Add LaughterSegmentation to contrib

[jimburtoft]

Description

Add LaughterSegmentation, a Wav2Vec2-based laughter detection model compiled with torch_neuronx.trace(). Includes a self-contained Jupyter notebook with compilation, single-core and DataParallel benchmarks, accuracy validation against CPU reference, and an end-to-end inference demo. Also documents a weight_norm parametrization crash workaround required on SDK 2.28+.

Model Information

Model Name: LaughterSegmentation (omine-me/LaughterSegmentation)
Model Architecture: Wav2Vec2ForAudioFrameClassification (~315M params, FP32)
Purpose: Audio frame classification -- detects laughter segments in speech audio
Checklist

Required Components

• Accuracy Test (test/integration/test_model.py)
  • 12 tests total: smoke tests, cosine similarity validation across 5 input types, frame-level prediction agreement, DataParallel correctness & speedup, throughput & latency benchmarks
  • All 12 pass on inf2.xlarge (SDK 2.28, PyTorch 2.9)
• README.md with the following sections:
  • Usage Example: Jupyter notebook workflow with bash quickstart
  • Compatibility Matrix: inf2.xlarge validated, trn2.3xlarge tested
  • Example Checkpoints: Link to HuggingFace model
  • Testing Instructions: pytest and standalone runner commands
• Source Code (src/)
  • N/A -- this model uses torch_neuronx.trace() directly rather than NxD Inference modeling code. All logic is in the notebook and test file.

Optional Components

• Unit Tests -- test/unit/ directory created (empty init.py, consistent with other contrib models)
  Folder Structure
  /contrib/models/LaughterSegmentation/
  README.md
  laughter_neuron_inf2.ipynb # notebook with executed outputs
  /test
  init.py
  /unit
  init.py
  /integration
  init.py
  test_model.py

Testing

How did you test this change?
Ran pytest on a fresh inf2.xlarge instance (sa-east-1) with the Deep Learning AMI Neuron (Ubuntu 24.04) 20260227 and the pre-installed PyTorch inference venv.
source /opt/aws_neuronx_venv_pytorch_inference_vllm_0_13/bin/activate
pip install safetensors pytest
pytest test_model.py --capture=tee-sys -v
Test Results:
test_model.py::TestModelLoads::test_neuron_model_loads PASSED [ 8%]
test_model.py::TestModelLoads::test_neuron_model_runs PASSED [ 16%]
test_model.py::TestAccuracy::test_cosine_similarity[random_normal] PASSED [ 25%]
test_model.py::TestAccuracy::test_cosine_similarity[quiet_noise] PASSED [ 33%]
test_model.py::TestAccuracy::test_cosine_similarity[loud_signal] PASSED [ 41%]
test_model.py::TestAccuracy::test_cosine_similarity[sine_440hz] PASSED [ 50%]
test_model.py::TestAccuracy::test_cosine_similarity[silence] PASSED [ 58%]
test_model.py::TestAccuracy::test_frame_agreement PASSED [ 66%]
test_model.py::TestDataParallel::test_data_parallel_runs PASSED [ 75%]
test_model.py::TestDataParallel::test_data_parallel_speedup PASSED [ 83%]
test_model.py::TestPerformance::test_throughput PASSED [ 91%]
test_model.py::TestPerformance::test_latency PASSED [100%]
======================= 12 passed, 2 warnings in 51.49s =======================

Key metrics:

• Cosine similarity: >= 0.999999 across all inputs
• Frame agreement: 100%
• Single core: 101.5 W/s, 9.85 ms p50 latency
• DataParallel (2 cores): 176.5 W/s, 1.88x speedup
  Compatibility

Tested with:

• Neuron SDK Version(s): 2.28
• Instance Type(s): inf2.xlarge
• PyTorch Version: 2.9.0
• Python Version: 3.12.3
  Additional Information
• This model uses torch_neuronx.trace() rather than NxD Inference, since it is an encoder-only classification model (not an autoregressive LLM). It does not have a src/ directory with modeling code.
• weight_norm workaround: Wav2Vec2 uses weight_norm parametrizations that crash torch_neuronx.trace() on SDK 2.28+. The notebook and test strip parametrizations before tracing. This affects any HuggingFace model using weight_norm.
• The notebook also runs on trn2.3xlarge by uncommenting the --lnc compiler arg.

Related Issues

None

vLLM Integration

• This model/feature is intended for use with vLLM
• Documentation includes vLLM registration instructions

---

By submitting this PR, I confirm that:

• I have read and followed the contributing guidelines (../contrib/CONTRIBUTING.md)
• This is a community contribution and may have limited testing compared to officially-supported models
• The code follows best practices and is well-documented
• All required components listed above are included