Add openai-gpt (GPT-1) contrib model port

contrib/models/openai-gpt/README.md

@@ -0,0 +1,111 @@
+# Contrib Model: openai-gpt
+
+NeuronX Distributed Inference implementation of OpenAI GPT-1.
+
+## Model Information
+
+- **HuggingFace ID:** `openai-community/openai-gpt`
+- **Model Type:** Decoder-only transformer (117M parameters)
+- **License:** Check HuggingFace model card
+
+## Architecture Details
+
+OpenAI GPT-1 has several architectural differences from later GPT variants:
+
+- **Post-norm:** LayerNorm is applied AFTER residual addition (not before, as in GPT-2/Neo)
+- **No final LayerNorm:** The last transformer block's post-norm handles normalization
+- **Conv1D weights:** HuggingFace checkpoint uses Conv1D format (transposed from nn.Linear)
+- **Fused QKV:** Single `c_attn` weight matrix for Q, K, V projections
+- **Standard GELU:** Uses erf-based GELU activation (not tanh-approximate `gelu_new`)
+- **Absolute position embeddings:** Learned position embeddings up to 512 tokens
+
+## Usage
+
+```python
+from transformers import AutoTokenizer
+from neuronx_distributed_inference.models.config import NeuronConfig
+
+# Import model classes from src
+from src.modeling_openai_gpt import NeuronOpenAIGPTForCausalLM, OpenAIGPTInferenceConfig
+
+model_path = "/path/to/openai-gpt/"
+compiled_model_path = "/path/to/compiled/"
+
+# Configure
+neuron_config = NeuronConfig(
+    tp_degree=1,
+    batch_size=1,
+    seq_len=128,
+    torch_dtype=torch.bfloat16,
+)
+
+config = OpenAIGPTInferenceConfig.from_pretrained(
+    model_path, neuron_config=neuron_config,
+)
+
+# Compile and load
+model = NeuronOpenAIGPTForCausalLM(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)
+```
+
+## Validation Results
+
+**Validated:** 2026-03-16
+**Configuration:** TP=1, batch_size=1, seq_len=128, bfloat16
+
+### Test Results
+
+| Test | Status | Result |
+|------|--------|--------|
+| Compilation | PASS | Context encoding + token generation compiled |
+| Token Matching | PASS | **100.0% match** (200/200 tokens, 10 prompts) |
+| Profiling | PASS | **437.4 tokens/sec** |
+
+**Status:** EXCELLENT
+
+## Compatibility Matrix
+
+| Instance/Version | 2.20+ | 2.19 and earlier |
+|------------------|-------|------------------|
+| Trn1             | Working | Not tested |
+| Inf2             | Not tested | Not tested |
+
+## Performance
+
+Profiled on trn1.32xlarge (single NeuronCore utilization):
+
+| Metric | Context Encoding | Token Generation |
+|--------|-----------------|------------------|
+| Throughput | - | 449.3 tok/s |
+| MBU (Memory) | 11.0% | 10.8% |
+| MFU (Compute) | 4.5% | 0.0% |
+
+*Batch size 1, sequence length 128, BF16 precision, TP=1*
+
+## Testing
+
+Run integration tests:
+
+```bash
+pytest contrib/models/openai-gpt/test/integration/test_model.py --capture=tee-sys
+```
+
+Or run manually:
+
+```bash
+cd contrib/models/openai-gpt
+python3 test/integration/test_model.py
+```
+
+## Example Checkpoints
+
+* openai-community/openai-gpt
+
+## Maintainer
+
+Neuroboros Team - Annapurna Labs