mirror of
https://github.com/vllm-project/vllm.git
synced 2025-10-20 14:53:52 +08:00
[Bugfix] Mapping physical device indices for e2e test utils (#8290)
This commit is contained in:
@ -356,12 +356,23 @@ def error_on_warning():
|
||||
yield
|
||||
|
||||
|
||||
def get_physical_device_indices(devices):
|
||||
visible_devices = os.environ.get("CUDA_VISIBLE_DEVICES")
|
||||
if visible_devices is None:
|
||||
return devices
|
||||
|
||||
visible_indices = [int(x) for x in visible_devices.split(",")]
|
||||
index_mapping = {i: physical for i, physical in enumerate(visible_indices)}
|
||||
return [index_mapping[i] for i in devices if i in index_mapping]
|
||||
|
||||
|
||||
@_nvml()
|
||||
def wait_for_gpu_memory_to_clear(devices: List[int],
|
||||
threshold_bytes: int,
|
||||
timeout_s: float = 120) -> None:
|
||||
# Use nvml instead of pytorch to reduce measurement error from torch cuda
|
||||
# context.
|
||||
devices = get_physical_device_indices(devices)
|
||||
start_time = time.time()
|
||||
while True:
|
||||
output: Dict[int, str] = {}
|
||||
|
Reference in New Issue
Block a user