Expertise
PyTorch developer — from training to serving, since 2024
Since 2024 I have trained models in PyTorch and adapted them with fine-tuning and LoRA/PEFT; the road that ends with the trained model behind an API is part of my work too.
- years
- 2
Me in the PyTorch ecosystem
PyTorch is where I touch the model itself: training, fine-tuning and adaptation with LoRA/PEFT. The output of training does not end there: the road to putting the model behind an API is part of my AI systems engineering work too.
Two tools sit around the framework: Python as my main language and CUDA on the GPU side. TensorFlow is part of my model-training work too; I see the two as tools I choose by the job, not as opposites.
This page describes the work I have done since 2024, not anything older.
What is the PyTorch Developer role?
PyTorch is an open-source deep learning framework used with Python: it provides the building blocks for tensor operations, automatic differentiation and the training loop. A PyTorch developer is the person who defines, trains and adapts a model with those building blocks.
What does the PyTorch Developer do?
I use PyTorch for model training, fine-tuning and adaptation with LoRA/PEFT. I run the training workload on the GPU with CUDA and move the trained model to a serving layer to put it behind an API.
Technologies I pair it with
-
PyTorch
The deep learning framework I use for model training, fine-tuning and adaptation with LoRA/PEFT.
-
Python
The main language of my model-training, fine-tuning and data-preparation work.
-
CUDA
The NVIDIA platform I run training and inference workloads on the GPU with.
-
OpenAI Triton
The language and compiler I use to write custom GPU kernels for inference optimization.
-
NVIDIA Dynamo-Triton
The model serving server, formerly NVIDIA Triton Inference Server, that I use to put trained models behind an API.
-
TensorFlow
The deep learning framework I use in my model-training work.
Frequently asked
7 questions
-
How long have you worked with PyTorch?
Since 2024, which is 2 years.
-
What does a PyTorch developer do?
They define, train and adapt a model in PyTorch: prepare the data, build the training loop, evaluate the result. In my work that means model training, fine-tuning and adaptation with LoRA/PEFT.
-
What is the difference between PyTorch and TensorFlow?
Both are deep learning frameworks. I use both in my model-training work; this page covers PyTorch and the [TensorFlow page](/expertise/tensorflow/) covers the other. I make no claim about which one is better; I think the work itself decides the choice.
-
What are LoRA and PEFT, and why do you use them?
PEFT (parameter-efficient fine-tuning) is the general name for methods that train a small set of parameters instead of updating every weight of a model; LoRA is one of them. I take this route when I adapt an existing model to my own work.
-
Do you run PyTorch on the GPU?
Yes. I run training and inference workloads on the GPU with NVIDIA's CUDA platform; the details are on the [CUDA page](/expertise/cuda/).
-
How do you serve a model you trained in PyTorch?
I use NVIDIA Dynamo-Triton, formerly NVIDIA Triton Inference Server, to put trained models behind an API. When inference optimization needs custom GPU kernels, I write them with OpenAI Triton.
-
Which language do you use PyTorch with?
Python. It is the main language of my model-training, fine-tuning and data-preparation work; the [Python page](/expertise/python/) covers the language itself.
Let us work together
If you have work in this ecosystem, tell me what you are building and we will talk about how to build it.