# Text generation (HuggingFace safetensors) — a coderai pod that serves LLMs.
#
# This is the profile that makes a LOCAL text LoRA usable on RunPod: vLLM and
# llama.cpp pods can only load adapters they can resolve themselves (a HF repo
# id), while a coderai pod is sent the adapter and applies it with PEFT.
transformers>=4.35.0
accelerate>=0.24.0
peft>=0.10.0
# QLoRA: the adapter is ordinary, but the BASE must load 4-bit the way it was
# trained, which is what bitsandbytes does.
bitsandbytes>=0.43.0
sentencepiece>=0.1.99
protobuf>=3.20.0
