ColossalAI/docs/source/zh-Hans/features/zero_with_chunk.md

211 lines
9.7 KiB
Markdown
Raw Normal View History

# 基于Chunk内存管理的零冗余优化器 (ZeRO)
作者: [Hongxiu Liu](https://github.com/ver217), [Jiarui Fang](https://github.com/feifeibear), [Zijian Ye](https://github.com/ZijianYY)
**前置教程:**
- [booster使用](../basics/booster_api.md)
**示例代码**
- [Train GPT with Colossal-AI](https://github.com/hpcaitech/ColossalAI/tree/main/examples/language/gpt)
**相关论文**
- [ZeRO: Memory Optimizations Toward Training Trillion Parameter Models](https://arxiv.org/abs/1910.02054)
- [ZeRO-Offload: Democratizing Billion-Scale Model Training](https://arxiv.org/abs/2101.06840)
- [ZeRO-Infinity: Breaking the GPU Memory Wall for Extreme Scale Deep Learning](https://arxiv.org/abs/2104.07857)
- [DeepSpeed: System Optimizations Enable Training Deep Learning Models with Over 100 Billion Parameters](https://dl.acm.org/doi/10.1145/3394486.3406703)
- [PatrickStar: Parallel Training of Pre-trained Models via Chunk-based Memory Management](https://arxiv.org/abs/2108.05818)
## 引言
零冗余优化器 (ZeRO) 通过对三个模型状态(优化器状态、梯度和参数)进行划分而不是复制他们,消除了数据并行进程中的内存冗余。该方法与传统的数据并行相比,内存效率得到了极大的提高,而计算粒度和通信效率得到了保留。
1. **分片优化器状态**: 优化器状态 (如 [Adam optimizer](https://arxiv.org/abs/1412.6980), 32位的权重,
以及一二阶动量估计) 被划分到各个进程中, 因此每个进程只更新其分区。
2. **分片梯度**: 在梯度在数据并行进程组内进行 reduction 后, 梯度张量也被划分,这样每个进程只存储与其划分的优化器状态对应的梯度。 注意, Colossal-AI 将梯度转换为 FP32 格式以参与更新参数。
3. **分片参数**: 16位的模型参数被划分到一个数据并行组的进程中。
4. **[Gemini](../advanced_tutorials/meet_gemini.md)**: 对于参数、梯度、优化器状态的动态异构内存空间管理器。
此外我们还将介绍基于Chunk内存管理的零冗余优化器。
在使用零冗余优化器 (ZeRO)时我们通过切分参数的方式对模型进行分布式存储这种方法的优点是每个节点的内存负载是完全均衡的。但是这种方式有很多缺点。首先通信时需要申请一块临时内存用来通信通信完毕释放这回导致存在内存碎片化的问题。其次以Tensor为粒度进行通信会导致网络带宽无法充分利用。通常来说传输的消息长度越长带宽利用率越高。
利用ColossalAI v0.1.8引入了Chunk机制我们可以提升ZeRO的性能。我们将运算顺序上连续的一组参数存入一个Chunk中Chunk即一段连续的内存空间每个Chunk的大小相同。Chunk方式组织内存可以保证PCI-e和GPU-GPU之间网络带宽的高效利用减小了通信次数同时避免潜在的内存碎片。
在v0.1.8之前ZeRO在进行参数聚合时通信成本较高如果一个参数在连续的几次计算中被使用多次即会发生多次通信效率较低。这种情况在使用Checkpoint时非常常见参数在计算backward时会重计算一遍forward。这种情况下ZeRO的效率便不高。
以GPT为例其Checkpoint会应用在每一个GPT Block上每一个GPT Block包含一个Self-Attention层和MLP层。在计算Backward时会依次计算Self-Attention层、MLP层的forward然后依次计算MLP层、Self-Attention层的backward。如使用Chunk机制我们将Self-Attention层和MLP层放在同一个Chunk中在每个GPT Block的backward的中便无需再通信。
除此之外由于小Tensor的通信、内存移动没法完全利用NVLINK、PCIE带宽而且每次通信、内存移动都有kernel launch的开销。使用了Chunk之后可以把多次小Tensor的通信、内存移动变为一次大Tensor的通信、内存移动既提高了带宽利用也减小了kernel launch的开销。
我们提供了轻量级的Chunk搜索机制帮助用户自动找到内存碎片最小的Chunk尺寸。
## 使用
### GeminiDDP
我们将运用`GeminiDDP`的方式来使用基于Chunk内存管理的ZeRO。这是我们新包装的torch.Module ,它使用 ZeRO-DP 和 Gemini其中ZeRO 用于并行Gemini 用于内存管理。
[gemini] improve compatibility and add static placement policy (#4479) * [gemini] remove distributed-related part from colotensor (#4379) * [gemini] remove process group dependency * [gemini] remove tp part from colo tensor * [gemini] patch inplace op * [gemini] fix param op hook and update tests * [test] remove useless tests * [test] remove useless tests * [misc] fix requirements * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [misc] update requirements * [gemini] refactor gemini optimizer and gemini ddp (#4398) * [gemini] update optimizer interface * [gemini] renaming gemini optimizer * [gemini] refactor gemini ddp class * [example] update gemini related example * [example] update gemini related example * [plugin] fix gemini plugin args * [test] update gemini ckpt tests * [gemini] fix checkpoint io * [example] fix opt example requirements * [example] fix opt example * [example] fix opt example * [example] fix opt example * [gemini] add static placement policy (#4443) * [gemini] add static placement policy * [gemini] fix param offload * [test] update gemini tests * [plugin] update gemini plugin * [plugin] update gemini plugin docstr * [misc] fix flash attn requirement * [test] fix gemini checkpoint io test * [example] update resnet example result (#4457) * [example] update bert example result (#4458) * [doc] update gemini doc (#4468) * [example] update gemini related examples (#4473) * [example] update gpt example * [example] update dreambooth example * [example] update vit * [example] update opt * [example] update palm * [example] update vit and opt benchmark * [hotfix] fix bert in model zoo (#4480) * [hotfix] fix bert in model zoo * [test] remove chatglm gemini test * [test] remove sam gemini test * [test] remove vit gemini test * [hotfix] fix opt tutorial example (#4497) * [hotfix] fix opt tutorial example * [hotfix] fix opt tutorial example
