Add MoLFormer molecular transformer contrib model

[jimburtoft]

Description

Add IBM MoLFormer-XL-both-10pct (~47M params, encoder-only molecular transformer) compiled with torch_neuronx.trace() for AWS Inferentia2. MoLFormer generates 768-dim embeddings from SMILES molecular strings, used in drug discovery and cheminformatics. Includes Jupyter notebook (compile, verify, benchmark), 17-test pytest suite, and README with measured results.

Model Information

Model Name: MoLFormer-XL-both-10pct
Model Architecture: Encoder-only transformer (custom, trust_remote_code=True)
Purpose: Molecular embedding generation from SMILES strings

Checklist

Please ensure your PR includes the following items. Refer to the contrib/CONTRIBUTING.md for detailed guidelines.

Required Components

• Accuracy Test (ex. test/integration/test_model.py)
  • At least one integration test that validates model accuracy
  • Uses logit validation or equivalent accuracy verification
  • Test can compile and run the model on Neuron
• README.md with the following sections:
  • Usage Example: Clear code example showing how to use the model
  • Compatibility Matrix: Table showing tested Neuron SDK versions and instance types (Trn1/Trn2/Inf2)
  • Example Checkpoints: Links to compatible model checkpoints (e.g., HuggingFace Hub)
  • Testing Instructions: Command to run the test suite for the model
• Source Code (src/)
  • N/A -- encoder-only model uses torch_neuronx.trace() directly (no NxDI modeling code). Inference logic is in the notebook and test file, following the same pattern as contrib/models/LaughterSegmentation/.

Optional Components

• Unit Tests (CPU or Neuron-based)
  • Tests for individual modeling components
  • Located in test/unit/ directory

Folder Structure

Confirm your contribution follows this structure:
/contrib/models/MoLFormer/
README.md
molformer_neuron_inf2.ipynb
/test
init.py
/integration
init.py
test_model.py
Note: No src/ directory -- this is an encoder-only traced model (same pattern as LaughterSegmentation). The modeling/inference code lives in the notebook and test file.

Testing

How did you test this change?
Deployed inf2.xlarge (2 NeuronCores) in sa-east-1 with SDK 2.28 DLAMI (Ubuntu 24.04). Compiled model with torch_neuronx.trace() using --auto-cast matmult --auto-cast-type bf16. Ran full pytest suite and standalone runner. Tested accuracy against CPU reference across 5 SMILES molecules (caffeine, aspirin, acetaminophen, cyclohexane, benzene). Benchmarked single-core, DataParallel, and various batch sizes.
Test Results:
============================= test session starts ==============================
platform linux -- Python 3.12.3, pytest-9.0.2, pluggy-1.6.0
collected 17 items
test_model.py::TestModelLoads::test_neuron_model_loads PASSED
test_model.py::TestModelLoads::test_output_shape PASSED
test_model.py::TestAccuracy::test_cosine_similarityCn1c(=O)c2c(ncn2C)n(C)c1=O PASSED
test_model.py::TestAccuracy::test_cosine_similarityCC(=O)Oc1ccccc1C(=O)O PASSED
test_model.py::TestAccuracy::test_cosine_similarityCC(=O)NC1=CC=C(O)C=C1 PASSED
test_model.py::TestAccuracy::test_cosine_similarityC1CCCCC1 PASSED
test_model.py::TestAccuracy::test_cosine_similarityc1ccccc1 PASSED
test_model.py::TestAccuracy::test_max_diffCn1c(=O)c2c(ncn2C)n(C)c1=O PASSED
test_model.py::TestAccuracy::test_max_diffCC(=O)Oc1ccccc1C(=O)O PASSED
test_model.py::TestAccuracy::test_max_diffCC(=O)NC1=CC=C(O)C=C1 PASSED
test_model.py::TestAccuracy::test_max_diffC1CCCCC1 PASSED
test_model.py::TestAccuracy::test_max_diffc1ccccc1 PASSED
test_model.py::TestDataParallel::test_dp_runs PASSED
test_model.py::TestDataParallel::test_dp_speedup PASSED
test_model.py::TestPerformance::test_single_core_throughput PASSED
test_model.py::TestPerformance::test_single_core_latency PASSED
test_model.py::TestPerformance::test_dp_throughput PASSED
============================= 17 passed in 41.20s ==============================

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

• MoLFormer requires trust_remote_code=True (custom modeling code from HuggingFace hub)
• The deterministic_eval=True flag is MoLFormer-specific (not a standard HF arg)
• --auto-cast matmult is critical for FP32 models: 65% single-core throughput gain (400 -> 661 inf/s), 54% smaller compiled model (160 MB -> 74 MB)
• Cosine similarity 0.999995 vs CPU reference (matmult bf16); 1.000000 for FP32
• Peak throughput: 1,660 inf/s (BS=4, DP=2); best low-latency: 1.50ms P50 (BS=1, single core)
• DataParallel scales 1.97x with 2 NeuronCores
• Neuron traced models return dict (without torchscript=True) or tuple (with it); test helpers handle both formats

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
• 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