fix(llamacpp): make the default CUDA image build (slim toolkit + nvcc on PATH)

build_llama_cpp_image's default CUDA image failed to build on constrained image builders,
for two independent reasons found driving a real GPU serve end to end:

1. Full `cuda-toolkit-12-8` metapackage. It drags in cuda-nsight-systems (a GUI profiler:
   Java/GTK/X11, ~2 GB) that a headless build never uses, bloating the image and OOM-killing
   resource-constrained builders mid-install. Replace it with the compile-only subset
   (cuda-nvcc + cudart/driver stubs + cuBLAS/cuRAND dev, ~2.8 GB).

2. `No CMAKE_CUDA_COMPILER could be found`. The image's PATH env (which includes
   $CUDA_HOME/bin, where nvcc lives) is applied *after* the cmake configure/build RUN steps,
   so nvcc is not on PATH during the build. Set PATH inline on the cmake configure and build
   commands.

With both fixes the default `build_llama_cpp_image()` compiles llama-server and the app serves
an OpenAI-compatible /v1 endpoint on a GPU node (verified on a live control plane).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Signed-off-by: Michael Hotan <mike@union.ai>

Author: mhotan

plugins/llamacpp/src/flyteplugins/llamacpp/_constants.py

@@ -5,7 +5,11 @@
 LLAMA_SERVER_BINARY = f"{LLAMA_CPP_INSTALL_DIR}/build/bin/llama-server"
 
 CUDA_HOME = "/usr/local/cuda-12.8"
-CUDA_TOOLKIT_PACKAGE = "cuda-toolkit-12-8"
+# Compile-only CUDA subset for a headless llama.cpp (GGML_CUDA=ON) build: nvcc + the
+# cudart/driver stubs + cuBLAS/cuRAND dev headers. The full `cuda-toolkit-12-8`
+# metapackage also pulls in a GUI profiler and its desktop toolchain (~2 GB) that a
+# headless build never uses, needlessly bloating the image and its build.
+CUDA_TOOLKIT_PACKAGE = "cuda-nvcc-12-8 cuda-cudart-dev-12-8 cuda-driver-dev-12-8 libcublas-dev-12-8 libcurand-dev-12-8"
 # CUDA stubs let the linker resolve libcuda.so on GPU-less build machines; the real
 # driver library is injected by the container runtime on the serving node.
 CUDA_STUB_LIB = f"{CUDA_HOME}/lib64/stubs"

plugins/llamacpp/src/flyteplugins/llamacpp/_image.py

@@ -95,6 +95,9 @@ def build_llama_cpp_image(
             ]
         )
         cmake_configure = (
+            # nvcc lives at $CUDA_HOME/bin; the image's PATH env is applied after these
+            # build RUNs, so set it inline or cmake reports "No CMAKE_CUDA_COMPILER".
+            f"PATH={CUDA_HOME}/bin:$PATH "
             f"LIBRARY_PATH={CUDA_STUB_LIB}:$LIBRARY_PATH "
             f"cmake -S {LLAMA_CPP_INSTALL_DIR} -B {LLAMA_CPP_INSTALL_DIR}/build "
             "-DBUILD_SHARED_LIBS=OFF -DGGML_CUDA=ON "
@@ -127,7 +130,7 @@ def build_llama_cpp_image(
             # fatal EBADENGINE; override it so install proceeds (Node here satisfies
             # the UI's real Vite requirement).
             _run_script(
-                f"PATH={NODE_HOME}/bin:$PATH npm_config_engine_strict=false "
+                f"PATH={NODE_HOME}/bin:{CUDA_HOME}/bin:$PATH npm_config_engine_strict=false "
                 f"{cmake_build_prefix}"
                 f"cmake --build {LLAMA_CPP_INSTALL_DIR}/build --config Release "
                 "-j $(nproc) --target llama-server"