microsoft/DeepSpeed

Pull Requests Commits

fix gradient accumulation

Masahiro Tanaka committed 1 year ago

cbda88da

udpate fp32 master params of bf16 optimizer after allreduce

Masahiro Tanaka committed 1 year ago

dc801066

fix gradient clipping

Masahiro Tanaka committed 1 year ago

794e9921

Update deprecated HuggingFace function (#5144)

mrwyattii committed 1 year ago

Verified a37e59b5

Fix UserWarning: The torch.cuda.*DtypeTensor constructors are no long… (#5018)

ShukantPal committed 1 year ago

Verified 177dc143

TestEmptyParameterGroup: replace fusedAdam with torch.optim.AdamW (#5139)

nelyahu committed 1 year ago

Verified a3be0d44

Added missing `isinstance` check in PR 5112 (#5142)

bm-synth committed 1 year ago

Verified 9bd62e0b

Method `run_map_reduce` to fix errors when running `run_map` followed by `run_reduce` (#5131)

bm-synth committed 1 year ago

Verified 95be499a

Update nv-inference torch version (#5128)

loadams committed 1 year ago

Verified 39f6ee59

Check overflow on device without host synchronization for each tensor (#5115)

BacharL committed 1 year ago

Verified 1c9e5ef2

tensorboard logging: avoid item() outside gas to improve performance (#5135)

nelyahu committed 1 year ago

Verified fac1df99

remove mandatory `index` key from output of `metric_function` in `DataAnalysis` map operation (#5112)

bm-synth committed 1 year ago

Verified 2d0a6bc2

Avoid zero-sized microbatches for incomplete minibatches when doing curriculum learning (#5118)

bm-synth committed 1 year ago

Verified a7864846

Fix broadcast deadlock for incomplete batches in data sample for data analysis (#5117)

bm-synth committed 1 year ago

Verified 2b411103

Solve tensor vs numpy dtype conflicts in data efficiency map-reduce. (#5108)

bm-synth committed 1 year ago

Verified 3e5d4004

add missing methods to MPS_Accelerator (#5134)

mrwyattii committed 1 year ago

Verified b20c4674

Update return codes in PyTest to properly error out if tests fail (#5122)

loadams committed 1 year ago

Verified 083197ea

Fix issue with zero-sized file after merging file on curriculum `map_reduce` (#5106)

bm-synth committed 1 year ago

Verified 9c696620

Update inference test model names (#5127)

mrwyattii committed 1 year ago

Verified d532f643

disable compile tests for torch<2.1 (#5121)

mrwyattii committed 1 year ago

Verified d67d4e5b

removed if condition in `if collate_fn is None` (#5107)

bm-synth committed 1 year ago

Verified 9aaead55

BF16 optimizer: Improve device utilization by immediate grad update (#4975)

deepcharm committed 1 year ago

Verified 3c0bd312

[NPU]ZeRO-Infinity feature compatibility (#5077)

misstek committed 1 year ago

Verified 0d9cfa01

Stop tracking backward chain of broadcast (ZeRO3) (#5113)

tohtana committed 1 year ago

Verified f295aea0

Update version.txt after 0.13.2 release (#5119)

mrwyattii committed 1 year ago

Verified 41bc9fee

Remove optimizer step on initialization (#5104)

tohtana committed 1 year ago

Verified 18179807

Add HIP conversion file outputs to .gitignore (#5111)

lekurile committed 1 year ago

Verified 25a02047

Skip Triton import for AMD (#5110)

lekurile committed 1 year ago

Verified d04a8386

Fix alignment of optimizer states when loading (#5105)

tohtana committed 1 year ago

Verified 0a10bd42

Support `exclude_frozen_parameters` for `zero_to_fp32.py` script (#4979)

andstor committed 1 year ago

Verified 2518cc42

Older