Reference environment
Field Note: Split Compute and GPU Runtimes
A sanitized reference architecture separating Pulsar application services from dedicated GPU model and media runtimes.
Key considerations
Application isolation
The compute node owns workflow and application state.
GPU specialization
Model and media hosts are sized and maintained separately.
Independent recovery
Application and runtime services can be diagnosed without rebuilding the full stack.
Architecture at a glance
Constraint
The reference environment needed private application data and operator controls without forcing the web stack, chat inference, and media workflows onto one GPU host.
Decision
Pulsar application services, Postgres, Valkey, workers, and object storage remained together on the compute node. Chat inference and ComfyUI used approved network paths to separately operated runtime hosts.
What improved
Application deployment and rollback no longer required model-runtime replacement. GPU services could be sized and maintained according to workload, while environment-specific application data remained isolated.
What operators still verify
Network reachability, TLS and proxy policy, runner model availability, queue age, object storage, backups, and release compatibility remain explicit promotion checks.
Field-note status
This is a sanitized Pulsar engineering reference environment. It is not a customer case study and includes no external endorsement or performance claim.
Private deployment consultation
Review Pulsar against your environment.
Bring the infrastructure, security boundaries, model runners, and use cases. The Pulsar team will map the appropriate deployment path.