Use CUDA 12.4 as default for release and nightly wheels (#12098)

This commit is contained in:
Michael Goin
2025-02-26 19:06:37 -08:00
committed by GitHub
parent a31614e386
commit ca377cf1b9
4 changed files with 25 additions and 9 deletions
+3 -4
View File
@@ -54,7 +54,7 @@ elif (sys.platform.startswith("linux") and torch.version.cuda is None
# fallback to cpu
VLLM_TARGET_DEVICE = "cpu"
MAIN_CUDA_VERSION = "12.1"
MAIN_CUDA_VERSION = "12.4"
def is_sccache_available() -> bool:
@@ -571,9 +571,8 @@ def get_requirements() -> List[str]:
cuda_major, cuda_minor = torch.version.cuda.split(".")
modified_requirements = []
for req in requirements:
if ("vllm-flash-attn" in req
and not (cuda_major == "12" and cuda_minor == "1")):
# vllm-flash-attn is built only for CUDA 12.1.
if ("vllm-flash-attn" in req and cuda_major != "12"):
# vllm-flash-attn is built only for CUDA 12.x.
# Skip for other versions.
continue
modified_requirements.append(req)