fourlhs/nano-gpt-z

fix: handle missing checkpoint gracefully, remove timeout param

refactor: replace Gradio with Flask for better reliability

debug: add logging to generate function

fix: dont skip sentence if only one sentence generated

fix: cast slider parameters to correct types (int/float)

feat: auto-download checkpoint from GitHub release

fix: improve checkpoint not found error message

debug: add detailed logging for checkpoint path discovery

docs: add HF Spaces README with configuration

fix: make checkpoint path flexible to support multiple locations

fix: update Dockerfile to expect checkpoint in root

chore: add HuggingFace Spaces deployment (Dockerfile, .dockerignore)

Merge remote changes

fix: fix export func bug in export.py

fix: fix naming bug in export.py

fix: fix loading base bug in evaluate.py

typo

fix: fix loading base model bug in finetune.py

fix(generation): fix top_p mask bug

new params for a new train

update comments

update load_checkpoint in train.py

update load_checkpoint in train.py

add torch.compile

start over

comment out torch.compile to find out whether thats why loss isnt decreasing

upd params on train.py

upd params on train.py

upd lr and min_lr

changed warmup_steps

changed batch_size in train.py

fix: remove GradScaler from train.py and change params

fix: adjust eval_interval and add token assertion

feat: increase pretraining tokens to 1B

feat: add bfloat16 mixed precision and torch.compile

refactor: use F.scaled_dot_product_attention for optimized attention

feat: add env variables and fix comments

feat: add build.sh

feat: update training pipeline, quantised WASM inference, and data prep. fix minor bugs.

feat: implement cpp inference architecture

add evaluation pipeline (perplexity + style shift)

feat: implement finetuning loop

add data pipeline: FineWeb-Edu pretraining + Gen Z finetuning + WikiText-103 coherence test

feat: add training loop

initial commit: prepare folder structure and implement model.py