13 Pull requests merged by 8 people

• fix(inference): Add missing dtype attribute to ParameterBase setter

  #7378 merged Jun 23, 2025

• Add support for ws=1 scenario

  #7379 merged Jun 23, 2025

• Fix dtype mismatch in TestParamPartitioningSkipInit

  #7377 merged Jun 23, 2025

• fix wandb.log() call by removing sync kwarg

  #7383 merged Jun 23, 2025

• Fix release of IPG buffer

  #7376 merged Jun 22, 2025

• Update latest news with DeepNVMe

  #7375 merged Jun 20, 2025

• Relax tolerances for FP8 unit test only for ROCm + FP16

  #7373 merged Jun 20, 2025

• Flops profiler support for F.interpolate

  #7353 merged Jun 20, 2025

• add Arctic Long Sequence Training paper reference

  #7372 merged Jun 20, 2025

• Enable torch.autocast with ZeRO

  #6993 merged Jun 19, 2025

• sequence parallel default dtype

  #7364 merged Jun 19, 2025

• Fix(scheduler): WarmupLR inherits optimizer lr when not specified

  #7360 merged Jun 19, 2025

• Restore real inputs for recompilation

  #7356 merged Jun 19, 2025

1 Pull request opened by 1 person

• fix #7188

  #7371 opened Jun 19, 2025

9 Issues closed by 5 people

• [REQUEST] Support for XLA/TPU

  #6901 closed Jun 24, 2025

• [BUG] AttributeError: 'UnembedParameter' object has no attribute 'dtype'

  #7260 closed Jun 23, 2025

• [BUG] WandbMonitor log() invocation broken with wandb 0.20.0

  #7381 closed Jun 23, 2025

• [BUG]Training

  #7319 closed Jun 20, 2025

• nv-sd CI test failure

  #7310 closed Jun 20, 2025

• [BUG] FLOPS compute **FAILS** for `F.interpolate` when using `scale_factor`

  #4504 closed Jun 20, 2025

• Bug when using optimizer and WarmupLR togather

  #7303 closed Jun 19, 2025

• [BUG] Universal checkpoint conversion - "Cannot find layer_01* files in there"

  #5776 closed Jun 17, 2025

• [BUG] No `universal_checkpoint_info` in the Accelerate+Deepspeed Checkpoint

  #5430 closed Jun 17, 2025

5 Issues opened by 5 people

• [BUG] Ulysses DistributedAttention silently produces incorrect output when #GPUs does not divide global sequence length

  #7384 opened Jun 23, 2025

• [BUG] deepspeed v0.17.1 con't run well on NPU platform!

  #7380 opened Jun 23, 2025

• [BUG] Memory leak when using adam_offload and save_checkpoint

  #7370 opened Jun 19, 2025

• [BUG] init_inference loads qwen3-32b model very slow but train model loads it quickly

  #7369 opened Jun 18, 2025

• FastPersist micro-benchmarks test results are inconsistent with expectations

  #7368 opened Jun 18, 2025

12 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.

• HF2UCP: Converting a `pytorch_model.bin` or `.safetensors` checkpoint to UCP

  #7212 commented on Jun 23, 2025 • 4 new comments

• [BUG] Qwen3: model loading failed when using meta device

  #7275 commented on Jun 18, 2025 • 0 new comments

• Functorch support: RuntimeError: In order to use an autograd.Function with functorch transforms

  #7323 commented on Jun 19, 2025 • 0 new comments

• AssertionError: no sync context manager is incompatible with gradientpartitioning logic of ZeRo stage 3

  #6793 commented on Jun 20, 2025 • 0 new comments

• Error when installing deepspeed with pip (Not sure if this is a bug or not)

  #7358 commented on Jun 23, 2025 • 0 new comments

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

  #7229 commented on Jun 23, 2025 • 0 new comments

• nv-torch-nightly-v100 CI test failure

  #7195 commented on Jun 24, 2025 • 0 new comments

• nv-nightly CI test failure

  #7140 commented on Jun 24, 2025 • 0 new comments

• [BUG] Receiving CUDA error: invalid argument using pytorch 2.7 with deepspeed 0.16.4 with Cuda 12.8

  #7150 commented on Jun 24, 2025 • 0 new comments

• [BUG] Universal Checkpoint Conversion: Resumed Training Behaves as If Model Initialized from Scratch

  #6691 commented on Jun 24, 2025 • 0 new comments

• Avoid graph break by enabling compile of record module

  #7362 commented on Jun 23, 2025 • 0 new comments

• Fix ZeRO stage 1 and add stage 2 support with DeepCompile

  #7366 commented on Jun 23, 2025 • 0 new comments