Views
No views yet
float8_e4m3fn, per-tensor scale) — 32GB → 8GB (4x)| Configuration | GPU | Peak VRAM | Status |
|---|---|---|---|
| BF16 baseline | A800 80GB | 73.8 GB | ✅ |
| BF16 baseline | RTX 4090 48GB | OOM (46.5 GB) | ❌ |
| FP8 + turbo3 | RTX 4090 48GB | 37.4 GB | ✅ |
| Chunk | Time/step |
|---|---|
| 0 | ~0.86s |
| 4 | ~2.4s |
| 7 | ~2.8s |
| Total | 196.8s (4 steps × 8 chunks) |
diffusion_pytorch_model.fp8.safetensors — FP8 quantized transformer weights (8GB)scripts/native_fp8_patch.py — FP8 Linear layer with torch._scaled_mmscripts/turbo3_integration.py — V cache PolarQuant compression (GPU optimized)scripts/run_fp8_turbo3_gpu.py — inference wrapperscripts/run_fp8_turbo3_gpu.sh — one-click launch scriptscripts/batch_inference.py — batch inference with random WASD posesvideos/ — generated video samples1# 1. Clone HY-WorldPlay
2git clone https://github.com/Tencent/HunyuanVideo.git
3cd HunyuanVideo
4
5# 2. Download this repo's FP8 weights
6# Place diffusion_pytorch_model.fp8.safetensors in your model directory
7
8# 3. Run inference
9bash scripts/run_fp8_turbo3_gpu.sh1import safetensors.torch
2import torch
3
4# Load FP8 quantized weights
5state_dict = safetensors.torch.load_file("diffusion_pytorch_model.fp8.safetensors")
6
7# Weights with dtype float8_e4m3fn are quantized
8# Corresponding *_scale tensors contain per-tensor scales
9# Dequantize: weight_bf16 = fp8_weight.to(bfloat16) * weight_scale| Optimization | Cosine Similarity | Verified |
|---|---|---|
| FP8 weights | > 0.999 | ✅ |
| V cache turbo3 (3-bit) | 0.983 | ✅ (A800 real KV cache) |
| FP8 + turbo3 combined | end-to-end video generated | ✅ |