7fa972be6a
version: sunnypilot v2025.003.000 (dev) date: 2026-02-09T02:04:38 master commit: 254f55ac15a40343d7255f2f098de3442e0c4a6f
15 lines
472 B
Python
15 lines
472 B
Python
from tinygrad import Tensor, Device, GlobalCounters
|
|
from tinygrad.helpers import Timing
|
|
|
|
N = 512
|
|
GPUS = 5
|
|
ds = tuple([f"{Device.DEFAULT}:{i+1}" for i in range(GPUS)])
|
|
t = [Tensor.ones(N, N, N, device=d).contiguous().realize() for d in ds]
|
|
|
|
for _ in range(10):
|
|
GlobalCounters.reset()
|
|
with Timing():
|
|
for ti in t:
|
|
ti.to_(ds[(ds.index(ti.device)+1+len(ds))%len(ds)])
|
|
# ti.to_(ds[(ds.index(ti.device)-1+len(ds))%len(ds)]) # reversed order
|
|
ti.realize() |