litgpt · Issues· 290 open

No checkpoints saved during pretraining in Google Colab despite log.log_model and save_interval set.

LLM.generate() can silently select a token excluded by top_k on CPU when sampling probabilities are float16

False positive warning about mixed precision in `merge_lora.py`

Reang

litgpt pretrain crashes or deadlocks when a rank's val_dataloader yields zero batches (multi-GPU FSDP)

Make `save_hyperparameters()` robust against different CLI entry points

Speculative decoding drops the first generated token

Stop sequence matching mishandles overlapping prefixes and buffered tokens

Unsafe Checkpoint Loading - This Is Bad

Support for SSM models (Mamba, Mamba2)

Initial and final evaluation in `finetune` scripts do not accumulate over devices

failure converting pretrained litgpt checkpoints to HF format: a reproducible example

Gradient Clipping Doesn't Work in Finetuning

Calculate loss at beginning and end of training

Add TinyStories to the pretraining docs