All issues

X-LoRA: disable_adapter does not disable the experts, and generate() freezes the classifier / switches the model to eval mode

#3635

· amogh-nagri-11 opened

on Sep 1, 2026

Fix state-dict round-trip for directly injected shared tuners

#3631

· CoralGarden52 opened

on Aug 31, 2026

TP + LoRA is broken on transformers 5.16: add_tensor_parallel_hooks_to_module and EmbeddingParallel were removed

#3628

· alanhuangyoo opened

on Aug 31, 2026

[Bug] ShadowPEFT LM-head trainability freeze after set_adapter — modules_to_save=["lm_head"] silently stops receiving grads

#3626

· Sravanjangam opened

on Aug 30, 2026

[Bug] ShadowPEFT hook-lifecycle leak: per-forward GPU growth + UAF on unload_shadow(copy=False)→delete_adapter

#3625

· Sravanjangam opened

on Aug 30, 2026

[Bug] ShadowPEFT sequence classification pools the wrong token with left padding

#3620

· aswanth-07 opened

on Aug 30, 2026

[Bug] Prefix tuning crashes for CPU-only device maps

#3619

· aswanth-07 opened

on Aug 30, 2026

[RFC, I can do a PR] Make `lora_alpha` a float

#3615

· sliedes opened

on Aug 28, 2026

layers_pattern does not match when the nn.ModuleList is at the root of the model

#3610

· LK-maker-007 opened

on Aug 26, 2026

Merging multiple DoRA adapters that share a target module does not match the unmerged forward pass

#3608

· LK-maker-007 opened

on Aug 26, 2026

LoHa/LoKr merge in train mode bakes stochastic rank_dropout masks into merged weights

#3586

· Sravanjangam opened

on Aug 23, 2026

save_pretrained writes an empty/truncated adapter checkpoint (silently) when another adapter name equals a base-model module path component

#3584

· Sravanjangam opened

on Aug 23, 2026