[gspark-etri]

Background

The repository is ai-computing/npu-train and trains with LoRA, full fine-tuning and MoE, but the Python package was still npu_lora. People run python3 -m npu_lora.<module> ... on shared hosts right now (for example npu_lora.moe_bench2, a script copied into the package directory), and some hosts have an editable install made before the rename, so the old name must keep working (#46).

Changes

• git mv npu_lora npu_train. The package uses relative imports, so the modules are unchanged apart from docstrings/usage lines; k8s.py now mounts the ConfigMap at /opt/npu-lora/npu_train and runs python3 -m npu_train.cli; the Helm job runs npu_train.cli.
• pyproject.toml: distribution npu-train; packages npu_train* and npu_lora*; CLI npu-train added, npu-lora kept; airm workload npu-lora now points at npu_train.airm:train. Helm install.pip installs npu-train @ git+....
• npu_lora/__init__.py is a compatibility shim:
  • emits DeprecationWarning on import;
  • installs a meta-path finder: npu_lora.X resolves to the same module object as npu_train.X (one copy of every class/global; __spec__ stays npu_train.X), and runpy gets the code from it, so python -m npu_lora.X runs npu_train.X as __main__;
  • a module that exists only under npu_lora/ (an untracked script left there on a host, e.g. moe_bench2.py) is still found there and can import npu_lora.* / relative names;
  • if npu_train is not importable (editable install made before the rename maps only npu_lora), it appends the sibling directory to sys.path;
  • top-level names (npu_lora.LoRATrainer, __version__) come from npu_train.
• tests/test_compat.py (new, 4 tests, each in a fresh interpreter): import + warning + module identity; python -m npu_lora.cli --help and python -m npu_lora.compare --help; a script inside npu_lora/ run with -m; an old-style editable finder that maps only npu_lora.
• Imports in tests, README (EN/KO), docs/, ai-rm.yaml comment, scripts/vendor-into-lmd-top.sh (copies npu_train/*.py, destination assets/npu-lora unchanged) and the STATUS goal line use npu_train.

Verification

• Checks / tests run: full CPU suite in a clean venv with this branch installed (pip install -e .; the host's old npu-lora editable install excluded so entry-point metadata comes from this branch).
• Commands: CUDA_VISIBLE_DEVICES= python -m pytest -q tests
• CI (cpu-tests, Python 3.11, CPU torch): 38 passed (run 39713975072).
• Results: 38 passed in 182 s (34 existing + 4 new in test_compat.py). Python 3.10.12, torch 2.10.0+cu128 (CPU only, no NPU/GPU used), transformers 5.1.0.
  • In the shared user environment (old npu-lora 0.1.0 editable install still registered) the same run gives 37 passed, 1 failed: test_airm.py::test_registry_loads_runner sees the stale entry point npu_lora.airm:train. That is the installed metadata, not the code; it goes away after reinstalling (see below). CI installs fresh.
  • Mutation check: removing the sys.path fallback fails test_old_editable_install_mapping_only_npu_lora; removing the __spec__ restore fails test_import_old_name_is_same_module_and_warns.
• Also checked: npu-train --help, npu-lora --help, python -m npu_lora.cli --help from outside the repo in the venv; npu_train.k8s.build(...) emits npu_train.cli and the /opt/npu-lora/npu_train mount; scripts/vendor-into-lmd-top.sh <tmpdir> copies the 17 modules. helm template not run (helm not installed on this host).
• Sensitive-data check (check_sensitive.py): clean, no new findings.

Compatibility and limits

• Kept on purpose (other systems depend on them; renaming them is a separate change): CLI command npu-lora, ai-rm workload name npu-lora, Prometheus metrics npu_lora_*, npu_lora_meta.json, NPU_LORA_RNGD* env vars, cache dir ~/.cache/npu-lora, Helm chart name/path deploy/helm/npu-lora, lmd-top vendoring path assets/npu-lora.
• Not changed: experiment records and scripts under experiments/ (they document what was run; their python3 -m npu_lora.* commands keep working through the shim).
• Existing hosts: after pulling, old commands keep working without reinstalling. To get the new metadata (entry point, npu-train command): pip uninstall -y npu-lora && pip install -e . (otherwise two distributions register the npu-lora command and workload).
• DeprecationWarning follows Python's default filters: shown when npu_lora is imported from __main__ code or under pytest / -W default, hidden for python -m npu_lora.X.

Related Issue

Closes #46

Checklist

• One purpose; unrelated formatting, dependency or API changes are in separate PRs
• Required checks pass locally or in CI
• No IPs, internal hostnames, tokens or site-specific paths (sensitive-data check run)
• Docs updated (EN, and KO where the repository keeps both)
• AI assistance stated, if used (written with Claude Code)

🤖 Generated with Claude Code