2023-08-24 01:29:25 +00:00
Gemini支持惰性初始化, 它可以节省多卡初始化大模型时的显存使用.
[gemini] improve compatibility and add static placement policy (#4479) * [gemini] remove distributed-related part from colotensor (#4379) * [gemini] remove process group dependency * [gemini] remove tp part from colo tensor * [gemini] patch inplace op * [gemini] fix param op hook and update tests * [test] remove useless tests * [test] remove useless tests * [misc] fix requirements * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [misc] update requirements * [gemini] refactor gemini optimizer and gemini ddp (#4398) * [gemini] update optimizer interface * [gemini] renaming gemini optimizer * [gemini] refactor gemini ddp class * [example] update gemini related example * [example] update gemini related example * [plugin] fix gemini plugin args * [test] update gemini ckpt tests * [gemini] fix checkpoint io * [example] fix opt example requirements * [example] fix opt example * [example] fix opt example * [example] fix opt example * [gemini] add static placement policy (#4443) * [gemini] add static placement policy * [gemini] fix param offload * [test] update gemini tests * [plugin] update gemini plugin * [plugin] update gemini plugin docstr * [misc] fix flash attn requirement * [test] fix gemini checkpoint io test * [example] update resnet example result (#4457) * [example] update bert example result (#4458) * [doc] update gemini doc (#4468) * [example] update gemini related examples (#4473) * [example] update gpt example * [example] update dreambooth example * [example] update vit * [example] update opt * [example] update palm * [example] update vit and opt benchmark * [hotfix] fix bert in model zoo (#4480) * [hotfix] fix bert in model zoo * [test] remove chatglm gemini test * [test] remove sam gemini test * [test] remove vit gemini test * [hotfix] fix opt tutorial example (#4497) * [hotfix] fix opt tutorial example * [hotfix] fix opt tutorial example
2023-08-24 01:29:25 +00:00
如果你的模型有 `N` billion 个参数,你的 GPU 内存为 `M` GB, 当 `4N >= M` 时,我们推荐使用 LazyInitContext。否则LazyInitContext 是可选的。
<!--- doc-test-ignore-start -->
```python
[gemini] improve compatibility and add static placement policy (#4479) * [gemini] remove distributed-related part from colotensor (#4379) * [gemini] remove process group dependency * [gemini] remove tp part from colo tensor * [gemini] patch inplace op * [gemini] fix param op hook and update tests * [test] remove useless tests * [test] remove useless tests * [misc] fix requirements * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [misc] update requirements * [gemini] refactor gemini optimizer and gemini ddp (#4398) * [gemini] update optimizer interface * [gemini] renaming gemini optimizer * [gemini] refactor gemini ddp class * [example] update gemini related example * [example] update gemini related example * [plugin] fix gemini plugin args * [test] update gemini ckpt tests * [gemini] fix checkpoint io * [example] fix opt example requirements * [example] fix opt example * [example] fix opt example * [example] fix opt example * [gemini] add static placement policy (#4443) * [gemini] add static placement policy * [gemini] fix param offload * [test] update gemini tests * [plugin] update gemini plugin * [plugin] update gemini plugin docstr * [misc] fix flash attn requirement * [test] fix gemini checkpoint io test * [example] update resnet example result (#4457) * [example] update bert example result (#4458) * [doc] update gemini doc (#4468) * [example] update gemini related examples (#4473) * [example] update gpt example * [example] update dreambooth example * [example] update vit * [example] update opt * [example] update palm * [example] update vit and opt benchmark * [hotfix] fix bert in model zoo (#4480) * [hotfix] fix bert in model zoo * [test] remove chatglm gemini test * [test] remove sam gemini test * [test] remove vit gemini test * [hotfix] fix opt tutorial example (#4497) * [hotfix] fix opt tutorial example * [hotfix] fix opt tutorial example
2023-08-24 01:29:25 +00:00
with LazyInitContext(default_device=torch.device('cuda')):
model = gpt2_medium(checkpoint=True)
```
