Implementation:Eric mitchell Direct preference optimization Disable Dropout

Knowledge Sources	Direct Preference Optimization
Domains	Regularization, Training_Stability
Last Updated	2026-02-08 02:00 GMT

Overview

Concrete tool for disabling all dropout layers in a PyTorch model provided by the direct-preference-optimization repository.

Description

The disable_dropout function iterates over all modules in a PyTorch model and sets the dropout probability to 0 for any nn.Dropout instance. This ensures deterministic forward passes during DPO and SFT training.

Usage

Call this function immediately after loading a model with AutoModelForCausalLM.from_pretrained and before passing the model to any trainer. Applied to both policy and reference models.

Code Reference

Source Location

• Repository: direct-preference-optimization
• File: utils.py
• Lines: 99-103

Signature

def disable_dropout(model: torch.nn.Module):
    """Disable dropout in a model."""
    for module in model.modules():
        if isinstance(module, torch.nn.Dropout):
            module.p = 0

Import

from utils import disable_dropout

I/O Contract

Inputs

Name	Type	Required	Description
model	torch.nn.Module	Yes	Any PyTorch model; all nn.Dropout submodules will have p set to 0

Outputs

Name	Type	Description
(in-place)	None	Modifies model in-place; all nn.Dropout.p values set to 0

Usage Examples

Disabling Dropout After Model Loading

import transformers
from utils import disable_dropout

model = transformers.AutoModelForCausalLM.from_pretrained("gpt2-large")
disable_dropout(model)

# Verify dropout is disabled
for module in model.modules():
    if isinstance(module, torch.nn.Dropout):
        assert module.p == 0

Related Pages

Implements Principle

• Principle:Eric_mitchell_Direct_preference_optimization_Dropout_Disabling

Requires Environment

• Environment:Eric_mitchell_Direct_preference_optimization_PyTorch_CUDA