*Batch size 1, sequence length 128, BF16 precision, TP=2*
+
+## Maintainer
+
+Neuroboros Team - Annapurna Labs
+
+**Last Updated:** 2026-03-18

contrib/models/GLM-4-9B-0414/src/__init__.py

@@ -0,0 +1 @@
+from .modeling_glm4 import NeuronGlm4ForCausalLM, Glm4InferenceConfig