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

Jump to bottom

Adding the compression tutorial on GPT distillation and quantization #2197

Merged
merged 3 commits into from
Aug 9, 2022
Merged

Adding the compression tutorial on GPT distillation and quantization #2197

Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
0948d60
Adding additional instructiosn in the compression tutorial on pre-tra…
minjiaz Aug 9, 2022
337b8df
Merge branch 'master' into minjiaz/compression-tutorial-addon
tjruwase Aug 9, 2022
bccc227
Merge branch 'master' into minjiaz/compression-tutorial-addon
jeffra Aug 9, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Jump to
Jump to file
Failed to load files.
Diff view
Diff view
20 changes: 8 additions & 12 deletions docs/_tutorials/model-compression.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -47,9 +47,7 @@ Epoch: 18 | Time: 12m 38s
Clean the best model, and the accuracy of the clean model is acc/mm-acc:0.8340295466123281/0.8339096826688365
```

<!-- TODO: uncomment the below task-agnostic tutorial after the code example is fixed and released. -->

<!-- To apply layer reduction for task-agnostic compression, we provide an example on how to do so in the GPT pre-training stage.
To apply layer reduction for task-agnostic compression, we provide an example on how to do so in the GPT pre-training stage.

Step 1: Obtain the latest version of the [Megatron-DeepSpeed](https://github.com/microsoft/Megatron-DeepSpeed).

Expand All @@ -69,16 +67,14 @@ Step 3: Run the example bash script such as `ds_pretrain_gpt_125M_dense_cl_kd.sh

Apart from the above configs, you may also need to modify the data path in the `data_options` so that the trainer knows the data location. To make things slightly easier, we provide several example scripts for running distillation for different model sizes, including 350M (`ds_pretrain_gpt_350M_dense_kd.sh`) and 1.3B models (`ds_pretrain_gpt_1.3B_dense_cl_kd.sh`). We also empirically found that a staged KD often led to a better pre-trained distilled model on downstream tasks. Therefore, we suggest an easy approach to early-stop KD by not setting `--kd` in the script provided (e.g., disabling KD in the remaining 40% of training).

Step 4 (optional): After distilling the model, one can also choose to further quantize the distilled model by running the script `125M-Int8-test-64gpu-distilled-group48.sh`, which quantizes both the weights and activations of a distilled model with INT8 quantizer (the weight and activation quantization are introduced in the following sections). We provide the perplexity results and zero-shot evaluation results (average across 10 tasks) in the following table.

| ---: | ---: | ---: |
| **GPT (125M)** | **PPL** | **Zero-shot evaluation** |
| Uncompressed | 8.92 | 32.01 |
| Pretraining distillation | 9.36 | 31.74 |
| Quantization only | 9.39 | 31.96 |
| Pretraining distillation + quantization | 9.86 | 31.32 | -->
Step 4: After distilling the model, one can also choose to further quantize the distilled model by running the script `125M-L10-Int8-test-64gpu-distilled-group48.sh`, which quantizes both the weights and activations of a distilled model with INT8 quantizer (the weight and activation quantization are introduced in the following sections). note that you need to set the `-reset-iteration` flag when performing the quantization. We provide the zero-shot perplexity result from WikiText-2 and LAMBADA in the following table.

<!-- TODO: Is the PPL the validation PPL or wikitext PPL? And there is no description about which 10 tasks are used, maybe it's better to just show one common task result (e.g., WikiText-2). -->
| ---: | ---: | ---: | ---: |
| **GPT (125M)** | **#Layers** | **wikitex2 perplexity** | **LAMBADA** |
| Uncompressed | 12 | 29.6 | 39.5 |
| Quantization only | 12 | 29.8 | 39.7 |
| Distillation only | 10 | 31.9 | 39.2 |
| Distillation + quantization | 10 | 32.28 | 38.7 |

### 1.2 Weight Quantization
**What is weight quantization**
Expand Down