deepspeedai/DeepSpeed

Pull Requests Commits

Fix gradient accumulation for Z2+offload (#6550)

tohtana committed 1 year ago

Verified c85c8703

[Accelerator] Cambricon MLU support (#6472)

Andy666G committed 1 year ago

Verified 0fbe96a5

DeepNVMe perf tuning (#6560)

tjruwase committed 1 year ago

Verified a5400974

Use msgpack for p2p comm (#6547)

tohtana committed 1 year ago

Verified 7622cd9e

Skip failing newly added tests in accelerate (#6574)

loadams committed 1 year ago

Verified 61de0171

Added Intel Gaudi to Accelerator Setup Guide (#6543)

ShifaAbu committed 1 year ago

Verified 2a56f533

Add conditional on torch version for scaled_dot_product_attention (#6517)

loadams committed 1 year ago

Verified 170b46e8

Avoid security issues of subprocess shell (#6498)

tjruwase committed 1 year ago

Verified 659f6be1

wrap include cuda_bf16.h with ifdef BF16_AVAILABLE (#6520)

oelayan7 committed 1 year ago

Verified c2748393

Revert "BF16 optimizer: Clear lp grads after updating hp grads in hook" (#6508)

nelyahu committed 1 year ago

Verified 8fa6b50b

fix environment variable export bug for MultiNodeRunner (#5878)

TideDra committed 1 year ago

Verified fc22d960

Fix the broken url link (#6500)

rogerxfeng8 committed 1 year ago

Verified 2a647c51

fix pipeline eval_batch micro_batches argument for schedule (#6484)

nelyahu committed 1 year ago

Verified 3b09d945

Op_builder->is_compatible quite warning (#6093)

terry-for-github committed 1 year ago

Verified 4f803852

HPU: add required ENV vars to acccelerator init (#6495)

nelyahu committed 1 year ago

Verified 857780a8

Update version.txt after 0.15.1 release (#6493)

loadams committed 1 year ago

Verified c210e601

Handle an edge case where `CUDA_HOME` is not defined on ROCm systems (#6488)

amorehead committed 1 year ago

Verified 10ba3dde

Safe usage of popen (#6490)

tjruwase committed 1 year ago

Verified 662a421b

Add default value to "checkpoint_folder" in "load_state_dict" of bf16_optimizer (#6446)

ljcc0930 committed 1 year ago

Verified ddd35718

DS_BUILD_OPS should build only compatible ops (#6489)

tjruwase committed 1 year ago

Verified 5d1a30c0

Fix patch for parameter partitioning in zero.Init() (#6388)

tohtana committed 1 year ago

Verified ddeb0c19

print warning if actual triton cache dir is on NFS, not just for default (#6487)

jrandall committed 1 year ago

Verified 9d17116f

DeepNVMe tutorial (#6449)

tjruwase committed 1 year ago

Verified 5df12a4a

bf16_optimizer: fixes to different grad acc dtype (#6485)

nelyahu committed 1 year ago

Verified cfc6ed37

Add workflow to build DS without torch to better test before releases (#6450)

loadams committed 1 year ago

Verified 9b7fc545

TestLowCpuMemUsage UT get device by device_name (#6397)

raza-sikander committed 1 year ago

Verified 89c4d9f5

Avoid gds build errors on ROCm (#6456)

rraminen committed 1 year ago

Verified a7ffe540

[CCL] fix condition issue in ccl.py (#6443)

YizhouZ committed 1 year ago

Verified 0cd9bf59

Change GDS to 1 AIO thread (#6459)

jomayeri committed 1 year ago

Verified f2739b4f

Allow triton==3.0.x for fp_quantizer (#6447)

siddartha-RE committed 1 year ago

Verified 4864991f

Newer Older