huggingface/transformers

Pull Requests Commits

allocate target device memory before using hmll

mfuntowicz committed 20 days ago

983c8be9

do not load model twice

mfuntowicz committed 21 days ago

f5a2e47e

mfuntowicz committed 21 days ago

1417782d

remove unecessary param for now

mfuntowicz committed 21 days ago

73c4b42f

improve on the capacity guard for safetensors binding

mfuntowicz committed 21 days ago

3b9e9ffe

missing refactoring

mfuntowicz committed 21 days ago

2efccf38

avoid branching for contiguous/scattered I/O over HmllLoadSpec

mfuntowicz committed 21 days ago

51322cfb

handle the case where we are not using hmll

mfuntowicz committed 22 days ago

ff1b9d9d

add HF_USE_HMLL guard against hmll loading

mfuntowicz committed 22 days ago

4f149009

attempt to avoid peak GPU memory usage upfront loading

mfuntowicz committed 22 days ago

3807d8c8

fix issue when using replicated weights should dispatch to fetch

mfuntowicz committed 22 days ago

e2484d19

more wip integration

mfuntowicz committed 23 days ago

45ee35a1

wip tensor parallel with HMLL path

mfuntowicz committed 23 days ago

4283c7f8

feat(tensor_parallel): split up the tp slice(s) computation and introduce a method to navigate slices over the flattened tensor view

mfuntowicz committed 23 days ago

c300ce78

use pytorch cuda memory arena

mfuntowicz committed 23 days ago

b5b64d9b

feat(hmll): load sharded safetensors

mfuntowicz committed 23 days ago

0b35dee3

dirty use of hmll to load weights

mfuntowicz committed 23 days ago

c1c3c965

Support non-gated experts (#44319)

IlyasMoutawwakil committed 23 days ago

Verified 28d02a31

Fix speculative tests that are flaky with SDPA (#44340)

Rocketknight1 committed 23 days ago

Verified 41888df5

[Bugfix] fix qwen3.5 no split module (#44382)

JJJYmmm committed 23 days ago

Verified c1838e8b

Fix mutable default arguments and resource leaks (#44287)

jashshah999 committed 23 days ago

Verified c4dd8180

fix(tokenizer): Fix MLukeTokenizer AttributeError post-v5 refactor (#44362)

harshaljanjani committed 23 days ago

Verified 7c4c9375

[VibeVoice ASR] Use updated padding cache for ASR model. (#44392)

ebezzam committed 23 days ago

Verified 744639ab

[higgs-audio-v2] fix sampling (#44386)

eustlb committed 23 days ago

Verified 8468dbf7

skip 2 invalid test cases for voxtral_realtime model (#44321)

kaixuanliu committed 23 days ago

Verified dcd3d720

Add VibeVoice ASR (#43625)

ebezzam committed 23 days ago

Verified f2f0d17d

Mamba-1/-2 init weights in mixer class (#43778)

kevinli573 committed 23 days ago

Verified 1391366f

add expectations for xpu for olmo_hybrid model (#44353)

kaixuanliu committed 23 days ago

Verified 9dbc5356

Model patching API (#43917)

IlyasMoutawwakil committed 23 days ago

Verified 6a92ebdd

[VITS] Add `speaking_rate` as an optionl forward argument (#43283)

gau-nernst committed 23 days ago

Verified e58be565

Older