v0.10.4

Changes:

• try again on GPU release error, hopefully fixes "Card N is still in use" message in most cases.

Full Changelog: v0.10.3...v0.10.4

v0.10.3

Changes:

• Relax depencies for the debian package to be compatible with older NVIDIA drivers

Full Changelog: v0.10.2...v0.10.3

v0.10.2

Changes:

• Rename debian package to gpu-claim

Full Changelog: v0.10.1...v0.10.2

v0.10.1

Changes:

• Correctly pass through environment variables like LD_LIBRARY_PATH or LD_PRELOAD (see commit e4a1f93 for details).

v0.10.0

Changes:

• Support the --card option, which lets you run a job in parallel on the GPUs which you already own.

v0.9.0

Changes:

• Major code refactoring
• Nicer handling of client <-> server connection losses

v0.8.0

Changes:

• Make sure GPUs are freed directly if the gpu process is killed (e.g. by VSCode)
• Remove the claim command, since nobody uses it anyway and it interferes with the above
• Set CUDA_VISIBLE_DEVICES with numeric IDs, anything else confuses older pytorch versions (torch.cuda.device_count() returns 0)
• Fix NVML struct versioning
• Print idle time in seconds rather than in minutes
• Decrease idle timeout to 1 min

v0.7.1

Changes:

• Use versioned NVML API - fixes ABI breakage on newest NVIDIA drivers (535.86.10)

v0.7.0

Changes:

• Isolate /dev/shm for each GPU container. This ensures correct cleanup (torch.multiprocessing, I'm looking at you!).
• Ensure that the initramfs is rebuilt after installation of this package, so that the correct nvidia flags are set.

v0.6.0

Changes:

• Force exit of all descendant user processes. Sometimes people use complicated python multiprocessing pipelines which leave stuck
  child processes behind. Using Linux PID namespaces, we can make sure that all processes are killed as soon as the main process exits.