[LLM] fix bug when masked_lm_loss is None in llama modeling.py

[cqulilujia]

Change-Id: I7b6ba9248e61bee24eb463698af26727394f023a

PR types

Bug fixes

PR changes

llama modeling

Description

如issue #8299 所示，llama模型LlamaPretrainingCriterion类中的masked_lm_loss为空时，无法计算loss = paddle.mean，mean不支持输入shape==[0]，本次修复恢复为之前版本的实现

[paddle-bot]

Thanks for your contribution!

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

lilujia seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[cqulilujia]

还有一种改法是：
masked_lm_loss = masked_lm_loss[masked_lm_loss > 0]
if (masked_lm_loss.shape[0] == 0):
loss = paddle.zeros([], dtype=masked_lm_loss.dtype)
loss.stop_gradient = False
else:
loss = paddle.mean(masked_lm_loss)

[Xreki]

还有一种改法是：
masked_lm_loss = masked_lm_loss[masked_lm_loss > 0]
if (masked_lm_loss.shape[0] == 0):
loss = paddle.zeros([], dtype=masked_lm_loss.dtype)
loss.stop_gradient = False
else:
loss = paddle.mean(masked_lm_loss)

这两种改法是等效的吗？能否通过加一些日子把问题数据找出来？

[cqulilujia]

还有一种改法是：
masked_lm_loss = masked_lm_loss[masked_lm_loss > 0]
if (masked_lm_loss.shape[0] == 0):
loss = paddle.zeros([], dtype=masked_lm_loss.dtype)
loss.stop_gradient = False
else:
loss = paddle.mean(masked_lm_loss)

这两种改法是等效的吗？能否通过加一些日子把问题数据找出来？

仅从目前我遇到的实例来看两种结果是一致的，从功能来看第二种更符合原始实现，另外我看这个PR #8342 已经修复了这个问题，用的第二种方式