The vllm.py tag matcher in package_plugins maps CPU, CUDA, and TPU variants all to accelerator="cuda" (line 74-75). This means the CPU variant looks for CUDA-tagged releases and won't match CPU-specific tags like v0.14.1+rhai1.cpu.

The relevant code in package_plugins/vllm.py:

if variant.startswith(("cuda", "cpu", "tpu")):
    accelerator = "cuda"  # CUDA, CPU, and TPU use cuda tags

This needs to be updated so that the CPU variant either has its own accelerator value or the matcher recognizes .cpu suffixed tags. Without this, the builder cannot resolve vllm versions tagged with .cpu suffixes (e.g. v0.14.1+rhai1.cpu) which are used for CPU-specific midstream releases.

Related to INFERENG-4612 and AIPCC-10275.