MakazhanAlpamys/Soup v0.53.3

v0.53.3 — grpo_fp16 routing + vision-GRPO VLM base probe

on GitHub

What's New

Two surgical fixes from the v0.50.0 GRPO Plus deferred-stub family. The four larger items (stability callback, variant losses, PRM trainer, multi-objective preference live combine) are scope-deferred to v0.53.4 — each warrants a focused release per the v0.40.x stub-then-live cadence.

• grpo_fp16: true now wires through GRPOTrainerWrapper. New _build_precision_kwargs(self) -> dict[str, bool] returns the right {fp16, bf16} per (device, grpo_fp16) matrix: non-CUDA (CPU/MPS/XPU) → both False (HF Trainer's mixed-precision kwargs are CUDA-specific), CUDA + grpo_fp16=True → fp16=True, bf16=False (unsloth parity), default CUDA → bf16 (legacy v0.50.0 path).
• Mutex cross-validator. SoupConfig._validate_grpo_fp16_amp_exclusive rejects the silent footgun grpo_fp16=True + auto_mixed_precision=True. The validator short-circuits when task != 'grpo' so the v0.50.0 stability task-gate diagnosis fires first, keeping the most actionable error in front of the user regardless of validator execution order.
• Vision-GRPO base-model probe. New KNOWN_VLM_REGEX covers 10 VLM families (Qwen2-VL / Qwen2.5-VL / QVQ / Pixtral / InternVL / Llama-3.2-Vision / LLaVA / MiniCPM-V / Idefics / ShareGPT4V / Fuyu) with word-boundary anchors that reject substring noise like "my-pixtralish". is_known_vlm_base returns False (never raises) on bad input. validate_vision_grpo_compat accepts a new optional base kwarg threaded from SoupConfig._validate_vision_grpo.
• Friendly schema rejection instead of a cryptic runtime AttributeError: 'vision_tower': a YAML pairing vision_grpo: true with a non-VLM checkpoint now fails at config-load with a message naming the expected families. Echoed base value is truncated to 64 chars before serialisation, mirroring the v0.34.0 crash.py redaction policy.
• +37 net new tests (7842 → 7879) in tests/test_v0533.py. Four review agents (python / code / security / tdd) ran; every HIGH / MEDIUM / LOW finding fixed.

Install / Upgrade

pip install --upgrade soup-cli

Security

• Vision-GRPO base probe rejects unknown VLMs at schema load (defense-in-depth — keeps the legacy attribute-error footgun off the runtime path).
• 64-char truncation on adversarial / oversized base values in the rejection error message — prevents log bloat and unredacted user input from surfacing in operator-facing tracebacks.
• is_known_vlm_base is defensive — returns False (never raises) on bool / non-string / empty / null-byte / >_MAX_BASE_NAME_LEN=512, mirroring v0.39.0 / v0.44.0 / v0.49.0 model-detection policy.

Known Limitations

• Scope-deferred — #127 GRPOStabilityCallback (7 stability knobs), #123 6 GRPO variant loss kernels (gspo / dapo / dr_grpo / bnpo / two_sided / rft), #126 PRMTrainerWrapper, and #68 PreferenceTrainerWrapper true weighted-loss combine all move to v0.53.4. Each requires deep TRL trainer subclassing and warrants its own focused release.
• VLM allowlist is static name-regex — a legitimate VLM published under an org whose checkpoint name lacks any of the 10 tokens is rejected at schema-load; users must omit vision_grpo: true until a future release adds a runtime model.config.vision_config probe.
• _build_precision_kwargs is GRPO-only — other RL trainers (PPO / RewardModel) follow their existing mixed-precision conventions; the grpo_fp16 flag is GRPO-gated by the v0.50.0 stability task-gate.