Currently each compute environment in DOCKER_COMPUTE_ENVIRONMENTS owns its full hardware resource list, including hardware metadata (init, deviceRequests, driverVersion, platform, etc.). This forces GPU (or any special hardware) to be duplicated across every environment that needs it, with no actual sharing — each env defines and tracks its own resources independently.
The goal is to move hardware resource definitions up to the Docker-connection level, so multiple environments can reference the same pool entry. This enables true sharing (e.g., one NVIDIA A100 with 2 GPUs split between a premium env and a standard env, with runtime contention handled automatically).
Currently each compute environment in
DOCKER_COMPUTE_ENVIRONMENTSowns its full hardware resource list, including hardware metadata (init,deviceRequests,driverVersion,platform, etc.). This forces GPU (or any special hardware) to be duplicated across every environment that needs it, with no actual sharing — each env defines and tracks its own resources independently.The goal is to move hardware resource definitions up to the Docker-connection level, so multiple environments can reference the same pool entry. This enables true sharing (e.g., one NVIDIA A100 with 2 GPUs split between a premium env and a standard env, with runtime contention handled automatically).