Skip to content

Commit

Permalink
Updated readme for 4-bit weight compression (#2237)
Browse files Browse the repository at this point in the history
  • Loading branch information
@ljaljushkin
ljaljushkin committed Nov 8, 2023
1 parent 96de397 commit 71a60de
Show file tree
Hide file tree
Showing 2 changed files with 192 additions and 13 deletions.
203 changes: 191 additions & 12 deletions docs/compression_algorithms/CompressWeights.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -6,14 +6,13 @@

The Weights Compression algorithm is aimed at compressing the weights of the models and can be used to optimize the model footprint and performance of large models where the size of weights is relatively larger than the size of activations, for example, Large Language Models (LLM). The algorithm compresses weights only for Linear and Embedding layers.

##### INT8 and NF4 modes
#### Supported modes

By default, weights are compressed to 8-bit integer data type - "INT8" mode.
OpenVINO backend has also an experimental support for "NF4" mode - compression to [nf4](https://arxiv.org/pdf/2305.14314v1.pdf) data type.
It goes with a grouped quantization, when small group of weights (e.g. 128) in the channel dimension share quantization parameters (scale).
First embedding and last linear layers are always compressed to 8-bit integer data type in the "NF4" mode.
Percent of the rest layers compressed to NF4 can be configured by "ratio" parameter.
E.g. ratio=0.9 means 90% of layers compressed to nf4 and the rest to 8-bit integer data type.
OpenVINO backend also supports 3 modes of mixed precision weight quantization with a 4-bit data type as a primary precision - INT4_SYM, INT4_ASYM and NF4. The primary precision in case of INT4_SYM mode is unsigned 4-bit integer and weights are quantized to it [symmetrically](https://github.com/openvinotoolkit/nncf/blob/develop/docs/compression_algorithms/Quantization.md#symmetric-quantization) with a fixed zero point equals to 8. In case of INT4_ASYM mode - also unsigned 4-bit integer, but weight are quantized to it [asymmetrically](https://github.com/openvinotoolkit/nncf/blob/develop/docs/compression_algorithms/Quantization.md#asymmetric-quantization) with a typical non-fixed zero point. In case of NF4 mode - [nf4](https://arxiv.org/pdf/2305.14314v1.pdf) data type without zero point.
All 4-bit modes have a grouped quantization support, when small group of weights (e.g. 128) in the channel dimension share quantization parameters (scale).
First embedding and last linear layers are always compressed to 8-bit integer data type.
Percent of the rest layers compressed to 4-bit can be configured by "ratio" parameter. E.g. ratio=0.9 means 90% of layers compressed to the corresponding 4-bit data type and the rest to 8-bit integer data type.

#### User guide

Expand All @@ -24,26 +23,206 @@ from nncf import compress_weights
compressed_model = compress_weights(model)
```

- Compress weights to nf4 data type with group size = 128, except first embedding and last linear layers - they are compressed to 8-bit integer data type.
- Compress weights symmetrically to 4-bit integer data type with group size = 128, except first embedding and last linear layers - they are compressed to 8-bit integer data type.

```python
from nncf import compress_weights
from nncf import CompressWeightsMode
compressed_model = compress_weights(model, mode=CompressWeightsMode.NF4)
compressed_model = compress_weights(model, mode=CompressWeightsMode.INT4_SYM)
```

- Compress weights of 90% of layers to nf4 with the group size 64, and the rest of layers to 8-bit integer data type.
- Generally, `INT4_SYM` mode is the fastest mixed-precision mode, but it may lead to a significant accuracy degradation or perplexity increase.
Compressing weights asymmetrically (`INT4_ASYM` mode) is the way to increase accuracy, however in turns it slows down inference a bit.
If the accuracy or perplexity is still not satisfying, there are 2 more hyper-parameters to tune: `group_size` and `ratio`.
Lower group size and less ratio of 4-bit layers usually improve accuracy at the sacrifice of inference speed.
Below is the example how to compress weights of 90% of layers to 4-bit integer asymmetrically with the group size 64, and
the rest of layers to 8-bit integer data type. The same parametrization is applicable for `INT4_SYM` mode.

