All issues

[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

IncrementalPCA: fit() accumulates instead of refitting, and overwrites its constructor arguments

#3617

· Sreekant13 opened

on Aug 29, 2026

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

#3615

· sliedes opened

on Aug 28, 2026

LN Tuning add_adapter moves existing adapters back to the base-layer device

#3612

· DaoyuanLi2816 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

eva_config / corda_config / lora_ga_config / arrow_config become plain dicts after save→from_pretrained round-trip (AttributeError far from cause)

#3583

· Sravanjangam opened

on Aug 23, 2026