InternVL3-78B-WSP-SFT-Compressed (8.68 GiB Multimodal Image + Video + Code/Security VLM)

This repository contains the Structurally Compressed, Weight-Space Programmed (WSP), and Jointly Multimodal Supervised Fine-Tuned (SFT) release of OpenGVLab/InternVL3-78B — compressed from 78.4B parameters (146.05 GiB) down to a single portable 8.68 GiB (9.32 GB) model.safetensors checkpoint (16.8x size reduction) that runs natively with Hugging Face transformers on both CPUs and GPUs.

---

Key Highlights & Compression Architecture

1. Multimodal Vision Encoder (InternViT-6B-448px-V2_5) Compression (45 $\rightarrow$ 8 Geodesic SLERP Layers, 12800 $\rightarrow$ 3200 MLP Neurons):
   • Preserves the full 3200-dimensional dynamic high-resolution patch embeddings (448x448 tiles + 0.5x pixel-shuffle) for Images, Multi-Frame Videos, UI/UX Wireframes, Network Topology Diagrams, and Dense OCR Screenshots.
   • Fuses 8 representative layer pairs ((0,1), (6,7), (12,13), (18,19), (24,25), (31,32), (37,38), (43,44)) via Geodesic Spherical Linear Interpolation (SLERP) with LayerScale depth-variance compensation and Wanda/L2-Product Structured Neuron Pruning (12,800 $\rightarrow$ 3,200).
2. Full-Width Multimodal Projector (mlp1) Preserved & Fine-Tuned:
   • Retains 100% of the 12800 -> 8192 Pixel-Shuffle Vision-Language MLP projector (LayerNorm(12800) -> Linear(12800, 8192) -> GELU -> Linear(8192, 8192)).
3. Language Model (Qwen2.5-72B) Compression (80 $\rightarrow$ 6 Geodesic SLERP Layers, 29568 $\rightarrow$ 4096 SwiGLU Neurons):
   • Preserves the full 8192-dimensional hidden state space and GQA (64Q / 8KV) attention across 6 SLERP-fused layer pairs ((1,2), (15,16), (31,32), (48,49), (64,65), (78,79)) with Wanda structured SwiGLU neuron pruning and Spectral Subspace Conditioning.
4. 8-Epoch Joint Multimodal & Domain Supervised Fine-Tuning (SFT):
   • Trained across 16 specialized curricula covering:
     • Multimodal Image & Video Processing (Spatio-temporal keyframe pipelines, UI/UX wireframe-to-React/Tailwind synthesis, Cloud Network Topology diagram security audits, HTTP/JWT visual capture triage)
     • Software Engineering & Distributed System Design (Consistent hashing, transactional outbox, Raft consensus, Python asyncio)
     • Network Engineering & CEH (OSI/TCP-IP defense, eBPF/XDP SYN flood mitigation, Zero-Trust Kubernetes NetworkPolicies)
     • Penetration Testing Methodology & Bug Bounty / Web Security Remediation (SSRF CWE-918, TOCTOU race conditions CWE-362, Prototype Pollution, OAuth2 PKCE, GraphQL hardening, Cloud IAM least-privilege)

---

Quick Start

1. Run with the Bundled Multimodal Runner (run_internvl3.py)

pip install torch torchvision transformers safetensors pillow opencv-python-headless

# Image + Prompt
python run_internvl3.py --image examples/image1.jpg --prompt "Describe this image in detail."

# Video + Prompt
python run_internvl3.py --video examples/red-panda.mp4 --prompt "Summarize the temporal events across the video frames."

# Text-only Coding / Cybersecurity Prompt
python run_internvl3.py --prompt "Design a DNS-rebinding-safe SSRF validator in Python."

2. Run Directly via Hugging Face transformers

import torch
from transformers import AutoModel, AutoTokenizer

model_id = "EverythingIsForGod/InternVL3-78B-WSP-SFT-Compressed"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True, use_fast=False)
model = AutoModel.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
    trust_remote_code=True,
    use_flash_attn=False,
).eval()

response = model.chat(
    tokenizer,
    pixel_values=None,
    question="Explain how to prevent TOCTOU race conditions in PostgreSQL.",
    generation_config=dict(max_new_tokens=256, do_sample=False),
)
print(response)