Related issue: sgl-project/sglang#18108
Root cause: CUDA Driver Version: 470.82.01 (too old), which leads to failure in version detectection in line 148
|
def _get_cuda_target() -> str: |
|
"""Get the CUDA target architecture flag.""" |
|
if "TVM_FFI_CUDA_ARCH_LIST" in os.environ: |
|
arch_list = os.environ["TVM_FFI_CUDA_ARCH_LIST"].split() # e.g., "8.9 9.0a" |
|
flags = [] |
|
for arch in arch_list: |
|
if len(arch.split(".")) != 2: |
|
raise ValueError(f"Invalid CUDA architecture: {arch}") |
|
major, minor = arch.split(".") |
|
flags.append(f"-gencode=arch=compute_{major}{minor},code=sm_{major}{minor}") |
|
return " ".join(flags) |
|
else: |
|
try: |
|
status = subprocess.run( |
|
args=["nvidia-smi", "--query-gpu=compute_cap", "--format=csv,noheader"], |
|
capture_output=True, |
|
check=True, |
|
) |
|
compute_cap = status.stdout.decode("utf-8").strip().split("\n")[0] |
|
major, minor = compute_cap.split(".") |
|
return f"-gencode=arch=compute_{major}{minor},code=sm_{major}{minor}" |
|
except Exception: |
|
# fallback to a reasonable default |
|
return "-gencode=arch=compute_70,code=sm_70" |
I wonder if users can manually specify the cuda major/minor version when building jit modules.
