Commits huggingface/transformers

[docs] Mention models whose attention bypasses the attention interface

3outeille committed 4 minutes ago

12ac024c

[docs] Add a context parallelism page

3outeille committed 28 minutes ago

dba74190

[distributed] Test context parallelism against the same weights on full sequences

3outeille committed 34 minutes ago

635bb98e

[distributed] Document context parallelism in the load and config docstrings

3outeille committed 34 minutes ago

5964f5d8

[distributed] Refuse context parallelism in Trainer

3outeille committed 34 minutes ago

95e96cb3

[distributed] Split the sequences across the cp group when the model runs

3outeille committed 34 minutes ago

8051af49

[distributed] Refuse models whose layers or head context parallelism would break

3outeille committed 34 minutes ago

cb227c6d

[distributed] Add the cp axis to the device mesh and shard FSDP2 over fsdp x cp

3outeille committed 34 minutes ago

931c06b8

[distributed] Add context parallelism sizes to DistributedConfig

3outeille committed 46 minutes ago

faf2d3e7

[distributed] Reinstall the context parallel attention wrapper when its key was registered again

3outeille committed 46 minutes ago

ba95ec66

[distributed] Add tests for the context parallelism implementations

3outeille committed 1 hour ago

4890896b

[distributed] Add the context parallelism base and the Ulysses implementation

3outeille committed 1 hour ago

f4cb9761

[torch 2.15] Update qwen3_vl_moe FA2 integration test expected values (#50481)

ydshieh committed 2 hours ago

Verified f262a8cd

fix(gemma3n): Clamp the AltUp magnitudes instead of using a CPU epsilon tensor (#50457)

harshaljanjani committed 8 hours ago

Verified 431e9e9d

[Mistral Large 4] Fix flaky test_left_padding_compatibility (#50479)

ydshieh committed 16 hours ago

Verified de35c2b1

[torch 2.15] Update integration test expected values for A10 (cuda 8.6) (#50478)

ydshieh committed 18 hours ago

Verified 0f7a8768

[Docker] Update Nvidia CI and Flash Attn CI images to torch 2.15 / cu132 (#50467)

ydshieh committed 1 day ago

Verified a6f5c867

Clarify warmup_steps error message to warmup_steps >= 0 (#50364)

adi274903 committed 2 days ago

Verified 0f53828e

[Gemma4] Stop prefill at last KV-owning layer when logits_to_keep=1 (#49640)

pculliton committed 2 days ago

Verified 2c6a00f9

Fix duplicated docstring (#50446)

Cyrilvallez committed 2 days ago

Verified 073a202c

chore(ci): refresh workflow references (#50436) (#50437)

hf-security-analysis[bot] committed 2 days ago

Verified 04b0d424

[serge] Fix 2 integration tests for model `pegasus` failing with `other` (other (2)) (#49507)

sergereview[bot] committed 2 days ago

Verified d9246b06

[distributed - TPU] Fix tied-weight comparisons on TPUs when using TP (#50406)

tengomucho committed 2 days ago

Verified 92b21f8d

vLLM CI: fix install and KV-cache memory, add failure report job (#50436)

ydshieh committed 2 days ago

Verified 826edb70

Harden our Jinja sandboxing (#50431)

Rocketknight1 committed 2 days ago

Verified 316217d3

Weight loading: conversion ops declare many-to-many support, and four small loader fixes (#49773)

IlyasMoutawwakil committed 2 days ago

Verified cea26862

Fix DeepSeek-V4 compression cache reordering for beam search (#50348)

taking-lying-flat committed 2 days ago

Verified 22c26e79

Allow saving an unchanged generation config (#49914)

benthecarman committed 2 days ago

Verified 7fb5f73b

[loading] Fix mismatched keys being silently ignored (#50433)

Cyrilvallez committed 2 days ago

Verified b6bb35a3

Capture only the projected hidden states in VLM feature extraction (#49594)

hmellor committed 2 days ago

Verified c5bd76ec