python/FunASR-XL.git

FUNASR训练

fix: resolve unexpected 'out of memory' issue in multi-GPU setup (#2373)

BienBoy

2025-02-01 c1e365fea09aafda387cac12fdff43d28c598979

refs
用户	BienBoy <92378515+BienBoy@users.noreply.github.com>
	星期六, 二月 1, 2025 23:29 +0800
提交者	GitHub <noreply@github.com>
	星期六, 二月 1, 2025 23:29 +0800
提交	c1e365fea09aafda387cac12fdff43d28c598979
目录	12717845f124e4deaa1a226627cffb9484cc67f4 目录 \| zip \| gz
parent	c4e70144920401a4f5d9b569929b6914af5d324c 查看 \| 对比

fix: resolve unexpected 'out of memory' issue in multi-GPU setup (#2373)

Fixed a bug where calling torch.cuda.empty_cache() caused extra memory usage on 'cuda:0', leading to unexpected 'out of memory' errors in multi-GPU environments.

Reference:
- https://github.com/pytorch/pytorch/issues/25752
- https://github.com/pytorch/pytorch/issues/144025

4个文件已修改

	funasr/auto/auto_model.py	3 ●●●●● 对比 \| 查看 \| 原始文档 \| blame \| 历史
	funasr/bin/train.py	3 ●●●●● 对比 \| 查看 \| 原始文档 \| blame \| 历史
	funasr/bin/train_ds.py	3 ●●●●● 对比 \| 查看 \| 原始文档 \| blame \| 历史
	funasr/models/language_model/rnn/decoders.py	3 ●●●●● 对比 \| 查看 \| 原始文档 \| blame \| 历史