```python
from nncf import compress_weights
from nncf import CompressWeightsMode
compressed_model = compress_weights(model, mode=CompressWeightsMode.NF4, group_size=64, ratio=0.9)
compressed_model = compress_weights(model, mode=CompressWeightsMode.INT4_ASYM, group_size=64, ratio=0.9)
```

##### Limitations
- `NF4` mode can be considered for improving accuracy, but currently models quantized to nf4 should not be faster models
quantized to 8-bit integer. Here's the example how to compress weights to nf4 data type with group size = 128.
Different `group_size` and `ratio` are also supported.

```python
from nncf import compress_weights
from nncf import CompressWeightsMode
compressed_model = compress_weights(model, mode=CompressWeightsMode.NF4)
```

#### Evaluation results

Here is the perplexity and model size before and after weight compression for different language models on the [Lambada OpenAI dataset](https://github.com/openai/gpt-2/issues/131#issuecomment-497136199).
`g32` refers to the group size equals to 32, `r60` - to the ratio equals to 0.6.

<table>
<thead>
<tr>
<th class="tg-0pky">Model</th>
<th class="tg-0pky">Mode</th>
<th class="tg-0pky">Perplexity</th>
<th class="tg-0pky">Perplexity <br>Increase</th>
<th class="tg-0pky">Model Size <br>(Gb)</th>
</tr>
</thead>
<tbody>
<tr>
<td class="tg-0pky">databricks/dolly-v2-3b</td>
<td class="tg-0pky">fp32</td>
<td class="tg-0pky">5.01</td>
<td class="tg-0pky">0</td>
<td class="tg-0pky">10.3</td>
</tr>
<tr>
<td class="tg-0pky">databricks/dolly-v2-3b</td>
<td class="tg-0pky">int8</td>
<td class="tg-0pky">5.07</td>
<td class="tg-0pky">0.05</td>
<td class="tg-0pky">2.6</td>
</tr>
<tr>
<td class="tg-0pky">databricks/dolly-v2-3b</td>
<td class="tg-0pky">int4_asym_g32_r50</td>
<td class="tg-0pky">5.28</td>
<td class="tg-0pky">0.26</td>
<td class="tg-0pky">2.2</td>
</tr>
<tr>
<td class="tg-0pky">databricks/dolly-v2-3b</td>
<td class="tg-0pky">nf4_g128_r60</td>
<td class="tg-0pky">5.19</td>
<td class="tg-0pky">0.18</td>
<td class="tg-0pky">1.9</td>
</tr>
<tr>
<td class="tg-0pky">facebook/opt-6.7b</td>
<td class="tg-0pky">fp32</td>
<td class="tg-0pky">4.25</td>
<td class="tg-0pky">0</td>
<td class="tg-0pky">24.8</td>
</tr>
<tr>
<td class="tg-0pky">facebook/opt-6.7b</td>
<td class="tg-0pky">int8</td>
<td class="tg-0pky">4.27</td>
<td class="tg-0pky">0.01</td>
<td class="tg-0pky">6.2</td>
</tr>
<tr>
<td class="tg-0pky">facebook/opt-6.7b</td>
<td class="tg-0pky">int4_asym_g64_r80</td>
<td class="tg-0pky">4.32</td>
<td class="tg-0pky">0.07</td>
<td class="tg-0pky">4.1</td>
</tr>
<tr>
<td class="tg-0pky">facebook/opt-6.7b</td>
<td class="tg-0pky">nf4_g64</td>
<td class="tg-0pky">4.35</td>
<td class="tg-0pky">0.1</td>
<td class="tg-0pky">3.6</td>
</tr>
<tr>
<td class="tg-0pky">meta-llama/Llama-2-7b-chat-hf</td>
<td class="tg-0pky">fp32</td>
<td class="tg-0pky">3.28</td>
<td class="tg-0pky">0</td>
<td class="tg-0pky">25.1</td>
</tr>
<tr>
<td class="tg-0pky">meta-llama/Llama-2-7b-chat-hf</td>
<td class="tg-0pky">int8</td>
<td class="tg-0pky">3.29</td>
<td class="tg-0pky">0.01</td>
<td class="tg-0pky">6.3</td>
</tr>
<tr>
<td class="tg-0pky">meta-llama/Llama-2-7b-chat-hf</td>
<td class="tg-0pky">int4_asym_g128_r80</td>
<td class="tg-0pky">3.41</td>
<td class="tg-0pky">0.14</td>
<td class="tg-0pky">4.0</td>
