Skip to content

Releases: pytorch/torchx

v0.7.0

16 Jul 21:06
299ad04
Compare
Choose a tag to compare

v0.6.0

17 Oct 16:33
6153315
Compare
Choose a tag to compare

v0.5.0

03 Apr 23:17
f722001
Compare
Choose a tag to compare

v0.4.0

30 Dec 01:45
6fb2a3f
Compare
Choose a tag to compare

v0.3.0

27 Oct 23:23
Compare
Choose a tag to compare

v0.2.0

15 Jun 23:05
Compare
Choose a tag to compare

v0.1.2

29 Mar 01:54
Compare
Choose a tag to compare

Full Changelog: v0.1.1...v0.1.2

Milestone: https://github.com/pytorch/torchx/milestones/3

  • PyTorch 1.11 Support
  • Python 3.10 Support
  • torchx.workspace
    • TorchX now supports a concept of workspaces. This enables seamless launching
      of jobs using changes present in your local workspace. For Docker based
      schedulers, we automatically build a new docker container on job launch
      making it easier than ever to run experiments. #333
  • torchx.schedulers
    • Ray #329
    • AWS Batch #381
      • Newly added AWS Batch scheduler makes it easy to launch jobs in AWS with minimal infrastructure setup.
    • Slurm
      • Slurm jobs will by default launch in the current working directory to match local_cwd and workspace behavior. #372
      • Replicas now have their own log files and can be accessed programmatically. #373
      • Support for comment, mail-user and constraint fields. #391
      • Workspace support (prototype) - Slurm jobs can now be launched in isolated experiment directories. #416
    • Kubernetes
      • Support for running jobs under service accounts. #408
      • Support for specifying instance types. #433
    • All Docker-based Schedulers (Kubernetes, Batch, Docker)
      • Added bind mount and volume supports #420, #426
      • Bug fix: Better shm support for large dataloader #429
      • Support for .dockerignore and custom Dockerfiles #401
    • Local Scheduler
      • Automatically set CUDA_VISIBLE_DEVICES #383
      • Improved log ordering #366
  • torchx.components
    • dist.ddp
      • Rendezvous works out of the box on all schedulers #400
      • Logs are now prefixed with local ranks #412
      • Can specify resources via the CLI #395
      • Can specify environment variables via the CLI #399
    • HPO
  • torchx.cli
    • .torchxconfig
      • You can now specify component argument defaults .torchxconfig c37cfd7
      • ~/.torchxconfig can now be used to set user level defaults. #378
      • --workspace can be configured #397
    • Color change and bug fixes #419
  • torchx.runner
    • Now supports workspace interfaces. #360
    • Returned lines now preserve whitespace to provide support for progress bars #425
    • Events are now logged to torch.monitor when available. #379
  • torchx.notebook (prototype)
    • Added new workspace interface for developing models and launching jobs via a Jupyter Notebook. #356
  • Docs
    • Improvements to clarify TorchX usage w/ workspaces and general cleanups.
    • #374, #402, #404, #407, #434

v0.1.1

18 Nov 02:20
Compare
Choose a tag to compare

v0.1.0

21 Oct 17:59
Compare
Choose a tag to compare

v0.1.0rc1

18 Oct 18:40
Compare
Choose a tag to compare