T5Base_fp8 / README.md
codemichaeld's picture
Upload README.md with huggingface_hub
ce144ee verified
|
raw
history blame
1.54 kB
---
library_name: diffusers
tags:
- fp8
- safetensors
- quantization
- precision-recovery
- diffusion
- converted-by-gradio
---
# FP8 Model with Precision Recovery
- **Source**: `https://huggingface.co/LifuWang/DistillT5`
- **File**: `model.safetensors`
- **FP8 Format**: `E5M2`
- **Correction Mode**: per_tensor
- **Correction File**: `model-correction.safetensors`
- **FP8 File**: `model-fp8-e5m2.safetensors`
## Usage (Inference)
```python
from safetensors.torch import load_file
import torch
# Load FP8 model and correction factors
fp8_state = load_file("model-fp8-e5m2.safetensors")
correction_state = load_file("model-correction.safetensors") if os.path.exists("model-correction.safetensors") else {}
# Reconstruct high-precision weights
reconstructed = {}
for key in fp8_state:
fp8_weight = fp8_state[key].to(torch.float32)
# Apply correction if available
correction_key = f"correction.{key}"
if correction_key in correction_state:
correction = correction_state[correction_key].to(torch.float32)
reconstructed[key] = fp8_weight + correction
else:
reconstructed[key] = fp8_weight
# Use reconstructed weights in your model
model.load_state_dict(reconstructed)
```
## Correction Modes
- **Per-Channel**: Computes mean correction per output channel (best for most layers)
- **Per-Tensor**: Single correction value per tensor (lightweight)
- **None**: No correction (pure FP8)
> Requires PyTorch ≥ 2.1 for FP8 support. For best quality, use the correction file during inference.