我使用 bertscore 作为我的项目的指标，我遇到了一个我无法理解的 GPU 泄漏。我已经能够用这个简单的脚本重现这个问题，查看我的 GPU 上的 GPU 使用情况。
评估版本为0.2.2，python=3.8.
import evaluate
predictions = ['a' for _ in range(30)]
references = ['b' for _ in range(30)]
# nvidia-smi output
# Sat Apr 29 12:25:16 2023
# +-----------------------------------------------------------------------------+
# | NVIDIA-SMI 510.108.03 Driver Version: 510.108.03 CUDA Version: 11.6 |
# |-------------------------------+----------------------+----------------------+
# | GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
# | Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
# | | | MIG M. |
# |===============================+======================+======================|
# | 0 NVIDIA GeForce ... Off | 00000000:07:00.0 Off | N/A |
# | N/A 41C P8 2W / N/A | 1356MiB / 8192MiB | 1% Default |
# | | | N/A |
# +-------------------------------+----------------------+----------------------+
# ...
# nothing for python
bertscore = evaluate.load("bertscore")
scores = bertscore.compute(predictions=predictions, references=references, lang='en', batch_size=64)['f1']
# nvidia-smi output:
# +-----------------------------------------------------------------------------+
# | NVIDIA-SMI 510.108.03 Driver Version: 510.108.03 CUDA Version: 11.6 |
# |-------------------------------+----------------------+----------------------+
# | GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
# | Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
# | | | MIG M. |
# |===============================+======================+======================|
# | 0 NVIDIA GeForce ... Off | 00000000:07:00.0 Off | N/A |
# | N/A 41C P8 2W / N/A | 3106MiB / 8192MiB | 1% Default |
# | | | N/A |
# +-------------------------------+----------------------+----------------------+
# ...
# | 0 N/A N/A 1087985 C python 1751MiB |
# +-----------------------------------------------------------------------------+
del scores
del bertscore
# nvidia-smi output:
# +-----------------------------------------------------------------------------+
# | NVIDIA-SMI 510.108.03 Driver Version: 510.108.03 CUDA Version: 11.6 |
# |-------------------------------+----------------------+----------------------+
# | GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
# | Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
# | | | MIG M. |
# |===============================+======================+======================|
# | 0 NVIDIA GeForce ... Off | 00000000:07:00.0 Off | N/A |
# | N/A 41C P8 2W / N/A | 3106MiB / 8192MiB | 1% Default |
# | | | N/A |
# +-------------------------------+----------------------+----------------------+
# ...
# | 0 N/A N/A 1087985 C python 1751MiB |
# +-----------------------------------------------------------------------------+
有谁知道是什么导致了这个问题？我预计在删除分数和模型后 GPU 利用率会下降。