Sub-issue of #1364.
Depends on: #1365 (configurable binder plugins).
Implement a new binder plugin that injects a GPU_MEMORY_LIMIT environment variable into each container requesting a shared GPU. The value (in bytes) depends on the request type:
- GPU memory request — value is the explicitly requested memory amount
- GPU fraction request — value is computed after the scheduling decision, based on the fraction and the actual GPU's total memory on the assigned node
This enables external GPU memory isolation runtimes (e.g., HAMI Core) to enforce hard memory limits.
Sub-issue of #1364.
Depends on: #1365 (configurable binder plugins).
Implement a new binder plugin that injects a
GPU_MEMORY_LIMITenvironment variable into each container requesting a shared GPU. The value (in bytes) depends on the request type:This enables external GPU memory isolation runtimes (e.g., HAMI Core) to enforce hard memory limits.