$ the-wire · showcase
TRL 1.15 lands a fused LM head, transformers finishes its RoPE cleanup
By RepoJournal · Filed · About Hugging Face · Composed from the cited sources · methodology
11 people shipped this
Yih-Dar
@ydshieh
1 cited
Raushan Turganbay
@zucchini-nlp
1 cited
qgallouedec
@qgallouedec
1 cited
albertvillanova
@albertvillanova
1 cited
sirahd
@sirahd
1 cited
Rocketknight1
@Rocketknight1
Xiaolong Ma
@maxiaolong03
Pauline Bailly-Masson
@paulinebm
jackboyla
@jackboyla
tavallaie
@tavallaie
Andrés Marafioti
@andimarafioti
TRL's fused LM head trades the full logits tensor for a tiled Triton kernel, while transformers closes out the deprecated rope device argument across 11 more model families.
v1.15.0 huggingface/trl
SFT, DPO, KTO, GRPO, RLOO and Distillation now score tokens through a Triton kernel that projects hidden states through the LM head in tiles and reduces straight to per-token log-probs and entropy, never materializing the [batch, seq, vocab] logits tensor. TRL measures this as up to 6.9x longer sequences on the same GPU; the same release drops the deprecated use_liger_kernel argument from the d...
Complete PR #49228: remove unused device=None from rope __init__ in 11 model families (#49285) huggingface/transformers
PR #49228 removed device=None from the rope __init__ signatures in most models, but 11 model families with custom rope classes kept the argument while no longer passing it to rope_init_fn, breaking test_model_rope_s. Custom rope implementations in those families now match the standard signature.
:rotating_light: SubConfigSpec for centralized init of subconfigs (#48282) huggingface/transformers
SubConfigSpec centralizes how subconfigs are constructed, including skipping models with bad defaults and a draft fix for saving nested configs. The commit sequence runs through batched model conversions and CI churn, with more fixes flagged as still to come.
Read chat templates shipped as named variants by their default variant huggingface/trl
Cohere checkpoints ship their chat template as named variants, which transformers loads as a dict; has_generation_markers crashed on that dict with a TypeError at init. TRL now reads the `default` variant, so SFTTrainer with assistant_only_loss=True initializes against checkpoints such as CohereLabs/aya-expanse-8b.
feat(cas_types): report the stored shard hash from a shard upload huggingface/xet-core
UploadShardResponse::shard_hash is now populated with what the server actually wrote, since it re-serializes the shard and keys the object on the result, making the uploader's own hash a prediction. Client::upload_shard returns Option<MerkleHash> instead of inventing one, and the field is omitted by servers predating it. Elsewhere: PP-DocLayoutV4 model support, a clamp fix in its resize path, a...
