Contrib: FLUX.1-lite-8B-alpha (native FLUX.1 compatibility)

[jimburtoft]

Summary

• FLUX.1-lite-8B-alpha (Freepik) is architecturally identical to FLUX.1-dev with 8 double-stream MMDiT blocks instead of 19. All other components (CLIP + T5-XXL encoders, VAE, scheduler, RoPE) are the same.
• NxDI's first-party FLUX.1 implementation reads num_layers from the model's config.json at runtime, so FLUX.1-lite works out of the box with no custom modeling code.
• This contrib provides a standalone generation script, integration tests, and documentation demonstrating native compatibility.

Validation Results (trn2.3xlarge, LNC=2, TP=4)

Metric	Value
Resolution	1024x1024
Inference steps	25
E2E generation time	5.91s avg
Pipeline steps/sec	4.23
Backbone forward/sec	4.49
Compilation time	~128s

Checklist

Model Type

• Diffusion/image generation model

Contribution Contents

• README with model info, benchmarks, usage instructions
• src/ directory with generation script
• test/integration/test_model.py with 3 passing tests
• Sample output image
• vLLM integration (N/A -- diffusion model)

Testing

• All code tested on Neuron hardware (trn2.3xlarge)
• All numbers in README are measured, not estimated
• Integration tests pass: smoke test, image generation, timing

SDK Compatibility

• Neuron SDK 2.29 (DLAMI 20260410)
• NxD Inference 0.9
• PyTorch 2.9