[gemini] improve compatibility and add static placement policy (#4479) * [gemini] remove distributed-related part from colotensor (#4379) * [gemini] remove process group dependency * [gemini] remove tp part from colo tensor * [gemini] patch inplace op * [gemini] fix param op hook and update tests * [test] remove useless tests * [test] remove useless tests * [misc] fix requirements * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [misc] update requirements * [gemini] refactor gemini optimizer and gemini ddp (#4398) * [gemini] update optimizer interface * [gemini] renaming gemini optimizer * [gemini] refactor gemini ddp class * [example] update gemini related example * [example] update gemini related example * [plugin] fix gemini plugin args * [test] update gemini ckpt tests * [gemini] fix checkpoint io * [example] fix opt example requirements * [example] fix opt example * [example] fix opt example * [example] fix opt example * [gemini] add static placement policy (#4443) * [gemini] add static placement policy * [gemini] fix param offload * [test] update gemini tests * [plugin] update gemini plugin * [plugin] update gemini plugin docstr * [misc] fix flash attn requirement * [test] fix gemini checkpoint io test * [example] update resnet example result (#4457) * [example] update bert example result (#4458) * [doc] update gemini doc (#4468) * [example] update gemini related examples (#4473) * [example] update gpt example * [example] update dreambooth example * [example] update vit * [example] update opt * [example] update palm * [example] update vit and opt benchmark * [hotfix] fix bert in model zoo (#4480) * [hotfix] fix bert in model zoo * [test] remove chatglm gemini test * [test] remove sam gemini test * [test] remove vit gemini test * [hotfix] fix opt tutorial example (#4497) * [hotfix] fix opt tutorial example * [hotfix] fix opt tutorial example
2023-08-24 01:29:25 +00:00
<!--- doc-test-ignore-end -->
我们提供了 `Booster` API它用户友好。我们推荐你使用 `Booster` API。如果您仍然想使用底层 API您可以继续阅读本节其他内容。
[gemini] improve compatibility and add static placement policy (#4479) * [gemini] remove distributed-related part from colotensor (#4379) * [gemini] remove process group dependency * [gemini] remove tp part from colo tensor * [gemini] patch inplace op * [gemini] fix param op hook and update tests * [test] remove useless tests * [test] remove useless tests * [misc] fix requirements * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [misc] update requirements * [gemini] refactor gemini optimizer and gemini ddp (#4398) * [gemini] update optimizer interface * [gemini] renaming gemini optimizer * [gemini] refactor gemini ddp class * [example] update gemini related example * [example] update gemini related example * [plugin] fix gemini plugin args * [test] update gemini ckpt tests * [gemini] fix checkpoint io * [example] fix opt example requirements * [example] fix opt example * [example] fix opt example * [example] fix opt example * [gemini] add static placement policy (#4443) * [gemini] add static placement policy * [gemini] fix param offload * [test] update gemini tests * [plugin] update gemini plugin * [plugin] update gemini plugin docstr * [misc] fix flash attn requirement * [test] fix gemini checkpoint io test * [example] update resnet example result (#4457) * [example] update bert example result (#4458) * [doc] update gemini doc (#4468) * [example] update gemini related examples (#4473) * [example] update gpt example * [example] update dreambooth example * [example] update vit * [example] update opt * [example] update palm * [example] update vit and opt benchmark * [hotfix] fix bert in model zoo (#4480) * [hotfix] fix bert in model zoo * [test] remove chatglm gemini test * [test] remove sam gemini test * [test] remove vit gemini test * [hotfix] fix opt tutorial example (#4497) * [hotfix] fix opt tutorial example * [hotfix] fix opt tutorial example
2023-08-24 01:29:25 +00:00
使用 `GeminiDDP` 包装模型。
[gemini] improve compatibility and add static placement policy (#4479) * [gemini] remove distributed-related part from colotensor (#4379) * [gemini] remove process group dependency * [gemini] remove tp part from colo tensor * [gemini] patch inplace op * [gemini] fix param op hook and update tests * [test] remove useless tests * [test] remove useless tests * [misc] fix requirements * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [misc] update requirements * [gemini] refactor gemini optimizer and gemini ddp (#4398) * [gemini] update optimizer interface * [gemini] renaming gemini optimizer * [gemini] refactor gemini ddp class * [example] update gemini related example * [example] update gemini related example * [plugin] fix gemini plugin args * [test] update