[Build] Switch default CUDA to 13.0, update CUDA architecture lists, clean up stale build-args (#39878)

Signed-off-by: Shengqi Chen <harry-chen@outlook.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
Shengqi Chen
2026-04-23 15:51:28 +08:00
committed by GitHub
co-authored by Claude Opus 4.6
parent 9c2492e501
commit 3ed5231c6a
10 changed files with 98 additions and 92 deletions
+2 -2
View File
@@ -79,7 +79,7 @@ if TYPE_CHECKING:
VLLM_MEDIA_CONNECTOR: str = "http"
VLLM_MM_HASHER_ALGORITHM: str = "blake3"
VLLM_TARGET_DEVICE: str = "cuda"
VLLM_MAIN_CUDA_VERSION: str = "12.9"
VLLM_MAIN_CUDA_VERSION: str = "13.0"
VLLM_FLOAT32_MATMUL_PRECISION: Literal["highest", "high", "medium"] = "highest"
VLLM_BATCH_INVARIANT: bool = False
MAX_JOBS: str | None = None
@@ -493,7 +493,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
"VLLM_TARGET_DEVICE": lambda: os.getenv("VLLM_TARGET_DEVICE", "cuda").lower(),
# Main CUDA version of vLLM. This follows PyTorch but can be overridden.
"VLLM_MAIN_CUDA_VERSION": lambda: (
os.getenv("VLLM_MAIN_CUDA_VERSION", "").lower() or "12.9"
os.getenv("VLLM_MAIN_CUDA_VERSION", "").lower() or "13.0"
),
# Controls PyTorch float32 matmul precision mode within vLLM workers.
# Valid options mirror torch.set_float32_matmul_precision