# Language-model adviser dependencies. Isolated from the core package: nothing here
# enters requirements.txt or the wheel. CI never loads model weights — the schema/serializer/
# scoring/eval/tool-loop tests all run against the deterministic stub AdviserModel and need only
# the core simulator + gymnasium + pydantic (already core/RL deps).
#
# The training stack (transformers/peft/trl/datasets/accelerate) is required ONLY for the
# `slm.train` smoke SFT (marked slow/manual) and for running a real local model through the
# adviser/eval harness. Install it explicitly when you want to train or run weights.

# --- data generation & scoring (reuses the RL env) ---
gymnasium==1.3.0

# --- schema / config ---
pydantic>=2.0
pyyaml>=6.0

# --- training backend (HF canonical path) ---
# Pin ranges, not exact versions, so the same file works across the wide hardware matrix the plan
# targets (laptop GPU → cluster). Required only for training runs (`slm.train`) and for running a
# real local model; the schema/serializer/scoring/eval/tool-loop paths and CI never need them.
transformers>=5.14.1
peft>=0.19.1
trl>=1.8.0
datasets>=5.0.0
accelerate>=1.14.0
torch>=2.13.0

# --- optional Apple-silicon backend (nice-to-have) ---
# mlx>=0.15
# mlx-lm>=0.16

# --- optional API-backed upper-bound baseline ---
# anthropic>=0.34