gemini ckpt tests * [gemini] fix checkpoint io * [example] fix opt example requirements * [example] fix opt example * [example] fix opt example * [example] fix opt example * [gemini] add static placement policy (#4443) * [gemini] add static placement policy * [gemini] fix param offload * [test] update gemini tests * [plugin] update gemini plugin * [plugin] update gemini plugin docstr * [misc] fix flash attn requirement * [test] fix gemini checkpoint io test * [example] update resnet example result (#4457) * [example] update bert example result (#4458) * [doc] update gemini doc (#4468) * [example] update gemini related examples (#4473) * [example] update gpt example * [example] update dreambooth example * [example] update vit * [example] update opt * [example] update palm * [example] update vit and opt benchmark * [hotfix] fix bert in model zoo (#4480) * [hotfix] fix bert in model zoo * [test] remove chatglm gemini test * [test] remove sam gemini test * [test] remove vit gemini test * [hotfix] fix opt tutorial example (#4497) * [hotfix] fix opt tutorial example * [hotfix] fix opt tutorial example
2023-08-24 01:29:25 +00:00
<!--- doc-test-ignore-start -->
```python
[gemini] improve compatibility and add static placement policy (#4479) * [gemini] remove distributed-related part from colotensor (#4379) * [gemini] remove process group dependency * [gemini] remove tp part from colo tensor * [gemini] patch inplace op * [gemini] fix param op hook and update tests * [test] remove useless tests * [test] remove useless tests * [misc] fix requirements * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [misc] update requirements * [gemini] refactor gemini optimizer and gemini ddp (#4398) * [gemini] update optimizer interface * [gemini] renaming gemini optimizer * [gemini] refactor gemini ddp class * [example] update gemini related example * [example] update gemini related example * [plugin] fix gemini plugin args * [test] update gemini ckpt tests * [gemini] fix checkpoint io * [example] fix opt example requirements * [example] fix opt example * [example] fix opt example * [example] fix opt example * [gemini] add static placement policy (#4443) * [gemini] add static placement policy * [gemini] fix param offload * [test] update gemini tests * [plugin] update gemini plugin * [plugin] update gemini plugin docstr * [misc] fix flash attn requirement * [test] fix gemini checkpoint io test * [example] update resnet example result (#4457) * [example] update bert example result (#4458) * [doc] update gemini doc (#4468) * [example] update gemini related examples (#4473) * [example] update gpt example * [example] update dreambooth example * [example] update vit * [example] update opt * [example] update palm * [example] update vit and opt benchmark * [hotfix] fix bert in model zoo (#4480) * [hotfix] fix bert in model zoo * [test] remove chatglm gemini test * [test] remove sam gemini test * [test] remove vit gemini test * [hotfix] fix opt tutorial example (#4497) * [hotfix] fix opt tutorial example * [hotfix] fix opt tutorial example
2023-08-24 01:29:25 +00:00
model = GeminiDDP(model, hidden_dim=hidden_dim, min_chunk_size_m=min_chunk_size_m)
```
[gemini] improve compatibility and add static placement policy (#4479) * [gemini] remove distributed-related part from colotensor (#4379) * [gemini] remove process group dependency * [gemini] remove tp part from colo tensor * [gemini] patch inplace op * [gemini] fix param op hook and update tests * [test] remove useless tests * [test] remove useless tests * [misc] fix requirements * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [misc] update requirements * [gemini] refactor gemini optimizer and gemini ddp (#4398) * [gemini] update optimizer interface * [gemini] renaming gemini optimizer * [gemini] refactor gemini ddp class * [example] update gemini related example * [example] update gemini related example * [plugin] fix gemini plugin args * [test] update gemini ckpt tests * [gemini] fix checkpoint io * [example] fix opt example requirements * [example] fix opt example * [example] fix opt example * [example] fix opt example * [gemini] add static placement policy (#4443) * [gemini] add static placement policy * [gemini] fix param offload * [test] update gemini tests * [plugin] update gemini plugin * [plugin] update gemini plugin docstr * [misc] fix flash attn requirement * [test] fix gemini checkpoint io test * [example] update resnet example result (#4457) * [example] update bert example result (#4458) * [doc] update gemini doc (#4468) * [example] update gemini related examples (#4473) * [example] update gpt example * [example] update dreambooth example * [example] update vit * [example] update opt * [example] update palm * [example] update vit and opt benchmark * [hotfix] fix bert in model zoo (#4480) * [hotfix] fix bert in model zoo * [test] remove chatglm gemini test * [test] remove sam gemini test * [test] remove vit gemini test * [hotfix] fix opt tutorial example (#4497) * [hotfix] fix opt tutorial example * [hotfix] fix opt tutorial example
