Commits huggingface/transformers

Collapse the masks to one tensor only for models with a single layer type

qgallouedec committed 41 minutes ago

c548a7b4

Merge branch 'cb-linear-attention' into cb-cache-layout

qgallouedec committed 43 minutes ago

1aa1ba90

Keep the engine's masks as a dict on hybrid models

qgallouedec committed 43 minutes ago

2d67d9b7

Declare cache rows and states by name, and keep every layer's states in one allocator

qgallouedec committed 52 minutes ago

b07b9f91

Size the paged cache from the rows the layers declare, not from the model config

qgallouedec committed 1 hour ago

7d8ca02e

Name the declaration cache_states: the states a layer asks the cache to keep per request

qgallouedec committed 2 hours ago

e05b2713

Type the allocator as the linear one where the cache reads its state views

qgallouedec committed 2 hours ago

237b61ba

Leave the warmup stream sync to its own pull request

qgallouedec committed 2 hours ago

b17183b5

Let the linear attention layers declare the states the paged cache keeps for them

qgallouedec committed 2 hours ago

ab7c8fe5

Wait for each warmup forward before the next one rewrites the static inputs

qgallouedec committed 5 hours ago

64def275

Continuous batching for linear attention layers

qgallouedec committed 6 hours ago

2b96d3db

Replace torch.FloatTensor with torch.Tensor in the model files (#50183)

qgallouedec committed 7 hours ago

Verified fd3cf014

Keep image inputs on the first chunk of a chunked prefill (#49753)

digit50 committed 7 hours ago

Verified 0c7f5ab8

Restore BLT's 512-byte sliding windows (entropy patcher + local encoder/decoder) (#49188)

hyunW3 committed 8 hours ago

Verified baa3646e

[Video] Require one metadata entry per video and refuse return_metadata=False (#50060)

paulpak58 committed 15 hours ago

Verified f2754311

:rotating_light: Finally drop rope deltas from ModelOutput! (#50153)

zucchini-nlp committed 16 hours ago

Verified d6c0ec4b

Docs: note in the v5 migration guide that hidden_states[-1] of vision models is pre-norm (#50303)

ravediamond committed 18 hours ago

Verified e540ae45

Fix model_input_names singleton issue causing shared state (#42051)

yashwantbezawada committed 18 hours ago

Verified 41ce391e

fix(fetcher): case-sensitive submodule check on case-insensitive filesystems (Windows) (#50326)

ydshieh committed 20 hours ago

Verified 3568df44

[tests] Pin the chat-template date in Llama's FA static-cache test (#50296)

tarekziade committed 20 hours ago

Verified a6a77004

fix(cosmos3_edge): bump backend_equivalence_atol for Pillow 13.x (miss from #50105) (#50327)

ydshieh committed 21 hours ago

Verified 9719c27f

feat(ci): runtime dep recorder — track transformers imports per test file (#50312)

ydshieh committed 22 hours ago

Verified 5572e702

Fix a cache test (#50320)

Cyrilvallez committed 1 day ago

Verified 8cfd4c5a

Keep overrides of the deprecated gradient checkpointing methods working (#50310)

qgallouedec committed 1 day ago

Verified 4a696b1c

Improve Mistral Large 4 cache sharing (#50305)

Cyrilvallez committed 1 day ago

Verified 4b7a509a

Test the dispatched FA3 kernel per device (#50304)

Abdennacer-Badaoui committed 1 day ago

Verified 581e59a9

Fix Qwen4-Exp eager attention NaNs with left padding (#49923)

taking-lying-flat committed 1 day ago

Verified 23ab8643

fix(pipelines): preserve batched generation metadata (#50210)

BlazeSR10hawk committed 1 day ago

Verified 11d50074

upgrade version for mamba-ssm and fla kernel (#50294)

kaixuanliu committed 1 day ago

Verified 944d028d

perf(pipelines): lazy-load pipeline submodules in pipelines/__init__.py (#50173)

ydshieh committed 1 day ago

Verified f7184029