Source snapshot of github.com/Scottcjn/ltx-video-power8, also on GitLab with full history. Code only: nothing here runs on Hugging Face infrastructure.
LTX-Video 13B on IBM POWER8
Run state-of-the-art AI video generation on IBM POWER8 servers with 320GB RAM
Features • Quick Start • Hardware • Results • Technical Details
Why This Matters
- First Ever: No one has run a 13B parameter video diffusion model on PowerPC before
- Legacy Hardware Revival: Your IBM POWER8 server can now generate AI videos
- 320GB RAM Advantage: Fits entire model in memory - no GPU required
- Open Source: Full pipeline, not just inference scripts
Features
- QK-Norm Key Mapping for distilled 13B model architecture
- Latent Packing/Unpacking for transformer input format
- POWER8 Stack Workaround for multi-threading stability
- Hybrid Threading - 5.7× speedup with smart thread switching
Quick Start
# Clone the repo
git clone https://github.com/Scottcjn/ltx-video-power8.git
cd ltx-video-power8
# Download models (you need both)
# 1. 13B distilled model → ~/models/ltx-video-13b/ltxv-13b-0.9.8-distilled.safetensors
# 2. Full LTX-Video model → ~/models/ltx-video-full/
# Run inference
cd scripts
python3 ltx_13b_hybrid.py
Hardware Requirements
| Component | Minimum | Recommended |
|---|---|---|
| CPU | IBM POWER8 | POWER8 S824 (dual 8-core) |
| RAM | 64GB | 320GB |
| Storage | 50GB | 100GB SSD |
| OS | Ubuntu 20.04 | Ubuntu 20.04 LTS |
Model Architecture
The 13B distilled LTX-Video model uses different key names than standard diffusers:
| Checkpoint Key | Diffusers Key |
|---|---|
patchify_proj.* |
proj_in.* |
attn*.q_norm.* |
attn*.norm_q.* |
attn*.k_norm.* |
attn*.norm_k.* |
adaln_single.* |
time_embed.* |
Our scripts handle this mapping automatically.
Scripts
ltx_13b_full.py
Complete single-threaded pipeline. Safe but slow.
- Resolution: 256×256 | Frames: 9 | Steps: 4
- Time: ~30 seconds
ltx_13b_hybrid.py ⭐ Recommended
Hybrid multi-threaded pipeline with 5.7× speedup.
- Multi-threading for transformer (safe on POWER8)
- Single-thread for VAE decode (required)
- Time: ~65 seconds vs 370 seconds
ltx_13b_hires.py
High-resolution generation.
- Resolution: 512×512 | Frames: 17 | Steps: 8
- Time: ~54 minutes
Results
| Configuration | Resolution | Frames | Time | Output |
|---|---|---|---|---|
| Single-thread | 256×256 | 9 | 30s | 511KB |
| Hybrid | 256×256 | 9 | 65s | 531KB |
| High-res | 512×512 | 17 | 54min | 3.9MB |
Example Outputs
 256×256 Preview "A glowing crystal rotating slowly in darkness" |
 512×512 High Resolution "A majestic phoenix rising from flames" |
Technical Details
POWER8 Stack Corruption Workaround
POWER8 exhibits stack smashing errors with multi-threaded PyTorch in certain code paths (particularly VAE decode):
# Force single-threaded for VAE
os.environ["OMP_NUM_THREADS"] = "1"
torch.set_num_threads(1)
The hybrid script dynamically switches between multi-threaded (transformer) and single-threaded (VAE) modes.
Latent Packing
The diffusers transformer expects pre-patchified input:
def pack_latents(latents, patch_size=1, patch_size_t=1):
# [B, C, F, H, W] → [B, F*H*W, C]
...
RoPE Dimensions
Rotary position embeddings expect latent-space dimensions, not video dimensions:
# Correct: pass latent dimensions
num_frames = (FRAMES - 1) // 8 + 1 # latent frames
height = RESOLUTION // 32 # latent height
width = RESOLUTION // 32 # latent width
Dependencies
torch>=2.1.0
diffusers>=0.32.0
transformers>=4.39.0
safetensors
pillow
numpy
Related Projects
| Project | Description |
|---|---|
| llama-cpp-power8 | AltiVec/VSX optimized llama.cpp for POWER8 |
| nvidia-power8-patches | Modern NVIDIA drivers for POWER8 via OCuLink |
| power8-projects | Ubuntu 22.04 build, PSE LLM, Darwin cross-compile |
Acknowledgments
- Lightricks for the LTX-Video model
- Hugging Face for diffusers library
- IBM for POWER8 architecture documentation
License
MIT License
Made with ⚡ by Elyan Labs
Proving that legacy hardware can run cutting-edge AI