2023-08-24 01:29:25 +00:00
<!--- doc-test-ignore-end -->
`hidden dim`是DNN的隐藏维度。用户可以提供这个参数来加快搜索速度。如果用户在训练前不知道这个参数也可以。 我们将使用默认值 1024。`min_chunk_size_m`是以兆2^20为单位的最小块大小。如果参数的总大小仍然小于最小块大小则所有参数将被压缩为一个小块。
初始化优化器。
[gemini] improve compatibility and add static placement policy (#4479) * [gemini] remove distributed-related part from colotensor (#4379) * [gemini] remove process group dependency * [gemini] remove tp part from colo tensor * [gemini] patch inplace op * [gemini] fix param op hook and update tests * [test] remove useless tests * [test] remove useless tests * [misc] fix requirements * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [misc] update requirements * [gemini] refactor gemini optimizer and gemini ddp (#4398) * [gemini] update optimizer interface * [gemini] renaming gemini optimizer * [gemini] refactor gemini ddp class * [example] update gemini related example * [example] update gemini related example * [plugin] fix gemini plugin args * [test] update gemini ckpt tests * [gemini] fix checkpoint io * [example] fix opt example requirements * [example] fix opt example * [example] fix opt example * [example] fix opt example * [gemini] add static placement policy (#4443) * [gemini] add static placement policy * [gemini] fix param offload * [test] update gemini tests * [plugin] update gemini plugin * [plugin] update gemini plugin docstr * [misc] fix flash attn requirement * [test] fix gemini checkpoint io test * [example] update resnet example result (#4457) * [example] update bert example result (#4458) * [doc] update gemini doc (#4468) * [example] update gemini related examples (#4473) * [example] update gpt example * [example] update dreambooth example * [example] update vit * [example] update opt * [example] update palm * [example] update vit and opt benchmark * [hotfix] fix bert in model zoo (#4480) * [hotfix] fix bert in model zoo * [test] remove chatglm gemini test * [test] remove sam gemini test * [test] remove vit gemini test * [hotfix] fix opt tutorial example (#4497) * [hotfix] fix opt tutorial example * [hotfix] fix opt tutorial example
2023-08-24 01:29:25 +00:00
<!--- doc-test-ignore-start -->
```python
optimizer = GeminiAdamOptimizer(model, lr=1e-3, initial_scale=2**5)
```
[gemini] improve compatibility and add static placement policy (#4479) * [gemini] remove distributed-related part from colotensor (#4379) * [gemini] remove process group dependency * [gemini] remove tp part from colo tensor * [gemini] patch inplace op * [gemini] fix param op hook and update tests * [test] remove useless tests * [test] remove useless tests * [misc] fix requirements * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [misc] update requirements * [gemini] refactor gemini optimizer and gemini ddp (#4398) * [gemini] update optimizer interface * [gemini] renaming gemini optimizer * [gemini] refactor gemini ddp class * [example] update gemini related example * [example] update gemini related example * [plugin] fix gemini plugin args * [test] update gemini ckpt tests * [gemini] fix checkpoint io * [example] fix opt example requirements * [example] fix opt example * [example] fix opt example * [example] fix opt example * [gemini] add static placement policy (#4443) * [gemini] add static placement policy * [gemini] fix param offload * [test] update gemini tests * [plugin] update gemini plugin * [plugin] update gemini plugin docstr * [misc] fix flash attn requirement * [test] fix gemini checkpoint io test * [example] update resnet example result (#4457) * [example] update bert example result (#4458) * [doc] update gemini doc (#4468) * [example] update gemini related examples (#4473) * [example] update gpt example * [example] update dreambooth example * [example] update vit * [example] update opt * [example] update palm * [example] update vit and opt benchmark * [hotfix] fix bert in model zoo (#4480) * [hotfix] fix bert in model zoo * [test] remove chatglm gemini test * [test] remove sam gemini test * [test] remove vit gemini test * [hotfix] fix opt tutorial example (#4497) * [hotfix] fix opt tutorial example * [hotfix] fix opt tutorial example
