Add git-base contrib model port

contrib/models/git-base/README.md

@@ -0,0 +1,144 @@
+# Contrib Model: Git-base
+
+NeuronX Distributed Inference implementation of Microsoft's Git (Generative Image-to-text Transformer).
+
+## Model Information
+
+- **HuggingFace ID:** `microsoft/git-base`
+- **Model Type:** Vision-language model (text decoder + vision encoder compiled on Neuron)
+- **Parameters:** ~130M (text decoder) + ~86M (CLIP ViT-B/16 vision encoder)
+- **License:** MIT
+
+## Architecture Details
+
+Git uses a BERT-style text decoder with several distinguishing features:
+
+- **Post-LayerNorm residual blocks** (BERT-style, not pre-LN like GPT/LLaMA): LayerNorm is applied after the residual addition, not before
+- **Learned absolute position embeddings** (no rotary embeddings)
+- **Embedding LayerNorm** applied after combining token + position embeddings
+- **Separate Q/K/V projections with bias** in all attention and MLP layers
+- **GELU activation** in MLP
+- **No final layer norm** (post-LN per block handles normalization)
+- **CLIP ViT-B/16 vision encoder** compiled on Neuron, projects to text hidden size via visual_projection
+
+| Property | Value |
+|----------|-------|
+| Hidden Size | 768 |
+| Num Attention Heads | 12 (MHA) |
+| Num Hidden Layers | 6 |
+| Intermediate Size | 3072 |
+| Vocab Size | 30522 |
+| Vision Encoder | CLIP ViT-B/16@224, 197 tokens (196 patches + 1 CLS) |
+
+## Available Implementations
+
+### 1. Text-only (`modeling_git.py`)
+Only the text decoder is compiled on Neuron. Vision encoder is skipped.
+
+### 2. Vision+Text (`modeling_git_vision.py`)
+Both CLIP vision encoder and text decoder are compiled as separate NEFFs on Neuron. Uses `NeuronBaseForImageToText` infrastructure.
+
+## Validation Results
+
+**Validated:** 2026-03-19
+**Configuration:** TP=1, batch_size=1, text_seq=256, vision_seq=197, fp32
+
+### Test Results
+
+| Test | Status | Result |
+|------|--------|--------|
+| Compilation | PASS | Text + Vision NEFFs compiled in 95s |
+| Teacher-Forced Match | PASS | **95.13% avg** (5 random images, 20 tokens each) |
+| Greedy Token Matching | PASS | **46.41% avg** (2/5 images 100%, cascading divergence on others) |
+
+### COCO Image Captioning Demo
+
+Captions generated from real COCO val2017 photos (greedy decoding, max 30 tokens):
+
+| Image | HF Golden | Neuron |
+|-------|-----------|--------|
+| Two cats on a couch | two cats laying on a pink blanket | two cats laying on a couch |
+| Person skateboarding | a woman bending over | a woman bending over |
+| Bus on street | a skateboarder doing a trick | skateboarder doing a trick |
+| Kitchen / food | kitchen with a breakfast bar | kitchen with a breakfast bar |
+| Solid red (synthetic) | the red light of the light | the red light of the car |
+| Checkerboard (synthetic) | black and white checkered pattern | checkered pattern in the middle |
+
+## Usage (Vision+Text)
+
+```python
+import torch
+from neuronx_distributed_inference.models.config import NeuronConfig
+from src.modeling_git_vision import (
+    GitVisionInferenceConfig,
+    NeuronGitForCausalLMVision,
+)
+
+model_path = "/path/to/git-base/"
+compiled_path = "/path/to/compiled_vision/"
+
+text_nc = NeuronConfig(
+    tp_degree=1, batch_size=1, ctx_batch_size=1, tkg_batch_size=1,
+    max_context_length=256, seq_len=256, n_active_tokens=1,
+    torch_dtype=torch.float32, padding_side="right",
+    enable_bucketing=True, save_sharded_checkpoint=True,
+)
+vision_nc = NeuronConfig(
+    tp_degree=1, batch_size=1, ctx_batch_size=1, tkg_batch_size=1,
+    seq_len=197, n_active_tokens=197,
+    torch_dtype=torch.float32, padding_side="right",
+    enable_bucketing=False, buckets=[197], save_sharded_checkpoint=True,
+)
+
+config = GitVisionInferenceConfig.from_pretrained(
+    model_path, text_neuron_config=text_nc, vision_neuron_config=vision_nc,
+)
+model = NeuronGitForCausalLMVision(model_path, config)
+model.compile(compiled_path)
+model.load(compiled_path)
+
+# Image captioning with forward_atomic_prefill + TKG loop
+# See contrib/run_token_match_git_vision.py for full example
+```
+
+## 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 | - | 714.4 tok/s |
+| MBU (Memory) | 11.7% | 12.4% |
+| MFU (Compute) | 4.3% | 0.1% |
+
+*Batch size 1, sequence length 128, BF16 precision, TP=1*
+
+## Testing
+
+Run integration tests:
+
+```bash
+pytest contrib/models/git-base/test/integration/test_model.py --capture=tee-sys
+```
+
+Or run manually:
+
+```bash
+cd contrib/models/git-base
+python3 test/integration/test_model.py
+```
+
+## Example Checkpoints
+
+* microsoft/git-base
+
+## Maintainer
+
+Neuroboros Team - Annapurna Labs