Skip to content

Split-GPU Architecture

Run LLM on GPU 0 + Graphistry on GPU 1.

Architecture

GPU 0 (T4 - 15GB)          GPU 1 (T4 - 15GB)
┌──────────────────┐      ┌──────────────────┐
│  llama-server    │      │  RAPIDS cuDF     │
│  GGUF Model      │ ────>│  cuGraph         │
│  LLM Inference   │      │  Graphistry[ai]  │
│  ~5-12GB VRAM    │      │  Network Viz     │
└──────────────────┘      └──────────────────┘

Configuration

from llcuda.graphistry import SplitGPUConfig
import os

# Assign GPUs
config = SplitGPUConfig(
    llm_gpu=0,      # GPU 0 for LLM
    graph_gpu=1     # GPU 1 for Graphistry
)

# Set CUDA device for llama-server
os.environ['CUDA_VISIBLE_DEVICES'] = '0'

# Start llama-server on GPU 0
# (Graphistry will use GPU 1)

Use Cases

  1. Knowledge Graph Extraction
  2. LLM generates entities/relationships

  3. Graphistry visualizes graphs

  4. Interactive Analysis

  5. LLM answers questions

  6. Graphistry shows data patterns

  7. Multi-Modal Workflows

  8. Text generation (GPU 0)
  9. Graph analytics (GPU 1)