使用 cupy 时内存不足

Question

当我使用 cupy 处理一些大数组时，出现内存不足错误，但是当我检查 nvidia-smi 以查看内存使用情况时，它没有达到我的 GPU 内存的限制，我用的是nvidia geforce RTX 2060，显存是6GB，这是我的代码：

import cupy as cp

mempool = cp.get_default_memory_pool()
print(mempool.used_bytes())              # 0
print(mempool.total_bytes())             # 0

a = cp.random.randint(0, 256, (10980, 10980)).astype(cp.uint8)
a = a.ravel()
print(a.nbytes)                          # 120560400
print(mempool.used_bytes())              # 120560640
print(mempool.total_bytes())             # 602803712
# when I finish create this array, the nvidia-smi shows like this
#+-----------------------------------------------------------------------------+
 | NVIDIA-SMI 430.86       Driver Version: 430.86       CUDA Version: 10.2     |
 |-------------------------------+----------------------+----------------------+
 | GPU  Name            TCC/WDDM | Bus-Id        Disp.A | Volatile Uncorr. ECC |
 | Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
 |===============================+======================+======================|
 |   0  GeForce RTX 2060   WDDM  | 00000000:01:00.0  On |                  N/A |
 | N/A   46C    P8     9W /  N/A |   1280MiB /  6144MiB |      1%      Default |
 +-------------------------------+----------------------+----------------------+

# but then I run this command, and error cames out
s_values, s_idx, s_counts = cp.unique(
    a, return_inverse=True, return_counts=True)
# and the error shows
# cupy.cuda.memory.OutOfMemoryError: out of memory to allocate 964483584 bytes (total 5545867264 bytes)
# the nvidia-smi shows
# +-----------------------------------------------------------------------------+
  | NVIDIA-SMI 430.86       Driver Version: 430.86       CUDA Version: 10.2     |
  |-------------------------------+----------------------+----------------------+
  | GPU  Name            TCC/WDDM | Bus-Id        Disp.A | Volatile Uncorr. ECC |
  | Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
  |===============================+======================+======================|
  |   0  GeForce RTX 2060   WDDM  | 00000000:01:00.0  On |                  N/A |
  | N/A   45C    P8     9W /  N/A |   5075MiB /  6144MiB |      3%      Default |
  +-------------------------------+----------------------+----------------------+

好像有足够的space可以使用，为什么会出现这个错误，是因为我的GPU没有足够的内存，还是因为我的代码错误或者我没有分配记忆正确。

Answer 1

964,483,584 不是比您的 mempool.total_bytes() 602,803,712 大吗？

如评论中所述，您可以分批进行，而不是一次完成整个计算。

Answer 2

您可以使用 dask 来执行与它代表您执行并行化相同的操作，而且即使数据不适合 RAM，您也永远不会运行内存不足。我附上 link 作者自己提供了如何做的解释。

from dask.distributed import  Client,LocalCluster
import dask.array as da
import numpy as np

cluster = LocalCluster() #using multiple CPUs in the machine/cluster
client = Client(cluster)
client
rs = da.random.RandomState(RandomState=np.random.RandomState)
x = rs.random((100000,40000),chunks=(10000,400)) #29.80GB of ndarray
x #just ensure that the chunk size is small #30.52MB chunk
da.exp(x).mean().compute() #do not try to return ndarray with element-wise transformation, instead always try to get the reduced form. 
da.exp(x) # Do not run this line as it will lead to

在最后一行，dask 尝试将输出保存在内存中。由于输出的数量级为 29+GB，因此您将运行内存不足。 Youtube link for the explanation of the above code by the author of dask

使用 cupy 时内存不足

out of memory when using cupy

python

cupy