Add InternVL3-8B-Instruct contrib model

[jimburtoft]

Summary

• Add InternVL3-8B-Instruct VLM contrib (InternViT-300M vision encoder + Qwen2.5-7B text backbone)
• Validated on trn2.3xlarge TP=4 with SDK 2.29: 75.1 tok/s decode, 138ms TTFT, cosine 0.9984 vs CPU
• 1.85x faster output throughput vs NVIDIA L40S GPU
• Includes vision encoder compilation, text backbone with vision embedding injection, vLLM integration patches, accuracy tests, and GPU benchmark scripts

Model Details

Component	Details
Vision encoder	InternViT-300M-448px-V2.5 (24 layers, traced via torch_neuronx)
Projector	Pixel shuffle + 2-layer MLP
Text backbone	Qwen2.5-7B (NxDI NeuronBaseForCausalLM)
Framework	NeuronBaseForImageToText

Files

• src/ — 3-file VLM implementation (main, text, vision)
• test/ — Integration tests
• vllm/ — vLLM-neuron patches for serving
• Benchmark/validation scripts (accuracy, scaling, NKI, GPU comparison)

Validation

• CTE logit comparison vs CPU FP32: cosine=0.9984, top-1 match
• TKG text generation: correct output
• Multimodal (text + image): end-to-end working
• Sequence lengths: 2K-32K validated