</tr>
<tr>
<td class="tg-0pky">meta-llama/Llama-2-7b-chat-hf</td>
<td class="tg-0pky">nf4_g128</td>
<td class="tg-0pky">3.41</td>
<td class="tg-0pky">0.13</td>
<td class="tg-0pky">3.5</td>
</tr>
<tr>
<td class="tg-0pky">togethercomputer/RedPajama-INCITE-7B-Instruct</td>
<td class="tg-0pky">fp32</td>
<td class="tg-0pky">4.15</td>
<td class="tg-0pky">0</td>
<td class="tg-0pky">25.6</td>
</tr>
<tr>
<td class="tg-0pky">togethercomputer/RedPajama-INCITE-7B-Instruct</td>
<td class="tg-0pky">int8</td>
<td class="tg-0pky">4.17</td>
<td class="tg-0pky">0.02</td>
<td class="tg-0pky">6.4</td>
</tr>
<tr>
<td class="tg-0pky">togethercomputer/RedPajama-INCITE-7B-Instruct</td>
<td class="tg-0pky">nf4_ov_g32_r60</td>
<td class="tg-0pky">4.28</td>
<td class="tg-0pky">0.13</td>
<td class="tg-0pky">5.1</td>
</tr>
<tr>
<td class="tg-0pky">togethercomputer/RedPajama-INCITE-7B-Instruct</td>
<td class="tg-0pky">int4_asym_g128</td>
<td class="tg-0pky">4.17</td>
<td class="tg-0pky">0.02</td>
<td class="tg-0pky">3.6</td>
</tr>
<tr>
<td class="tg-0pky">meta-llama/Llama-2-13b-chat-hf</td>
<td class="tg-0pky">fp32</td>
<td class="tg-0pky">2.92</td>
<td class="tg-0pky">0</td>
<td class="tg-0pky">48.5</td>
</tr>
<tr>
<td class="tg-0pky">meta-llama/Llama-2-13b-chat-hf</td>
<td class="tg-0pky">int8</td>
<td class="tg-0pky">2.91</td>
<td class="tg-0pky">0</td>
<td class="tg-0pky">12.1</td>
</tr>
<tr>
<td class="tg-0pky">meta-llama/Llama-2-13b-chat-hf</td>
<td class="tg-0pky">int4_sym_g64_r80</td>
<td class="tg-0pky">2.98</td>
<td class="tg-0pky">0.06</td>
<td class="tg-0pky">8.0</td>
</tr>
<tr>
<td class="tg-0pky">meta-llama/Llama-2-13b-chat-hf</td>
<td class="tg-0pky">nf4_g128</td>
<td class="tg-0pky">2.95</td>
<td class="tg-0pky">0.04</td>
<td class="tg-0pky">6.6</td>
</tr>
</tbody>
</table>

#### Limitations

- The algorithm is supported for OpenVINO and PyTorch models.
- The compression applies in-place.
- The compressed model is not trainable.
- NF4 mode, grouped quantization and mixed nf4-int8 precision selection is available for OpenVINO backend only.
- INT4_SYM, INT4_ASYM and NF4 modes, grouped quantization and mixed precision selection is available for OpenVINO backend only.
- NF4 support is experimental - models quantized to nf4 should not be faster models quantized to 8-bit integer.

#### Additional resources

- [LLM Weight Compression](https://docs.openvino.ai/nightly/weight_compression.html)
- [Optimize and Deploy Generative AI Models using Hugging Face Optimum Intel](https://docs.openvino.ai/nightly/gen_ai_guide.html)
- [Optimum Intel documentation](https://huggingface.co/docs/optimum/intel/inference)
2 changes: 1 addition & 1 deletion nncf/quantization/algorithms/weight_compression/openvino_backend.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -334,7 +334,7 @@ def _get_bitwidth_distribution_str(all_weight_params: List[WeightNodeParams]) ->
num_bits_vs_num_weights_map[num_bits] = (n_total, n_internal)

# Table creation
header = ["Num bits (N)", "% all weight", "% internal weights"]
header = ["Num bits (N)", "% all parameters (layers)", "% internal parameters (layers)"]
rows = []
for bitwidth, (n_total, n_internal) in num_bits_vs_num_weights_map.items():
rows.append(
Expand Down

0 comments on commit 71a60de

Please sign in to comment.