# Pinned to what runs on the dev box (RTX 5090, CUDA 12.8). Versions are
# observed-baselines, not strict floors — bump as upstream/drivers move.

# core
torch>=2.9
# Hard-pin: the upstream modeling_qwen3.py (commit df86ca9) reads
# `config.rope_theta` directly; transformers 5.x moved that into a
# `rope_parameters` dict, silently breaking attention freqs at sample time.
transformers==4.57.1
safetensors>=0.7

# HF model load (NEOChatModel uses trust_remote_code)
huggingface_hub>=0.26
sentencepiece
einops
pillow
numpy

# low-VRAM training
bitsandbytes>=0.45    # 8-bit / 4-bit base + paged AdamW8bit
peft>=0.13            # not strictly required (we have our own LoRA), pulled in for utility helpers

# config + dataset
pyyaml
pydantic>=2.0
pyarrow>=14         # ArrowT2IDataset (parquet shards) for 1M+ scale datasets

# tests
pytest>=7