Lance-2080ti Breathes New Life Into Modded Turing GPUs For AI Video

    
        By vramkickedin    
     | 
    
            June 2, 2026 at 12:44 pm        
    
     | 
    
        2 min read

Lance-2080ti is a hardware-optimized fork of ByteDance’s Lance model, tailored specifically for RTX 2080 Ti cards with 22GB of memory. It runs video and image generation without relying on BF16 precision or FlashAttention, instead using FP16 inference and compatible attention fallback paths. The release includes carefully tested single-GPU and dual-GPU configurations that squeeze stable performance out of older Turing architecture.

Open-source developer lvyufeng built this adaptation to help the many homelab builders who use affordable, modded 2080 Ti cards. Standard Lance workflows fail on these GPUs because they lack native BF16 support and cannot run FlashAttention. This fork removes that dependency and adds memory controls, multi-GPU parallelism, and validated presets so users can generate content locally without buying newer hardware.

Purpose-built for Turing GPUs

Key Features

Runs without BF16 or FlashAttention on Turing.
FP16 inference with SDPA/eager attention paths.
Gradio memory controls for 22GB single-GPU setups.
Optional 2-GPU language-model layer parallelism.
VAE offload to a second GPU.
Validated 360p frame presets up to 49 frames.
Profiling tools and experimental SM75 custom ops.
Auto-switches image/video models in memory.

This tool is for budget-conscious enthusiasts who want to run modern generative AI locally on RTX 2080 Ti 22GB cards. It lets you generate short videos and images while keeping all data private, without cloud fees or a $1,500+ GPU upgrade. By following the tested launch scripts, you can reliably produce 360p clips with a single card or unlock 480p 49-frame videos when pairing two cards together.

What's different from the original Lance

The fork strips out FlashAttention and enforces FP16 pathways to stay within Turing’s hardware limits. Single-GPU users should expect 360p/97 frames to run out of memory, and 480p will fail even at 13 frames unless VAE decoding is shifted to a second GPU. The two-GPU pipeline parallelism places decoder layers 0–17 on GPU0 and layers 18–35 on GPU1, with VAE offload on the second card, cutting generation time almost in half for heavier workloads. All experimental SM75 custom ops remain optional and are not enabled by default—real-world correctness and speed should guide whether you flip those switches.