2023-08-24 01:29:25 +00:00
<!--- doc-test-ignore-end -->
[gemini] improve compatibility and add static placement policy (#4479) * [gemini] remove distributed-related part from colotensor (#4379) * [gemini] remove process group dependency * [gemini] remove tp part from colo tensor * [gemini] patch inplace op * [gemini] fix param op hook and update tests * [test] remove useless tests * [test] remove useless tests * [misc] fix requirements * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [misc] update requirements * [gemini] refactor gemini optimizer and gemini ddp (#4398) * [gemini] update optimizer interface * [gemini] renaming gemini optimizer * [gemini] refactor gemini ddp class * [example] update gemini related example * [example] update gemini related example * [plugin] fix gemini plugin args * [test] update gemini ckpt tests * [gemini] fix checkpoint io * [example] fix opt example requirements * [example] fix opt example * [example] fix opt example * [example] fix opt example * [gemini] add static placement policy (#4443) * [gemini] add static placement policy * [gemini] fix param offload * [test] update gemini tests * [plugin] update gemini plugin * [plugin] update gemini plugin docstr * [misc] fix flash attn requirement * [test] fix gemini checkpoint io test * [example] update resnet example result (#4457) * [example] update bert example result (#4458) * [doc] update gemini doc (#4468) * [example] update gemini related examples (#4473) * [example] update gpt example * [example] update dreambooth example * [example] update vit * [example] update opt * [example] update palm * [example] update vit and opt benchmark * [hotfix] fix bert in model zoo (#4480) * [hotfix] fix bert in model zoo * [test] remove chatglm gemini test * [test] remove sam gemini test * [test] remove vit gemini test * [hotfix] fix opt tutorial example (#4497) * [hotfix] fix opt tutorial example * [hotfix] fix opt tutorial example
2023-08-24 01:29:25 +00:00
<!--- doc-test-ignore-start -->
训练
```python
optimizer.zero_grad()
outputs = model(input_ids, attn_mask)
loss = criterion(outputs, input_ids)
optimizer.backward(loss)
optimizer.step()
```
[gemini] improve compatibility and add static placement policy (#4479) * [gemini] remove distributed-related part from colotensor (#4379) * [gemini] remove process group dependency * [gemini] remove tp part from colo tensor * [gemini] patch inplace op * [gemini] fix param op hook and update tests * [test] remove useless tests * [test] remove useless tests * [misc] fix requirements * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [misc] update requirements * [gemini] refactor gemini optimizer and gemini ddp (#4398) * [gemini] update optimizer interface * [gemini] renaming gemini optimizer * [gemini] refactor gemini ddp class * [example] update gemini related example * [example] update gemini related example * [plugin] fix gemini plugin args * [test] update gemini ckpt tests * [gemini] fix checkpoint io * [example] fix opt example requirements * [example] fix opt example * [example] fix opt example * [example] fix opt example * [gemini] add static placement policy (#4443) * [gemini] add static placement policy * [gemini] fix param offload * [test] update gemini tests * [plugin] update gemini plugin * [plugin] update gemini plugin docstr * [misc] fix flash attn requirement * [test] fix gemini checkpoint io test * [example] update resnet example result (#4457) * [example] update bert example result (#4458) * [doc] update gemini doc (#4468) * [example] update gemini related examples (#4473) * [example] update gpt example * [example] update dreambooth example * [example] update vit * [example] update opt * [example] update palm * [example] update vit and opt benchmark * [hotfix] fix bert in model zoo (#4480) * [hotfix] fix bert in model zoo * [test] remove chatglm gemini test * [test] remove sam gemini test * [test] remove vit gemini test * [hotfix] fix opt tutorial example (#4497) * [hotfix] fix opt tutorial example * [hotfix] fix opt tutorial example
2023-08-24 01:29:25 +00:00
