Compute Infrastructures for IA applications in the wild - second edition

With the advent of Chatbots, LLMs and other generative IA technologies, as well as other progresses in the IA field, there is an explosion of the demand for compute force. IA is no longer computer science: it is computational science. As such, it can no longer be done with casual, self-managed equipment. More advanced compute infrastructures are required both to satisfy user needs (in terms of compute power, GPU Ram capacity) and to ensure a decent utilization of the increasingly costly resources.

Content and topics

The purpose of this workshop is to gather people in charge of compute infrastructure (on-prem, cloud or hybrid) destined to support AI workloads (both training and inference). Being “in charge” means someone who does typically performs one or more of the tasks below

• Prepares and follows purchase orders for servers including GPUs.

• Racks the servers, ensure they are connected with adequate network / IO performance

• Install the OS, the drivers, some software stack

• Binds the servers with the login system, typically LDAP

• Monitor the utilization of the server over time

• Ensure there are no abuse

• Insert the server into a cluster (e.g. SLURM, K8s)

• Designs and implement cluster architecture

• Helps AI engineers to run their workloads according to the policies

• Provisions cloud GPUs resources

• Manages cloud platforms accesses

• Secures the cluster

• Dimensions the cluster and plan its evolution over the time

The workshop will be organized around 8-10 presentations, followed by a group discussion

The workshop has no specific registration, and walk-ins are welcome.

Schedule and speakers: to be announced by mid-February