Add GPT-Neo-1.3B contrib model port

contrib/models/gpt-neo-1.3B/README.md

@@ -0,0 +1,112 @@
+# Contrib Model: GPT-Neo-1.3B
+
+NeuronX Distributed Inference implementation of GPT-Neo-1.3B.
+
+## Model Information
+
+- **HuggingFace ID:** `EleutherAI/gpt-neo-1.3B`
+- **Model Type:** Decoder-only transformer
+- **Parameters:** 1.3B
+- **License:** MIT
+
+## Architecture Details
+
+- GPT-Neo uses **unscaled dot-product attention** (no 1/sqrt(d_k) division). This is handled by overriding `prep_qkv_tensors` to pre-multiply Q by sqrt(head_dim), canceling the base class's mandatory scaling.
+- GPT-Neo alternates between **local** and **global** attention layers. For NXDI, all layers use full (global) KV cache for correctness.
+- Uses **absolute learned position embeddings** (not rotary).
+- Uses **tied weights** between token embeddings and LM head.
+
+## Validation Results
+
+**Validated:** 2026-03-16
+**Configuration:** TP=1, batch_size=1, seq_len=128, bf16
+
+### Test Results
+
+| Test | Status | Result |
+|------|--------|--------|
+| Compilation | PASS | Compiler status PASS |
+| Greedy Token Matching | PASS | **100% match on 10/10 prompts** (1185 tokens) |
+| Teacher-Forced Match | PASS | **100% average** |
+| Throughput | PASS | **81.5 tokens/sec** |
+
+**Status:** EXCELLENT - Perfect greedy and teacher-forced match across all prompts.
+
+## Usage
+
+```python
+from transformers import AutoTokenizer
+from neuronx_distributed_inference.models.config import NeuronConfig
+from neuronx_distributed_inference.utils.hf_adapter import load_pretrained_config
+
+from src.modeling_gpt_neo import NeuronGPTNeoForCausalLM, GPTNeoInferenceConfig
+
+model_path = "/path/to/gpt-neo-1.3B/"
+compiled_model_path = "/path/to/compiled/"
+
+# Configure
+neuron_config = NeuronConfig(
+    tp_degree=1,
+    batch_size=1,
+    seq_len=128,
+    torch_dtype=torch.bfloat16,
+)
+
+config = GPTNeoInferenceConfig(
+    neuron_config,
+    load_config=load_pretrained_config(model_path),
+)
+
+# Compile and load
+model = NeuronGPTNeoForCausalLM(model_path, config)
+model.compile(compiled_model_path)
+model.load(compiled_model_path)
+
+# Generate
+tokenizer = AutoTokenizer.from_pretrained(model_path)
+# ... (see integration test for full example)
+```
+
+## Compatibility Matrix
+
+| Instance/Version | 2.20+ | 2.19 and earlier |
+|------------------|-------|------------------|
+| Trn1             | PASS | Not tested |
+| Inf2             | Not tested | Not tested |
+
+## Performance
+
+Profiled on trn1.32xlarge (single NeuronCore utilization):
+
+| Metric | Context Encoding | Token Generation |
+|--------|-----------------|------------------|
+| Throughput | - | 81.9 tok/s |
+| MBU (Memory) | 18.1% | 15.2% |
+| MFU (Compute) | 9.4% | 0.1% |
+
+*Batch size 1, sequence length 128, BF16 precision, TP=1*
+
+## Testing
+
+Run integration tests:
+
+```bash
+pytest contrib/models/gpt-neo-1.3B/test/integration/test_model.py --capture=tee-sys
+```
+
+Or run manually:
+
+```bash
+cd contrib/models/gpt-neo-1.3B
+python3 test/integration/test_model.py
+```
+
+## Example Checkpoints
+
+* EleutherAI/gpt-neo-1.3B
+
+## Maintainer
+
+Neuroboros Team - Annapurna Labs
+
+**Last Updated:** 2026-03-16