Commits huggingface/accelerate

Merge branch 'main' into use-supports-context-parallel

qgallouedec committed 25 days ago

Verified 2483a167

Fix accelerate env never reading the MPS chip name (#4221)

yupengtang committed 25 days ago

Verified 9e5d1de5

docs: fix mixed precision argument in examples (#4192)

noobyalan committed 25 days ago

Verified ede1c22b

Fix compile_regions running the uncompiled module (#4188)

hjinnkim committed 25 days ago

Verified a4d394de

Forward reduce_batch_size_fn through the find_executable_batch_size decorator form (#4160)

vineethsaivs committed 25 days ago

Verified e07f2555

docs: fix documented parameter names that do not match the signature (#4156)

massimiliano1991 committed 25 days ago

Verified 29eb9954

fix: handle FP8 dtypes in disk offload write/read path (#4151)

shoemoney committed 25 days ago

Verified 7b16bf3f

Fix memory accounting for integrated CUDA devices (#4187)

harshitakrishna18 committed 25 days ago

Verified 069b3d6d

Reach through the DeepSpeed optimizer wrapper in AcceleratedOptimizer.eval() (#4130)

vineethsaivs committed 25 days ago

Verified e965f100

Handle missing accelerate executable in env command (#4168)

tandede committed 25 days ago

Verified 6bd00de6

chore(deps): bump the actions group across 1 directory with 4 updates (#4182)

dependabot[bot] committed 25 days ago

Verified f627c45e

Report MPS as the accelerator in accelerate env. The accelerator detection chain in (#4158)

xquantize committed 25 days ago

Verified 004717a5

Add type hints to utils/memory.py (#4152)

RudrenduPaul committed 25 days ago

Verified 69db6aa8

Add missing type hints in checkpointing.py (#4153)

RudrenduPaul committed 25 days ago

Verified 8eb8f115

Fix distributedType -> distributed_type typo in downcast_bf16 guard (#4149)

uttam12331 committed 25 days ago

Verified 5d0176a9

Raise a clear error when FSDP is enabled on a mesh with no shard dimension (#4180)

qgallouedec committed 29 days ago

Verified fd132822

prepare_model: don't move DTensor-sharded models to the device (#4181)

qgallouedec committed 29 days ago

Verified f7cc8e9a

Add FSDP2 `activation_checkpointing_offload`: offload checkpointed layer inputs to pinned CPU memory (#4175)

qgallouedec committed 31 days ago

Verified 2ab9b2cc

Trim the context-parallel guard comment

qgallouedec committed 32 days ago

c19955a4

Ask transformers whether a model supports context parallelism

qgallouedec committed 32 days ago

10dd4ecf

Refuse context parallelism for models with sliding-window or chunked attention layers (#4177)

qgallouedec committed 32 days ago

Verified bd204169

Fix convert_model_to_fp8_ao converting the first and last linear layers (#4147)

vineethsaivs committed 32 days ago

Verified 2b55f5a2

FSDP2 activation checkpointing: wrap the matched transformer layer itself, not each of its children (#4172)

qgallouedec committed 43 days ago

Verified fd01e35c

[test] Update min version torch (#4170)

SunMarc committed 43 days ago

Verified 2583573a

Fix FSDP2 tied-embedding models on torch >= 2.13: put output embedding in the same fully_shard group (#4171)

qgallouedec committed 43 days ago

Verified f8012607

fix(security): remediate workflow vulnerability in .github/workflows/build_and_run_tests.yml (#4169)

hf-security-analysis[bot] committed 45 days ago

Verified 7fccde20

chore(deps): bump the actions group across 1 directory with 7 updates (#4167)

dependabot[bot] committed 45 days ago

Verified fc3d11b6

Ensure device-agnostic in docs (#4163)

kaixuanliu committed 45 days ago

Verified dc78670c

Make doc builds faster (#4103)

mishig25 committed 71 days ago

Verified 16cb6eb8

Fix FSDP sharded checkpoint path resolution (#4119)

HaomingSong committed 71 days ago

Verified e625c7a3