diff --git a/setup.py b/setup.py index 190d65d9fb..0ac53a45b5 100644 --- a/setup.py +++ b/setup.py @@ -1341,10 +1341,10 @@ def get_requirements() -> list[str]: # vllm-flash-attn is built only for CUDA 12.x. # Skip for other versions. continue - if "flashinfer-cubin" in req: - # Not on PyPI since 0.6.14 (only https://flashinfer.ai/whl), so - # it cannot be a wheel dependency; flashinfer falls back to - # fetching cubins at runtime when the package is absent. + if "flashinfer-cubin" in req or "flashinfer-jit-cache" in req: + # Not on PyPI (only https://flashinfer.ai/whl), so they + # cannot be wheel dependencies; flashinfer handles the + # absence of pre-compiled cubins/jit-cache at runtime. continue if "nvidia-cutlass-dsl[cu13]" in req and cuda_major == "12": # [cu13] extra is the default; strip it on CUDA 12 builds.