# 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). LIGHT core: the GGUF backend used to ask torch about the
# card; it asks NVML now (codai/backends/gpu_probe.py), so the 5 GB of torch
# that put this image at 11.6 GB is gone. The wheel links the CUDA runtime,
# cuBLAS and NCCL, which NVIDIA ships as wheels at the 12.8 it was built with.
#
# 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.
nvidia-cuda-runtime-cu12==12.8.90
nvidia-cublas-cu12==12.8.4.1
nvidia-nccl-cu12>=2.26
nvidia-ml-py>=12.0.0
sentencepiece>=0.1.99
protobuf>=3.20.0
