sdxl-trtx-bf16-ada

SDXL TensorRT-RTX: BF16-Ada 🚀 Ready

Optimized TensorRT-RTX engines for SDXL on Ada architecture with BF16 quantization.

🎯 This Repository

One variant, one download - only get exactly what you need!

• Model: SDXL
• Architecture: Ada (Compute Capability 8.0+)
• Quantization: BF16
• Memory: ~4-6GB VRAM
• Speed: ~4.1s (RTX 4090) for 1024x1024 generation

🚀 Quick Start

Automatic (Recommended)

# ImageAI server downloads automatically
curl -X POST "http://localhost:8001/generate" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "a beautiful landscape",
    "model": "sdxl-tensorrt_rtx:bf16",
    "width": 1024,
    "height": 1024
  }'

Manual Download

from huggingface_hub import snapshot_download

# Download this specific variant only
engines_path = snapshot_download(
    repo_id="imgailab/sdxl-trtx-bf16-ada"
)

# Engines are in: engines_path/engines/*.plan

Direct Integration

from imageai_server.tensorrt.nvidia_sdxl_pipeline import NVIDIASDXLPipeline

pipeline = NVIDIASDXLPipeline()
pipeline.load_engines(
    engine_dir=f"{engines_path}/engines",
    framework_model_dir=f"{engines_path}/framework",  
    onnx_dir=f"{engines_path}/onnx"
)
pipeline.activate_engines()

images, time_ms = pipeline.infer(
    prompt="a serene mountain landscape",
    height=1024,
    width=1024
)

📊 Performance

🔧 Requirements

Hardware

• GPU: Ada architecture
  • Ampere: RTX 3090, A100, etc.
  • Ada Lovelace: RTX 4090, etc.
  • Blackwell: H200, etc.
• VRAM: ~4-6GB minimum
• Compute Capability: 8.0+

Software

• TensorRT-RTX: 1.0.0.21+
• CUDA: 12.0+
• Python: 3.8+

📁 Repository Structure

sdxl-trtx-bf16-ada/
├── engines/           # TensorRT engine files
│   ├── *.plan        # Optimized engines
├── config.json       # Configuration metadata
└── README.md         # This file

🌐 Related Repositories

Other variants for SDXL:

• Ampere BF16\n- Ada FP8\n- Blackwell FP4\n- Blackwell FP8\n- Blackwell BF16\n

📝 License

Inherits license from base model: stabilityai/stable-diffusion-xl-base-1.0

🔄 Updates

• 2025-08-12: Initial release
• Optimized for single-variant downloads

---

Part of the ImageAI TensorRT-RTX engine collection