Implementation:Triton inference server Server Model Repository Directory Convention

Knowledge Sources	Triton Server Model Repository Docs
Domains	MLOps, Model_Serving
Last Updated	2026-02-13 17:00 GMT

Overview

Concrete filesystem convention for organizing model artifacts into Triton Inference Server's required directory hierarchy.

Description

The Model Repository Directory Convention defines the exact filesystem layout that Triton Inference Server expects when loading models. Each model occupies a named directory containing numbered version subdirectories and an optional config.pbtxt configuration file. The server scans this directory structure at startup (or on poll) to discover and load models.

Usage

Use this convention whenever preparing models for deployment on Triton Inference Server. This is the first step in any Triton deployment workflow — before launching the server, the model repository must be correctly structured.

Code Reference

Source Location

• Repository: triton-inference-server/server
• File: docs/user_guide/model_repository.md
• Lines: L36-72 (Repository Layout), L267-378 (Model file naming per backend)

Signature

# Create model repository with required hierarchy
mkdir -p <model-repository-path>/<model-name>/<version>/

# Example: Deploy an ONNX model
mkdir -p models/densenet_onnx/1/
cp model.onnx models/densenet_onnx/1/model.onnx

# Optional: Add configuration
cat > models/densenet_onnx/config.pbtxt << 'EOF'
name: "densenet_onnx"
platform: "onnxruntime_onnx"
max_batch_size: 0
input [
  { name: "data_0", data_type: TYPE_FP32, dims: [ 3, 224, 224 ] }
]
output [
  { name: "fc6_1", data_type: TYPE_FP32, dims: [ 1000 ] }
]
EOF

Import

# No import needed — this is a filesystem convention
# Consumed by: tritonserver --model-repository=<path>

I/O Contract

Inputs

Name	Type	Required	Description
model-repository-path	string (filesystem path)	Yes	Root directory for all models
model-name	string (directory name)	Yes	Name of the model (becomes the model identifier)
version	integer (directory name)	Yes	Numeric version directory (e.g., 1, 2, 3)
model-definition-file	file	Yes	Backend-specific model file (model.onnx, model.plan, model.pt, model.py)
config.pbtxt	protobuf text file	No	Model configuration (optional for auto-complete backends)

Outputs

Name	Type	Description
model repository	directory tree	Well-formed directory hierarchy readable by tritonserver
model identity	string	Model name derived from directory name
model versions	integer set	Available versions derived from numeric subdirectories

Usage Examples

Minimal ONNX Model Repository

# 1. Create the repository structure
mkdir -p model_repository/simple_onnx/1/

# 2. Copy the model file (backend auto-detects from file extension)
cp my_model.onnx model_repository/simple_onnx/1/model.onnx

# 3. Launch server (config.pbtxt auto-generated for ONNX)
tritonserver --model-repository=model_repository

Multi-Model Repository with Multiple Versions

# Create repository with two models, each with two versions
mkdir -p model_repository/text_classifier/1/
mkdir -p model_repository/text_classifier/2/
mkdir -p model_repository/image_detector/1/

cp classifier_v1.onnx model_repository/text_classifier/1/model.onnx
cp classifier_v2.onnx model_repository/text_classifier/2/model.onnx
cp detector.plan model_repository/image_detector/1/model.plan

# Add explicit configuration for TensorRT model
cat > model_repository/image_detector/config.pbtxt << 'EOF'
name: "image_detector"
platform: "tensorrt_plan"
max_batch_size: 8
input [
  { name: "input", data_type: TYPE_FP32, dims: [ 3, 640, 640 ] }
]
output [
  { name: "detections", data_type: TYPE_FP32, dims: [ 100, 6 ] }
]
EOF

Cloud Storage Repository (S3)

# Triton supports S3, GCS, and Azure Blob Storage
# Same directory structure, different path prefix
tritonserver --model-repository=s3://my-bucket/models

# Structure in S3:
# s3://my-bucket/models/
#     model_a/
#         config.pbtxt
#         1/
#             model.onnx

Related Pages

Implements Principle

• Principle:Triton_inference_server_Server_Model_Repository_Layout