1 Release published by 1 person

• v0.17.1 v0.17.1 Patch Release

  published Jun 9, 2025

15 Pull requests merged by 12 people

• Don't break set_start_method

  #7349 merged Jun 11, 2025

• s/UlyssesPlus/Arctic Long Sequence Training (ALST)/

  #7348 merged Jun 11, 2025

• Update version after 0.17.1 release

  #7345 merged Jun 10, 2025

• Move pytest pinning from individual tests to requirements-dev.txt until fixed.

  #7327 merged Jun 9, 2025

• Fix docs that are rendering Incorrectly

  #7344 merged Jun 9, 2025

• Improve overflow handling in ZeRO

  #6976 merged Jun 9, 2025

• Update folder name

  #7343 merged Jun 9, 2025

• Fix LoRA arxiv reference

  #7340 merged Jun 7, 2025

• fixed: Modified the topkgating function and modified the test_moe file for testing

  #7163 merged Jun 6, 2025

• DeepNVMe update

  #7215 merged Jun 6, 2025

• fp16 optimizer timers fix - TypeError: 'NoneType' object is not callable

  #7330 merged Jun 6, 2025

• Fix issue with symint input

  #7243 merged Jun 6, 2025

• Fix pytest version to 8.3.5 in hpu-gaudi actions

  #7337 merged Jun 5, 2025

• Update config_utils.py

  #7333 merged Jun 5, 2025

• Improve Ulysses Plus Docs

  #7335 merged Jun 5, 2025

9 Issues closed by 8 people

• [BUG] Deepspeed may set multiprocessing set_start_method, breaking existing applications

  #7347 closed Jun 11, 2025

• [BUG] AutoTP: incorrect total train batch size when using the huggingface trainer API

  #7298 closed Jun 10, 2025

• Model Checkpoint docs are incorrectly rendered on deepspeed.readthedocs.io

  #6747 closed Jun 9, 2025

• [BUG] Zero2 offload overflow

  #5241 closed Jun 9, 2025

• [REQUEST]Do deepspeed's Pipline Engine support `ep >1 and pp >1 and tp>1` theoretically now?

  #7336 closed Jun 7, 2025

• Trainer.train(resume_from_checkpoint=...) fails when using auto tensor parallel

  #7320 closed Jun 7, 2025

• [REQUEST] Equivalent of FSDP ignore_params or ignore_modules for DeepSpeed Zero 3

  #7271 closed Jun 6, 2025

• [BUG]AssertionError: found no DeviceMesh from dtensor args for c10d.broadcast_.default!

  #7338 closed Jun 6, 2025

• [BUG] DeepSpeed accuracy issue for torch.compile if activation checkpoint function not compiler disabled

  #6718 closed Jun 5, 2025

2 Issues opened by 2 people

• DeepSpeed async_io requires libaio-0.3.112 or newer, breaks on libaio-0.3.111 (e.g., Fedora/EL9)

  #7346 opened Jun 10, 2025

• [REQUEST]Do deepspeed support megatron's `--sequence_parallel`？

  #7339 opened Jun 6, 2025

20 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• Create COMMITTERS_RESPONSIBILITY.md

  #7300 commented on Jun 10, 2025 • 1 new comment

• Enable torch.autocast with ZeRO

  #6993 commented on Jun 6, 2025 • 0 new comments

• Update names of CPU Adam/Adagrad/Lion params to better match torch/GPU ops.

  #5382 commented on Jun 10, 2025 • 0 new comments

• [BUG] DeepCompile in ZeRO-1 fails to do the forward pass

  #7229 commented on Jun 11, 2025 • 0 new comments

• [BUG] DeepCompile in ZeRO-3 fails to do the forward pass

  #7228 commented on Jun 11, 2025 • 0 new comments

• Test hangs when world_size=4 and reuse_dist_env=True on PyTorch >= 2.7.0

  #7334 commented on Jun 11, 2025 • 0 new comments

• [BUG] DeepCompile: MemoryProfiling error /pytorch/build/aten/src/ATen/RegisterCUDA.cpp:7280: SymIntArrayRef expected to contain only concrete integers

  #7311 commented on Jun 11, 2025 • 0 new comments

• nv-nightly CI test failure

  #7140 commented on Jun 11, 2025 • 0 new comments

• nv-torch-nightly-v100 CI test failure

  #7195 commented on Jun 11, 2025 • 0 new comments

• [BUG] The NCCL timed out while using the zero3 model. How can I solve this problem?

  #5066 commented on Jun 10, 2025 • 0 new comments

• [BUG] deepspeed zero2 training hangon and timeout after a fixed step

  #7044 commented on Jun 10, 2025 • 0 new comments

• [BUG] how to achieve hybrid data and pipeline parallelism?

  #7280 commented on Jun 10, 2025 • 0 new comments

• [REQUEST]I want to use CPU-based distributed approach to train a small recommendation model. Is there a demo available for me to refer ?

  #7329 commented on Jun 10, 2025 • 0 new comments

• [REQUEST] Support for XLA/TPU

  #6901 commented on Jun 6, 2025 • 0 new comments

• [BUG]"DeepSpeedZeRoOffload missing '_restore_from_bit16_weights' method when loading checkpoints"

  #7272 commented on Jun 6, 2025 • 0 new comments

• [BUG] `Assert Error: assert buffer.grad is not None` & `RuntimeError: element 1 of tensors does not require grad and does not have a grad_fn` During pipeline parallelism

  #7270 commented on Jun 6, 2025 • 0 new comments

• [BUG]Training

  #7319 commented on Jun 6, 2025 • 0 new comments

• [BUG] AttributeError: 'Linear' object has no attribute 'ds_grads_remaining'

  #7203 commented on Jun 6, 2025 • 0 new comments

• [BUG]clip gradient not working

  #7220 commented on Jun 5, 2025 • 0 new comments

• [REQUEST] How can I use AutoTP for training without huggingface Trainer and accelerate ?

  #7179 commented on Jun 5, 2025 • 0 new comments