v0.8.4

Fixed

• GPU benchmark: vLLM failed to start with curand.h: No such file or directory.
  vLLM's FlashInfer sampler compiles a kernel the first time a model is served.
  The cuRAND, cuBLAS and NVRTC headers it needs were in the image, inside the
  PyTorch wheels, but not where nvcc looks. The image now links the missing
  headers and libraries from the wheels into /usr/local/cuda, so the image
  does not grow. The build also compiles a small check file, so a missing
  header now fails the build instead of the first LLM run on a GPU host.
  Rebuild GPU benchmark images to get the fix. Workaround for an existing
  image: run it with -e VLLM_USE_FLASHINFER_SAMPLER=0.

Changed

• start_container.sh from the GPU kit keeps kernels compiled by vLLM,
  FlashInfer and Triton in $RESULTS_DIR/.cache, so later runs start
  faster. Set CACHE_DIR= to put the cache somewhere else.
• start_container.sh takes extra podman run options in PODMAN_ARGS,
  for example PODMAN_ARGS="-e NCCL_DEBUG=INFO".