8 Pull requests merged by 5 people

• Update references to new X/Twitter handle

  #7110 merged Mar 5, 2025

• Fix fused_qkv print model ValueError

  #7109 merged Mar 4, 2025

• Avoid graph break due to unsupported frozenset

  #7105 merged Mar 4, 2025

• Only run pre-commit on the changes

  #7106 merged Mar 4, 2025

• Avoid graph breaks in torch.compile caused by inner classes in the backward hooks

  #7062 merged Mar 4, 2025

• Avoid graph breaks by disabling sourceless calls in instrument_w_nvtx

  #7081 merged Mar 3, 2025

• Use new dlpack api; Formatting fixes

  #7101 merged Mar 3, 2025

• Remove workflows for very old torch versions

  #7090 merged Feb 28, 2025

5 Pull requests opened by 5 people

• Update gaudi2 nightly,ci to latest 1.20.0 build

  #7093 opened Feb 28, 2025

• Variable batch size and LR scheduler

  #7104 opened Mar 3, 2025

• [Draft] Add support for seq split in Domino

  #7111 opened Mar 4, 2025

• fix keep_module_on_host

  #7112 opened Mar 6, 2025

• [XPU] Support XCCL on deepspeed side

  #7113 opened Mar 6, 2025

3 Issues closed by 3 people

• nv-sd CI test failure

  #7098 closed Mar 3, 2025

• [BUG]npu zero3 训练自定义模型时，报错Function SumBackward0 returned an invalid gradient at index 0

  #7078 closed Mar 3, 2025

• [BUG]Does MoQ mode deprecated in DeepSpeed? I run with MoQ config, but no quantization found in log

  #7091 closed Feb 28, 2025

8 Issues opened by 8 people

• Deepspeed on Power with CPU Acclerator and AuToTP

  #7108 opened Mar 4, 2025

• [REQUEST] An option for SUM gradient allreduce instead of MEAN

  #7107 opened Mar 4, 2025

• Support DualPipe training

  #7100 opened Mar 2, 2025

• About delay param update of ZeRO Offload

  #7099 opened Mar 2, 2025

• [REQUEST] Proposal for Enhancing ChatGPT's Response Quality During Training

  #7097 opened Mar 1, 2025

• LLama factory, quantized model and deepspeed compatibility

  #7096 opened Mar 1, 2025

• nv-nightly CI test failure

  #7095 opened Mar 1, 2025

• [BUG]Does MoQ mode deprecated in DeepSpeed? I run with MoQ config, but no quantization found in log

  #7092 opened Feb 28, 2025

22 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.

• Training multiple models

  #7018 commented on Mar 6, 2025 • 10 new comments

• Enabled high-performance Automatic Tensor Parallelism (auto TP) for the MoE models on multiple GPUs/HPUs

  #6964 commented on Mar 5, 2025 • 2 new comments

• Unpin once transformers latest is fixed

  #7088 commented on Mar 3, 2025 • 0 new comments

• Update Domino for Llama3

  #7084 commented on Mar 5, 2025 • 0 new comments

• Conditionally quote env vars

  #7071 commented on Mar 5, 2025 • 0 new comments

• Fix, pipeline model with moe cause error when send grad

  #7055 commented on Mar 5, 2025 • 0 new comments

• Enable ZeRO set/get APIs for NVMe offload

  #7046 commented on Mar 5, 2025 • 0 new comments

• Enable python 3.11 and 3.12 tests

  #7007 commented on Mar 5, 2025 • 0 new comments

• Enable torch.autocast with ZeRO

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

• Improve overflow handling in ZeRO

  #6976 commented on Mar 4, 2025 • 0 new comments

• Enabled configurable auto Tensor Parallelism (TP) for the inference of diverse models

  #6553 commented on Feb 28, 2025 • 0 new comments

• support autoTP with weight only quantization in DS inference path

  #4750 commented on Mar 5, 2025 • 0 new comments

• Getting requirements to build wheel: finished with status 'error'

  #7043 commented on Mar 5, 2025 • 0 new comments

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

  #6793 commented on Mar 4, 2025 • 0 new comments

• Deepspeed Inference not working on llama when input has padding and using kernel injection

  #3960 commented on Mar 4, 2025 • 0 new comments

• [REQUEST] Runable solution of RTX 5090 GPU + Linux Driver version + Pytorch version + Deepspeed version for LLM finetuning?

  #7042 commented on Mar 3, 2025 • 0 new comments

• [BUG] Deepspeed does not update the model when using "Qwen/Qwen2.5-3B" and is fine with ""Qwen/Qwen2.5-1.%B""

  #7077 commented on Mar 3, 2025 • 0 new comments

• Dynamic/variable batch size support

  #1051 commented on Mar 3, 2025 • 0 new comments

• [BUG] DS zero stage 1 or 2 communication uses reduce-scatter instead of All-reduce

  #7059 commented on Mar 1, 2025 • 0 new comments

• Suspected memory leak during zero3 training. oom eventually after several checkpoint

  #3582 commented on Mar 1, 2025 • 0 new comments

• Ascend 910B: attributeerror: 'deepspeedcpuadam' object has no attribute 'ds_opt_adam'

  #7061 commented on Feb 28, 2025 • 0 new comments

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

  #7044 commented on Feb 28, 2025 • 0 new comments