Search results

[BUG][IMPORTANT] zero_to_fp32.py consolidated weights are all zero after this commit

bugSomething isn't workingSomething isn't working

compression

Status: Open.

#6791 In microsoft/DeepSpeed;

· npuichigo opened on Nov 26, 2024

nv-nightly CI test failure

ci-failure

Status: Open.

#6790 In microsoft/DeepSpeed;

· github-actions opened on Nov 26, 2024

[BUG] RuntimeError: Input type (torch.cuda.FloatTensor) and weight type (CUDABFloat16Type) should be the same

bugSomething isn't workingSomething isn't working

training

Status: Open.

#6789 In microsoft/DeepSpeed;

· rileyhun opened on Nov 26, 2024

Question about throughput numbers in blog post

Status: Open.

#6781 In microsoft/DeepSpeed;

· rlanday opened on Nov 25, 2024

[REQUEST] Let ZeRO-offload use CPU and GPU parallelly

enhancementNew feature or requestNew feature or request

Status: Open.

#6778 In microsoft/DeepSpeed;

· fzyzcjy opened on Nov 23, 2024

[BUG] [Fix Suggestion] Uneven head sequence parallelism

bugSomething isn't workingSomething isn't working

training

Status: Open.

#6774 In microsoft/DeepSpeed;

· Eugene29 opened on Nov 21, 2024

[BUG] [Fix-Suggested] ZeRO Stage 3 Overwrites Module ID Attribute Causing Incorrect Expert Placement on GPUs

bugSomething isn't workingSomething isn't working

training

Status: Open.

#6772 In microsoft/DeepSpeed;

· traincheck-team opened on Nov 20, 2024

[BUG] [Fix-Suggested] Checkpoint Inconsistency When Freezing Model Parameters Before deepspeed.initialize

Status: Open.

#6771 In microsoft/DeepSpeed;

· traincheck-team opened on Nov 20, 2024

[BUG] [Fix-Suggested] KeyError in stage_1_and_2.py Due to Optimizer-Model Parameter Mismatch

Status: Open.

#6770 In microsoft/DeepSpeed;

· traincheck-team opened on Nov 20, 2024

grad is None

bugSomething isn't workingSomething isn't working

training

Status: Open.

#6768 In microsoft/DeepSpeed;

· suanflower opened on Nov 20, 2024

[BUG] clip_grad_norm for zero_optimization mode is not working

bugSomething isn't workingSomething isn't working

training

Status: Open.

#6767 In microsoft/DeepSpeed;

· chengmengli06 opened on Nov 20, 2024

[BUG]NCCL operation timeout when training with deepspeed_zero3_offload or deepspeed_zero3 on RTX4090

bugSomething isn't workingSomething isn't working

training

Status: Open.

#6756 In microsoft/DeepSpeed;

· MLS2021 opened on Nov 18, 2024