MakazhanAlpamys/Soup v0.40.1

v0.40.1 — QA Hardening

on GitHub

What's New

Patch release closing 14 bugs from the v0.40.0 QA pass on Windows + RTX 3050 4 GB, plus the deferred multi-objective preference loss runtime stub.

• UTF-8 stdio bootstrap (Windows) — one-line fix at the CLI entrypoint reconfigures stdout/stderr to UTF-8 before any Rich console init. Closes 6 separate QA findings (β / ✓ / box-drawing crashes on cp1251/cp1252). POSIX no-op.
• Multi-objective preference loss is live — training.preference_loss_weights: {dpo: 0.7, simpo: 0.3} no longer raises NotImplementedError. Builds the highest-weighted loss as primary; rejects BCO + paired loss blends at runtime (data-format incompatible). Pure-function math kernel exercised by 24 tests.
• Smarter defaults on small GPUs — soup quickstart auto-switches to SmolLM2-135M-Instruct on ≤6 GB VRAM (verified in 5 s on RTX 3050 4 GB); autopilot fallback now defaults to 1B instead of 7B and reads the local safetensors index when available.
• soup doctor upgrades — flags transformers ≥ 5.0.0 as INCOMPATIBLE; distinguishes "no GPU hardware" from "GPU hardware present, wrong torch wheel" (nvidia-smi succeeds but torch.cuda.is_available() returns False); detects dual-Python interpreter setups; uses importlib.metadata as version-probe fallback.
• CLI UX consistency — soup init --force non-interactively overwrites; --template help generated from the live registry; soup migrate <data.jsonl> errors with "did you pass the wrong file?"; soup eval custom -o now writes JSON regardless of --attach-to-registry; soup recipes show <typo> suggests close matches via difflib; soup data sample filenames embed strategy (no overwrite); JSONL loader auto-strips UTF-8 BOM.
• Schema strictness — root-level lora: in YAML (the LlamaFactory / Axolotl convention) now migrates into training.lora so nested validators (including init_strategy: random|pissa|olora) actually fire instead of being silently dropped.
• --find-lr actually runs the live loop — fixed broken load_local import that previously always silently fell through to a static placeholder curve.

Install / Upgrade

pip install -U soup-cli

Or via Docker (GHCR):

docker pull ghcr.io/makazhanalpamys/soup:v0.40.1

Security

• UTF-8 stdio bootstrap swallows (OSError, ValueError, AttributeError) on detached streams; never overrides user-set PYTHONIOENCODING.
• _remap_root_level_misplaced_keys does not mutate caller dict (shallow-copy policy mirroring v0.33.0 #47 / v0.40.0 Part B).
• _probe_cache_param_count rejects empty / null-byte model names before path construction.
• nvidia-smi GPU label is rich.markup.escaped before embedding in markup string (NVIDIA Quadro [T4] cannot break or inject markup).
• Dual-Python detector uses os.path.realpath (not Path.resolve()) for Windows 8.3 short-name compat.
• combine_losses rejects empty weights, propagates NaN loudly, rejects bool weight values.

Known Limitations

• Multi-objective preference loss is a primary-loss approximation — the highest-weighted loss is selected as the primary inner trainer; auxiliary losses are named in the advisory but do not yet contribute to the backward pass. Full per-batch weighted-loss combination across all named losses (one forward pass shared across DPO/SimPO/ORPO/IPO via TRL trainer subclassing) is deferred to v0.40.2.
• --trust-remote-code opt-in surface unchanged — non-SFT trainers (DPO / GRPO / KTO / ORPO / SimPO / IPO / PPO / RewardModel / Pretrain / Embedding / BCO) plus commands/{diff,export,merge,infer,generate}.py still hardcode trust_remote_code=True. Tracked under issue #63 for incremental expansion.
• CI / docs / UX papercuts deferred to v0.40.2 — H2 (data flag aliases), H3 (quickstart --output), N7 (infer/bench HF id auto-download), M4 (data dedup --threshold), M5 (runs --cwd-only), plus the post-v0.40.1 patch-chain items #36 / #50 / #51.

Test count

4720 passed / 3 skipped (was 4656 in v0.40.0 → +64 net new).