[Trainer] support sharding for trainer. #3352
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
ZHUI
commented
Sep 23, 2022
| fused_allreduce_gradients(list(model.parameters()), | ||
| None) | ||
|
|
||
| if self.do_grad_scaling: |
There was a problem hiding this comment.
@haohongxiang 这里的scaler使用体验，请与官方scaler一致。
|
|
||
| self.save_model(output_dir) | ||
|
|
||
| if self.sharding is not None: |
There was a problem hiding this comment.
@haohongxiang 提供接口，rank 0卡，收集参数到cpu
ZHUI
commented
Oct 9, 2022
paddlenlp/trainer/trainer_base.py
Outdated
Comment on lines
683
to
684
| if self.do_grad_scaling: | ||
| self.scaler.minimize(self.optimizer, tr_loss) | ||
| # TODO: fix sharding stage2 stage3 with original scaler useage. |
wawltor
reviewed
Oct 24, 2022
|
|
||
| The value of initial scale_loss for fp16. (default: 32768) | ||
|
|
||
| --sharding |
| |--------------------------------|-------|-------|-------------|------------------|-------------|-------------|------|-------| | ||
| | | mcc | acc | acc | pearson | acc | acc | acc | acc | | ||
| | T5-v1_1-base Paddle | 47.6845 | 94.38 | 84.31 | 87.74 | 88.05 | 85.39 | 90.518 | 65.70 | | ||
| | epoch | 100 | 10 | 100 | 100 | 3 | 3 | 10 | 100 | |
| if self.label2id: | ||
| label = self.label2id[label] | ||
| if pred not in self.label2id: | ||
| pred = 0 |
There was a problem hiding this comment.
这里为什么label为0的时候，pred = 1？
这里的逻辑可以再具体说一下
生成的label不在label list里面，最终预付label 0，这块的情况在encoder不会出现，这里能具体说一下数据指标怎么对齐了？
haohongxiang
previously approved these changes
Oct 25, 2022
ZHUI
commented
Oct 26, 2022
| next_tokens = paddle.argmax(probs, axis=-1).unsqueeze(-1) | ||
| next_scores = paddle.index_sample(probs, next_tokens) | ||
| next_scores = paddle.index_sample(probs.astype("float32"), | ||
| next_tokens) |
There was a problem hiding this comment.
这里的话，index_sample 没有 fp16/bf16 kernel
ZHUI
commented
Oct 26, 2022
| f"{self.dtype} not recognized. `dtype` should be set to either `paddle.float32` or `paddle.float16`" | ||
| ) | ||
| encoder_extended_attention_mask = ( | ||
| 1.0 - encoder_extended_attention_mask) * -1e4 |
ZHUI
commented
Oct 26, 2022
| labels.flatten()) | ||
| loss = loss_fct( | ||
| lm_logits.reshape( | ||
| shape=[-1, lm_logits.shape[-1]]).astype("float32"), |
There was a problem hiding this comment.
CrossEntropyLoss 没有fp16/bf16 kernel
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
PR types
New features
PR changes
APIs
Description
support sharding for trainer.
stage1: 可以支持
stage2：部分支持
stage3：暂不支持