feat/refactor partition strategy #13
Merged
Conversation
This file contains 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
merge upstream/develop
merge upstream develop
…t-0.x-activation-ckpt feat(isp.py): isp communicator support 0.x activation ckpt
feat(model/linear.py): update FeedForward class to internlm2
sunpengsdu
reviewed
Jan 31, 2024
configs/7B_sft.py
Outdated
| 1. size: int, the size of weight parallel. | ||
| 2. overlap: bool, enable/disable all_gather/reduce_scatter communication overlap, defaults to False. | ||
| 3. memory_pool: bool, enable/disable memory pool, defaults to False. | ||
| """ | ||
| parallel = dict( | ||
| zero1=dict(size=8, fsdp=False), |
There was a problem hiding this comment.
既然我们都用了自己的wp了，fsdp这个接口要不要就隐藏了，不在范例中体现了
sunpengsdu
reviewed
Jan 31, 2024
| pipeline=dict(size=1, interleaved_overlap=True), | ||
| sequence_parallel=False, | ||
| weight=dict(size=1, overlap=True, memory_pool=True), | ||
| ) |
There was a problem hiding this comment.
目前的config 7B_sft.py就是带wp的；测例的话可以加一个
There was a problem hiding this comment.
哦，鹏哥的意思加一个wp size大于1的样例
sunpengsdu
reviewed
Jan 31, 2024
| module_shapes: Dict[str, torch.Size] = None | ||
|
|
||
|
|
||
| class MemoryPool: |
There was a problem hiding this comment.
@huangting4201 @mwiacx 如果升级到pytorch lastest版本（假设使用vmm api的版本从rc版本变为正式版）后，memory pool是可以不用了?
sunpengsdu
approved these changes
Feb 1, 2024
| expert_parallel_size (int): Size of expert parallel. | ||
| """ | ||
|
|
||
| def __init__(self, *args, **kwargs): | ||
| super().__init__(*args, **kwargs) | ||
| self.rank_num_per_group = self.tensor_parallel_size * self.pipeline_parallel_size | ||
| self.num_group = self.world_size // self.rank_num_per_group | ||
| self.num_tensor_parallel_group = self.world_size // self.tensor_parallel_size |
|
|
||
| def _get_expert_parallel_ranks(self): | ||
| """ | ||
| Create expert and data parallel groups | ||
| Example: world_size = 8, model_parallel_size = 2, expert_parallel_size = 2 | ||
| Example: world_size = 8, tensor_parallel_size = 2, expert_parallel_size = 2 |
| ) | ||
|
|
||
|
|
||
| class ISPLinear(ColumnParallelLinear): |
There was a problem hiding this comment.
为 ISP 设linear 是为了加 communicator？
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Motivation
Refactor data(sequence), weight, gradients and os partition strategy.