Skip to content

Feat/pr2 ports nanogpt minimind - #22

Merged
CanGeng merged 2 commits into
masterfrom
feat/pr2-ports-nanogpt-minimind
Jun 27, 2026
Merged

Feat/pr2 ports nanogpt minimind#22
CanGeng merged 2 commits into
masterfrom
feat/pr2-ports-nanogpt-minimind

Conversation

@CanGeng

@CanGeng CanGeng commented Jun 27, 2026

Copy link
Copy Markdown
Owner

No description provided.

CanGeng added 2 commits June 27, 2026 11:39
Add examples/nanoGPT/ and examples/MiniMind/ — end-to-end lighttrain
ports of two well-known training repos.

nanoGPT (Karpathy):
  model.py: @register("model","nano_gpt") — identical arch (optional-bias
  LayerNorm, c_attn/c_proj naming, Flash Attn, from_pretrained for GPT-2
  weights, estimate_mfu, weight-decay split optimizer)
  components.py: @register BinaryMemmapDataset("memmap_bin") + StackCollator
  configs/: train_shakespeare_char / train_gpt2 / finetune_shakespeare
  data/: shakespeare_char + shakespeare (BPE) + openwebtext prepare scripts
  sample.py: standalone generation from lighttrain checkpoint or gpt2*

MiniMind (jingyaogong):
  model/: model_minimind.py (verbatim) + model_adapter.py (@register
  "minimind" + @register("tokenizer","hf_auto")) + model_lora.py
  tokenizer: vocab_size=6400 custom tokenizer
  dataset/lm_dataset.py: @register minimind_pretrain/sft/dpo/rlaif datasets
  + minimind_pad collator
  trainer/trainer_utils.py + rollout_engine.py (adapted/verbatim)
  configs/: pretrain / sft / lora / dpo / grpo / ppo / distillation (7 total)
  scripts/: chat_api / web_demo / serve_openai_api / eval_toolcall /
  convert_model (verbatim + sys.path adapted)

All gates green: ruff clean, mypy 575 files 0 errors, pytest 4491 passed.
Verbatim copies (model_minimind, model_lora, rollout_engine, prepare.py)
carry # ruff: noqa to preserve upstream style.

CanGeng
@CanGeng
CanGeng merged commit 485a975 into master Jun 27, 2026
3 checks passed
@CanGeng
CanGeng deleted the feat/pr2-ports-nanogpt-minimind branch June 27, 2026 03:44
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant