README.md

---
library_name: diffusers
tags:
- fp8
- safetensors
- precision-recovery
- mixed-method
- converted-by-gradio
---
# FP8 Model with Per-Tensor Precision Recovery
- **Source**: `https://huggingface.co/spacepxl/Wan2.1-VAE-upscale2x`
- **Original File**: `Wan2.1_VAE_upscale2x_imageonly_real_v1.safetensors`
- **FP8 Format**: `E5M2`
- **FP8 File**: `Wan2.1_VAE_upscale2x_imageonly_real_v1-fp8-e5m2.safetensors`
- **Recovery File**: `Wan2.1_VAE_upscale2x_imageonly_real_v1-recovery.safetensors`

## Recovery Rules Used
```json
[
  {
    "key_pattern": "vae",
    "dim": 4,
    "method": "diff"
  },
  {
    "key_pattern": "encoder",
    "dim": 4,
    "method": "diff"
  },
  {
    "key_pattern": "decoder",
    "dim": 4,
    "method": "diff"
  },
  {
    "key_pattern": "all",
    "method": "none"
  }
]
```

## Usage (Inference)
```python
from safetensors.torch import load_file
import torch

# Load FP8 model
fp8_state = load_file("Wan2.1_VAE_upscale2x_imageonly_real_v1-fp8-e5m2.safetensors")

# Load recovery weights if available
recovery_state = load_file("Wan2.1_VAE_upscale2x_imageonly_real_v1-recovery.safetensors") if "Wan2.1_VAE_upscale2x_imageonly_real_v1-recovery.safetensors" and os.path.exists("Wan2.1_VAE_upscale2x_imageonly_real_v1-recovery.safetensors") else {}

# Reconstruct high-precision weights
reconstructed = {}
for key in fp8_state:
    fp8_weight = fp8_state[key].to(torch.float32)  # Convert to float32 for computation
    
    # Apply LoRA recovery if available
    lora_a_key = f"lora_A.{key}"
    lora_b_key = f"lora_B.{key}"
    if lora_a_key in recovery_state and lora_b_key in recovery_state:
        A = recovery_state[lora_a_key].to(torch.float32)
        B = recovery_state[lora_b_key].to(torch.float32)
        # Reconstruct the low-rank approximation
        lora_weight = B @ A
        fp8_weight = fp8_weight + lora_weight
    
    # Apply difference recovery if available
    diff_key = f"diff.{key}"
    if diff_key in recovery_state:
        diff = recovery_state[diff_key].to(torch.float32)
        fp8_weight = fp8_weight + diff
    
    reconstructed[key] = fp8_weight

# Use reconstructed weights in your model
model.load_state_dict(reconstructed)
```

> **Note**: For best results, use the same recovery configuration during inference as was used during extraction.
> Requires PyTorch ≥ 2.1 for FP8 support.

## Statistics
- **Total layers**: 194
- **Layers with recovery**: 60
  - LoRA recovery: 0
  - Difference recovery: 60