Add CTRL contrib model port

contrib/models/ctrl/README.md

@@ -0,0 +1,118 @@
+# Contrib Model: CTRL
+
+NeuronX Distributed Inference implementation of Salesforce CTRL.
+
+## Model Information
+
+- **HuggingFace ID:** `Salesforce/ctrl`
+- **Model Type:** Decoder-only transformer (conditional text generation)
+- **Parameters:** 1.63B
+- **Architecture:** 48 layers, 16 heads, hidden_size=1280, dff=8192
+- **Position Encoding:** Sinusoidal (computed, not learned)
+- **Activation:** ReLU
+- **License:** BSD-3-Clause
+
+## Architecture Details
+
+CTRL uses control codes (e.g., `Links`, `Wikipedia`, `Science`) prepended to prompts to condition generation on a particular domain/style. Key differences from standard GPT-2:
+
+- Sinusoidal position embeddings (not learned)
+- Input embeddings scaled by sqrt(d_model)
+- Separate lm_head with bias
+- ReLU activation (not GELU)
+- Pre-LayerNorm architecture
+
+## Validation Results
+
+**Validated:** 2026-03-16
+**Configuration:** TP=1, batch_size=1, seq_len=128, bf16
+
+### Test Results
+
+| Test | Status | Result |
+|------|--------|--------|
+| Smoke Test | PASS | Model loads successfully |
+| Token Matching (greedy) | PASS | **99.06% match** (634/640 tokens) |
+| Token Matching (teacher-forced) | PASS | **99.84% match** |
+| Profiling | PASS | **59.2 tokens/sec** |
+
+**Status:** EXCELLENT
+
+## Usage
+
+```python
+import torch
+from neuronx_distributed_inference.models.config import NeuronConfig
+from modeling_ctrl import NeuronCTRLForCausalLM, CTRLInferenceConfig
+
+model_path = "/path/to/ctrl/"
+compiled_model_path = "/path/to/compiled/"
+
+# Configure
+neuron_config = NeuronConfig(
+    tp_degree=1,
+    batch_size=1,
+    seq_len=128,
+    torch_dtype=torch.bfloat16,
+)
+
+config = CTRLInferenceConfig.from_pretrained(
+    model_path, neuron_config=neuron_config,
+)
+
+# Compile and load
+model = NeuronCTRLForCausalLM(model_path, config)
+model.compile(compiled_model_path)
+model.load(compiled_model_path)
+
+# Generate with control code
+from transformers import AutoTokenizer
+tokenizer = AutoTokenizer.from_pretrained(model_path)
+inputs = tokenizer("Wikipedia The theory of relativity", return_tensors="pt")
+# ... (see integration test for full example)
+```
+
+## Compatibility Matrix
+
+| Instance/Version | 2.20+ | 2.19 and earlier |
+|------------------|-------|------------------|
+| Trn1             | PASS (99.84% TF) | Not tested |
+| Inf2             | Not tested | Not tested |
+
+## Performance
+
+Profiled on trn1.32xlarge (single NeuronCore utilization):
+
+| Metric | Context Encoding | Token Generation |
+|--------|-----------------|------------------|
+| Throughput | - | 59.3 tok/s |
+| MBU (Memory) | 16.7% | 15.0% |
+| MFU (Compute) | 7.6% | 0.1% |
+
+*Batch size 1, sequence length 128, BF16 precision, TP=1*
+
+## Testing
+
+Run integration tests:
+
+```bash
+pytest contrib/models/ctrl/test/integration/test_model.py --capture=tee-sys
+```
+
+Or run manually:
+
+```bash
+cd contrib/models/ctrl
+python3 test/integration/test_model.py
+```
+
+## Example Checkpoints
+
+* Salesforce/ctrl
+
+## Maintainer
+
+Neuroboros Team - Annapurna Labs
+
+**Last Updated:** 2026-03-16
+**Validation Job:** SLURM 9589