<!--- doc-test-ignore-end -->
> ⚠️ 注意:请不要使用`loss.backward()`,规范写法是`optimizer.backward(loss)`。
### 训练GPT
在此例程中, 我们使用 `Hugging Face Transformers`,并以 `GPT2 Medium` 为例。你必须在允许该例程前安装 `transformers`
为了简单起见,我们在这里只使用随机生成的数据。
首先我们只需要引入`Huggingface transformers` 的 `GPT2LMHeadModel`来定义我们的模型,不需要用户进行模型的定义与修改,方便用户使用。
定义GPT模型
```python
class GPTLMModel(nn.Module):
def __init__(self,
hidden_size=768,
num_layers=12,
num_attention_heads=12,
max_seq_len=1024,
vocab_size=50257,
checkpoint=False):
super().__init__()
self.checkpoint = checkpoint
self.model = GPT2LMHeadModel(
GPT2Config(n_embd=hidden_size,
n_layer=num_layers,
n_head=num_attention_heads,
n_positions=max_seq_len,
n_ctx=max_seq_len,
vocab_size=vocab_size))
if checkpoint:
self.model.gradient_checkpointing_enable()
def forward(self, input_ids, attention_mask):
return self.model(input_ids=input_ids, attention_mask=attention_mask, use_cache=not self.checkpoint)[0]
def gpt2_medium(checkpoint=False):
return GPTLMModel(hidden_size=1024, num_layers=24, num_attention_heads=16, checkpoint=checkpoint)
```
定义损失函数:
```python
class GPTLMLoss(nn.Module):
def __init__(self):
super().__init__()
self.loss_fn = nn.CrossEntropyLoss()
def forward(self, logits, labels):
shift_logits = logits[..., :-1, :].contiguous()
shift_labels = labels[..., 1:].contiguous()
return self.loss_fn(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
```
写一个获得随机输入的函数:
```python
def get_data(batch_size, seq_len, vocab_size):
input_ids = torch.randint(0, vocab_size, (batch_size, seq_len), device=torch.cuda.current_device())
attention_mask = torch.ones_like(input_ids)
return input_ids, attention_mask
```
最后使用booster注入 Gemini + ZeRO DDP 特性, 并定义训练循环。由于我们在这个例子中对GPT进行预训练因此只使用了一个简单的语言模型损失函数
```python
from colossalai.nn.optimizer import HybridAdam
from colossalai.booster import Booster
[gemini] improve compatibility and add static placement policy (#4479) * [gemini] remove distributed-related part from colotensor (#4379) * [gemini] remove process group dependency * [gemini] remove tp part from colo tensor * [gemini] patch inplace op * [gemini] fix param op hook and update tests * [test] remove useless tests * [test] remove useless tests * [misc] fix requirements * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [misc] update requirements * [gemini] refactor gemini optimizer and gemini ddp (#4398) * [gemini] update optimizer interface * [gemini] renaming gemini optimizer * [gemini] refactor gemini ddp class * [example] update gemini related example * [example] update gemini related example * [plugin] fix gemini plugin args * [test] update gemini ckpt tests * [gemini] fix checkpoint io * [example] fix opt example requirements * [example] fix opt example * [example] fix opt example * [example] fix opt example * [gemini] add static placement policy (#4443) * [gemini] add static placement policy * [gemini] fix param offload * [test] update gemini tests * [plugin] update gemini plugin * [plugin] update gemini plugin docstr * [misc] fix flash attn requirement * [test] fix gemini checkpoint io test * [example] update resnet example result (#4457) * [example] update bert example result (#4458) * [doc] update gemini doc (#4468) * [example] update gemini related examples (#4473) * [example] update gpt example * [example] update dreambooth example * [example] update vit * [example] update opt * [example] update palm * [example] update vit and opt benchmark * [hotfix] fix bert in model zoo (#4480) * [hotfix] fix bert in model zoo * [test] remove chatglm gemini test * [test] remove sam gemini test * [test] remove vit gemini test * [hotfix] fix opt tutorial example (#4497) * [hotfix] fix opt tutorial example * [hotfix] fix opt tutorial example
2023-08-24 01:29:25 +00:00
from colossalai.lazy import LazyInitContext
from colossalai.booster.plugin import GeminiPlugin
def main():
args = parse_args()
BATCH_SIZE = 8
SEQ_LEN = 1024
VOCAB_SIZE = 50257
NUM_STEPS = 10
colossalai.launch_from_torch(config={})
# build criterion
criterion = GPTLMLoss()
optimizer = HybridAdam(model.parameters(), lr=0.001)
torch.manual_seed(123)
