Lightning-AI/pytorch-lightning

Training loss will increase drastically when resume from ckpts after adopting self-defined sampler in multi-GPU experiments [Fabric]

#21312

· Moondok opened

on Oct 24, 2025

LightingModel.save_checkpoint()

checkpointing

feature

#21280

· Karesto opened

on Oct 10, 2025

[CLI] predict(ckpt_path="best") raises ValueError even though ModelCheckpoint is configured under trainer.callbacks

#21254

· ilinvai opened

on Sep 29, 2025

Directly save checkpoint to specified disk location

checkpointing

question

#21253

· HarveyYan opened

on Sep 29, 2025

Feature request: Support splitting model weights and training states into separate checkpoint files

callback: model checkpoint

checkpointing

feature

#21170

· yilin404 opened

on Sep 7, 2025

FSDP with custom state_dict()

#21124

· LiuTaowen-Tony opened

on Aug 28, 2025

Can't run a model trained in MPS-system on CPU-only system.

#21094

· obbiondo opened

on Aug 18, 2025

Resume training from checkpoint that only save trainable parameters

checkpointing

feature

#21053

· jasonrichdarmawan opened

on Aug 11, 2025

With automatic_optimization disabled and checkpointing every n steps, the best checkpointed model is the model obtained after backpropagation and not the one used for computing the loss

bug

checkpointing

ver: 2.5.x

#20947

· Yann-CV opened

on Jun 26, 2025

fabric FSDP strategy save/load checkpoint does not support s3 url

#20749

· likesum opened

on Apr 24, 2025

Progress bar is broken when loading trainer state from checkpoint

#20603

· JLenzy opened

on Feb 25, 2025

Error learning rate when load ckpt for cotinue training if check_val_every_n_epoch > 1

#20495

· razgzy opened

on Dec 13, 2024

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Training loss will increase drastically when resume from ckpts after adopting self-defined sampler in multi-GPU experiments [Fabric]

LightingModel.save_checkpoint()

[CLI] predict(ckpt_path="best") raises ValueError even though ModelCheckpoint is configured under trainer.callbacks

Directly save checkpoint to specified disk location

Feature request: Support splitting model weights and training states into separate checkpoint files

FSDP with custom state_dict()

Can't run a model trained in MPS-system on CPU-only system.

Resume training from checkpoint that only save trainable parameters

With automatic_optimization disabled and checkpointing every n steps, the best checkpointed model is the model obtained after backpropagation and not the one used for computing the loss

fabric FSDP strategy save/load checkpoint does not support s3 url

Progress bar is broken when loading trainer state from checkpoint

Error learning rate when load ckpt for cotinue training if check_val_every_n_epoch > 1

Issues

Search results