[contrib] Add Bloom-1b7 NeuronX port

[dhwanw]

Description

NeuronX Distributed Inference port of bigscience/bloom-1b7, a 1.7B-parameter decoder-only transformer. Bloom uses ALiBi positional encoding (no position embeddings), fused interleaved QKV weights, embedding LayerNorm, and weight-tied lm_head. ALiBi bias is computed and added in custom perform_prefill and compute_for_token_gen overrides.

Model Information

Model Name: Bloom-1b7
Model Architecture: Decoder-only transformer with ALiBi attention, 16 MHA heads, 24 layers, LayerNorm (not RMSNorm), embedding LayerNorm, GELU (tanh approximation), tied embeddings
Purpose: Multilingual text generation

Checklist

Required Components

• Accuracy Test (test/integration/test_model.py)
  • Validates model generation and coherence
  • Performance benchmarks (TTFT, throughput)
  • 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
  • Example Checkpoints: Links to compatible model checkpoints
  • Testing Instructions: Command to run the test suite for the model
• Source Code (src/)
  • Modeling code following NxD Inference patterns

Optional Components

• Unit Tests (CPU or Neuron-based)

Folder Structure

/contrib/models/bloom-1b7/
  README.md
  /src
    modeling_bloom.py
  /test
    /integration
      test_model.py

Testing

Model was compiled and tested with TP=1, batch_size=1, seq_len=128, bfloat16 on trn1.32xlarge.

Test Results:

Test	Status	Result
Smoke Test	✅ PASS	Model loads successfully
Greedy Token Matching	✅ PASS	76.9% average (vs HF bfloat16 reference)
Throughput	✅ PASS	63.3 tok/s

Compatibility

Tested with:

• Neuron SDK Version(s): 2.22
• Instance Type(s): trn1.32xlarge
• PyTorch Version: 2.9
• Python Version: 3.10
• Configuration: TP=1, batch_size=1, seq_len=128, bfloat16

Additional Information

• ALiBi attention: No position embeddings. Per-head slopes are used to compute position bias added to attention scores. Implemented via custom perform_prefill and compute_for_token_gen overrides.
• Fused QKV: HF checkpoint stores interleaved [num_heads, 3, head_dim, hidden_size] QKV weights, split into separate Q, K, V during weight conversion.
• Embedding LayerNorm: Applied after token embeddings via get_model_output override.
• Weight tying: LM head tied to token embeddings.

Related Issues

N/A

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