huggingface/accelerate

Pull Requests Commits

Release: v0.18.0

sgugger committed 2 years ago

Verified ecd12888

Handle multiple tied parameters (#1241)

sgugger committed 2 years ago

Verified a826e444

Hardware Auto-Setup Example/Tutorial for Distributed Launch (#1227)

Caroline Chen committed 2 years ago

Verified 1fe27e7c

Change multinode to multigpu (#1247)

muellerzr committed 2 years ago

Verified c1a6c209

backfill ds plugin attributes when using ds_config (#1235)

pacman100 committed 2 years ago

Verified 8ebd6ab2

remove empty dicts while saving accelerate config (#1236)

pacman100 committed 2 years ago

Verified ea9b8547

extensions has been removed and replaced by customizations (#1075)

dbpprt committed 2 years ago

Verified 420ff21c

Make grad accum steps mutable on the Accelerator object (#1233)

muellerzr committed 2 years ago

Verified b1b33127

add additional check before deleting env variable (#1229)

Chris-hughes10 committed 2 years ago

Verified 6e4e8702

Silence dynamo_backend (#1226)

muellerzr committed 2 years ago

Verified a3065e18

docs: add finetuner to ppl who use accelerate (#1224)

Wang Bo committed 2 years ago

Verified 4eaf36e1

Fix get_logger kwarg documentation issue (#1222)

bcol23 committed 2 years ago

Verified e7bb060c

Fix bug in loading launch config (#1218)

neumyor committed 2 years ago

Verified a15d3074

FIx TPU gradient state (#1219)

muellerzr committed 2 years ago

Verified 7e7f3445

ds offload optim fix to use CPUAdam (#1208)

pacman100 committed 2 years ago

Verified 10c67463

Fix example in accumulate method (#1211)

VikParuchuri committed 2 years ago

Verified 82c2665c

Fix typo in TPU config (#1202)

muellerzr committed 2 years ago

Verified 2930cac6

Better error message when using multi-GPU and Accelerate on torch <1.9.1 (#1203)

muellerzr committed 2 years ago

Verified 901ab69a

Fix tied weights load (#1204)

sgugger committed 2 years ago

Verified 780e4aa3

Make the Scheduler adjust the steps taken relative to the gradient accumulation steps (#1187)

muellerzr committed 2 years ago

Verified e4620984

Fixup --fsdp (#1198)

muellerzr committed 2 years ago

Verified 017a98c0

[`Accelerator`] We should not call `to` on modules that wraps `accelerate` loaded models (#1172)

younesbelkada committed 2 years ago

Verified d1aa5581

Set drop last to ensure modulo16 restriction for fp8 (#1189)

ksivaman committed 2 years ago

Verified 41479fe4

Only convert linear layers with weights multiple of 16 (#1188)

sgugger committed 2 years ago

Verified eac5d13c

add `use_orig_params` to FullyShardedDataParallelPlugin (#1184)

pacman100 committed 2 years ago

Verified b228136c

Add documentation about PyTorch FSDP state dict behavior (#1181)

VikParuchuri committed 2 years ago

Verified 90deb748

Support special mapping of dtypes when preparing device map (#1179)

sgugger committed 2 years ago

Verified d9427087

fixed typo in launch.py tpu_pod_launcher (#1180)

hackpert committed 2 years ago

Verified 37831808

Add repr to AlignHook for easier debugging. (#1177)

sgugger committed 2 years ago

Verified ea836f30

Run accelerate_test in cli (#1176)

muellerzr committed 2 years ago

Verified a4c94762

Older