Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

fix description and format error in docs #106

Merged
merged 1 commit into from
Oct 20, 2023
Merged

fix description and format error in docs #106

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 2 additions & 2 deletions docs/user-tutorial/container-images.mdx
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,8 +25,8 @@ You can use MS-AMP image by `ghcr.io/microsoft/msamp:${tag}`, available tags are

| Tag | Description |
|-------------------|------------------------------------|
| v0.2.0-cuda12.1 | MS-AMP v0.9.0 with CUDA 12.1 |
| v0.2.0-cuda11.8 | MS-AMP v0.9.0 with CUDA 11.8 |
| v0.2.0-cuda12.1 | MS-AMP v0.2.0 with CUDA 12.1 |
| v0.2.0-cuda11.8 | MS-AMP v0.2.0 with CUDA 11.8 |

</TabItem>
</Tabs>
3 changes: 3 additions & 0 deletions docs/user-tutorial/optimization-level.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,6 +2,8 @@
id: optimization-level
---

# Optimization Level

Currently MS-AMP supports three optimization levels: O1 and O2 and O3. The three levels gradually incorporate 8-bit collective communcation, optimizer and distributed parallel training in an incremental manner. Users can directly set O1/O2 using `msamp.initialize` and set O3 in config file when using DeepSpeed.

- O1: We found that directly transitioning weight gradients from FP32 to FP8 in the Transformer Engine leads to a decrease in accuracy. However, this issue is resolved in O1 through the implementation of FP8 for weight gradients and AllReduce communication. This optimization also has the added benefits of saving GPU memory and reducing communication bandwidth.
Expand All @@ -11,6 +13,7 @@ Currently MS-AMP supports three optimization levels: O1 and O2 and O3. The three
- O3: This optimization level is specifically designed for FP8 support in distributed parallel training for large scale models. These frequently-used strategies include data parallelism, tensor parallelism, pipeline parallelism, sequence parallelism and ZeRO optimizer. ZeRO separates model weights into regular weights and master weights, with the former used for network forward/backward on each GPU, and the latter used for model updating in the optimizer. This separation allows us to use 8-bit data precision for regular weights and weight broadcasting, which reduces GPU memory and bandwidth usage even further.

Here are details of different MS-AMP optimization levels:

| Optimization Level | Computation(GEMM) | Comm | Weight | Master Weight | Weight Gradient | Optimizer States |
| ------------------- | ----------- | ----- | ------ | ------------- | --------------- | ---------------- |
| FP16 AMP | FP16 | FP32 | FP32 | N/A | FP32 | FP32+FP32 |
Expand Down
Loading