Add OLMo-7B-Instruct NeuronX port

contrib/models/olmo-7b-instruct/README.md

@@ -0,0 +1,60 @@
+# OLMo-7B-Instruct NeuronX Port
+
+NeuronX Distributed Inference port of [allenai/OLMo-7B-Instruct](https://huggingface.co/allenai/OLMo-7B-Instruct).
+
+## Architecture
+
+| Property | Value |
+|----------|-------|
+| Parameters | 6.9B |
+| Hidden size | 4096 |
+| Attention | MHA, 32 heads |
+| Layers | 32 |
+| Intermediate | 11008 (fused: 22016 in checkpoint) |
+| Vocab | 50304 (padded from 50280) |
+| Activation | SwiGLU |
+| Position encoding | RoPE |
+| Normalization | LayerNorm, non-affine (no learnable params) |
+| Weight tying | No |
+
+## Key Implementation Details
+
+- **Fused QKV splitting**: HF checkpoint has `att_proj [12288, 4096]` which is split into Q, K, V (each `[4096, 4096]`)
+- **Fused MLP splitting**: `ff_proj [22016, 4096]` split into up (first half) and gate (second half), each `[11008, 4096]`. OLMo SwiGLU convention: `x, gate = chunk(2); silu(gate) * x`
+- **Non-affine LayerNorm**: No norm weights in state dict; uses manual mean/var ops for Neuron traceability
+- **Config format**: `model_type: "hf_olmo"` uses non-standard field names (d_model, n_heads, n_layers, etc.)
+
+## Compile & Validate
+
+```bash
+# Submit as SLURM job
+sbatch run_validation.sh
+
+# Or run directly on a Neuron instance
+python compile_and_validate.py
+
+# Compile only
+python compile_and_validate.py --compile-only
+
+# Validate only (requires existing compiled model)
+python compile_and_validate.py --validate-only
+```
+
+## Configuration
+
+- **TP degree**: 2
+- **Batch size**: 1
+- **Sequence length**: 128
+- **Dtype**: bfloat16
+
+## Validation Results
+
+| Metric | Value |
+|--------|-------|
+| Greedy token match | 93.75% (300/320) |
+| Teacher-forced match | **99.38%** |
+| Prompts tested | 10 |
+| Tokens per prompt | 32 |
+| Perfect prompts | 9/10 (100% match) |
+
+Validated on NXDI 0.6.0, Neuron SDK 2.x.

contrib/models/olmo-7b-instruct/src/__init__.py

@@ -0,0 +1,3 @@
+from .modeling_olmo import NeuronOlmoForCausalLM, OlmoInferenceConfig
+
+__all__ = ["NeuronOlmoForCausalLM", "OlmoInferenceConfig"]