# Exact dependency versions the bind2_0 checkpoints were trained/exported with. # Source: local conda env `babylm-smoke` (C:/Users/yulin/.conda/envs/babylm-smoke), # queried via pip on 2026-07-15. # # Python: 3.11.15 # # NOTE on fla (flash-linear-attention): the inlined HF modeling code # (modeling_babylm.py) DOES import it — lazily, inside GDNBlock: # `from fla.layers import GatedDeltaNet` (only executed when arch == "bind2_0") # Since the bind2_0 exports instantiate GDN blocks, fla IS required at runtime # to load/run these checkpoints. Installed from PyPI as release 0.5.1 # (no commit-pin / direct-URL metadata present in the env; pulls fla-core 0.5.1). # # NOTE on triton: the env uses the Windows fork `triton-windows` # (github.com/woct0rdho/triton-windows); on Linux use the matching upstream # `triton` that your torch build requires. # torch build is CUDA 12.6 (`+cu126`); pick the equivalent build for your platform. torch==2.12.1+cu126 transformers==5.13.0 triton-windows==3.7.1.post27 flash-linear-attention==0.5.1 fla-core==0.5.1 safetensors==0.8.0 numpy==2.4.6