# build GPT model
[gemini] improve compatibility and add static placement policy (#4479) * [gemini] remove distributed-related part from colotensor (#4379) * [gemini] remove process group dependency * [gemini] remove tp part from colo tensor * [gemini] patch inplace op * [gemini] fix param op hook and update tests * [test] remove useless tests * [test] remove useless tests * [misc] fix requirements * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [misc] update requirements * [gemini] refactor gemini optimizer and gemini ddp (#4398) * [gemini] update optimizer interface * [gemini] renaming gemini optimizer * [gemini] refactor gemini ddp class * [example] update gemini related example * [example] update gemini related example * [plugin] fix gemini plugin args * [test] update gemini ckpt tests * [gemini] fix checkpoint io * [example] fix opt example requirements * [example] fix opt example * [example] fix opt example * [example] fix opt example * [gemini] add static placement policy (#4443) * [gemini] add static placement policy * [gemini] fix param offload * [test] update gemini tests * [plugin] update gemini plugin * [plugin] update gemini plugin docstr * [misc] fix flash attn requirement * [test] fix gemini checkpoint io test * [example] update resnet example result (#4457) * [example] update bert example result (#4458) * [doc] update gemini doc (#4468) * [example] update gemini related examples (#4473) * [example] update gpt example * [example] update dreambooth example * [example] update vit * [example] update opt * [example] update palm * [example] update vit and opt benchmark * [hotfix] fix bert in model zoo (#4480) * [hotfix] fix bert in model zoo * [test] remove chatglm gemini test * [test] remove sam gemini test * [test] remove vit gemini test * [hotfix] fix opt tutorial example (#4497) * [hotfix] fix opt tutorial example * [hotfix] fix opt tutorial example
2023-08-24 01:29:25 +00:00
with ColoInitContext(default_device=torch.device('cuda')):
model = gpt2_medium(checkpoint=True)
[gemini] improve compatibility and add static placement policy (#4479) * [gemini] remove distributed-related part from colotensor (#4379) * [gemini] remove process group dependency * [gemini] remove tp part from colo tensor * [gemini] patch inplace op * [gemini] fix param op hook and update tests * [test] remove useless tests * [test] remove useless tests * [misc] fix requirements * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [test] fix model zoo * [misc] update requirements * [gemini] refactor gemini optimizer and gemini ddp (#4398) * [gemini] update optimizer interface * [gemini] renaming gemini optimizer * [gemini] refactor gemini ddp class * [example] update gemini related example * [example] update gemini related example * [plugin] fix gemini plugin args * [test] update gemini ckpt tests * [gemini] fix checkpoint io * [example] fix opt example requirements * [example] fix opt example * [example] fix opt example * [example] fix opt example * [gemini] add static placement policy (#4443) * [gemini] add static placement policy * [gemini] fix param offload * [test] update gemini tests * [plugin] update gemini plugin * [plugin] update gemini plugin docstr * [misc] fix flash attn requirement * [test] fix gemini checkpoint io test * [example] update resnet example result (#4457) * [example] update bert example result (#4458) * [doc] update gemini doc (#4468) * [example] update gemini related examples (#4473) * [example] update gpt example * [example] update dreambooth example * [example] update vit * [example] update opt * [example] update palm * [example] update vit and opt benchmark * [hotfix] fix bert in model zoo (#4480) * [hotfix] fix bert in model zoo * [test] remove chatglm gemini test * [test] remove sam gemini test * [test] remove vit gemini test * [hotfix] fix opt tutorial example (#4497) * [hotfix] fix opt tutorial example * [hotfix] fix opt tutorial example
2023-08-24 01:29:25 +00:00
# Gemini + ZeRO DP
plugin = GeminiPlugin(max_norm=1.0, initial_scale=2**5)
booster = Booster(plugin=plugin)
model, optimizer, criterion, _, _ = booster.boost(model, optimizer, criterion)
torch.cuda.synchronize()
model.train()
for n in range(NUM_STEPS):
# we just use randomly generated data here
input_ids, attn_mask = get_data(BATCH_SIZE, SEQ_LEN, VOCAB_SIZE)
optimizer.zero_grad()
outputs = model(input_ids, attn_mask)
loss = criterion(outputs, input_ids)
booster.backward(loss, optimizer)
optimizer.step()
torch.cuda.synchronize()
```
> ⚠️ 注意如果你使用Gemini模块的话请不要使用我们之前提到过的[梯度累加](../features/gradient_accumulation.md)。
完整的例子代码可以在 [Train GPT with Colossal-AI](https://github.com/hpcaitech/ColossalAI/tree/main/examples/language/gpt). 获得。
<!-- doc-test-command: torchrun --standalone --nproc_per_node=1 zero_with_chunk.py -->