# llama.cpp as a coderai pod: llama-cpp-python built with CUDA for every pod
# GPU generation, driving GGUF models through coderai's own CUDA backend —
# the same kernels as the upstream ghcr.io/ggml-org/llama.cpp image, plus what
# a coderai pod gives (TLS on the direct path, /boot, seeding, local LoRAs,
# the Tasks page). FULL core: the GGUF backend asks torch about the card.
#
# The wheel itself is compiled in Dockerfile.capability-llama's toolchain
# stage (see the `llama.dockerfile` marker); this file lists what else the
# text path needs.
sentencepiece>=0.1.99
protobuf>=3.20.0
