* LLM: add readme to long-context examples. * add precision. * update wording. * add GPU type. * add Long-Context example to GPU examples. * fix comments. * update max input length. * update max length. * add output length. * fix wording.
33 lines
No EOL
1.7 KiB
Markdown
33 lines
No EOL
1.7 KiB
Markdown
# Running Long-Context generation using IPEX-LLM on Intel Arc™ A770 Graphics
|
|
|
|
Long-Context Generation is a critical aspect in various applications, such as document summarization, extended conversation handling, and complex question answering. Effective long-context generation can lead to more coherent and contextually relevant responses, enhancing user experience and model utility.
|
|
|
|
This folder contains examples of running long-context generation with IPEX-LLM on Intel Arc™ A770 Graphics(16GB GPU memory):
|
|
|
|
<!-- TODO: Maybe like this after adding more examples:
|
|
- [Single GPU](Single GPU): single GPU examples w & w/o batch.
|
|
- [Multiple GPU](Multiple GPU): multiple GPU examples w & w/o batch. -->
|
|
- [LLaMA2-32K](LLaMA2-32K): examples of running LLaMA2-32K models with INT4/FP8 precision.
|
|
- [ChatGLM3-32K](Chatglm3-32K): examples of running ChatGLM3-32K models with INT4/FP8 precision.
|
|
|
|
### Maximum Input Length for Different Models with INT4/FP8 Precision.
|
|
|
|
- **INT4**
|
|
|
|
| Model Name | Low Memory Mode | Maximum Input Length | Output Length |
|
|
| -- | -- | -- | -- |
|
|
| LLaMA2-7B-32K | Disable | 10K | 512 |
|
|
| | Enable | 12K | 512 |
|
|
| ChatGLM3-6B-32K | Disable | 9K | 512 |
|
|
| | Enable | 10K | 512 |
|
|
|
|
- **FP8**
|
|
|
|
| Model Name | Low Memory Mode | Maximum Input Length | Output Length |
|
|
| -- | -- | -- | -- |
|
|
| LLaMA2-7B-32K | Disable | 7K | 512 |
|
|
| | Enable | 9K | 512 |
|
|
| ChatGLM3-6B-32K | Disable | 8K | 512|
|
|
| | Enable | 9K | 512 |
|
|
|
|
> Note: If you need to run longer input or use less memory, please set `IPEX_LLM_LOW_MEM=1` to enable **low memory mode**, which will enable memory optimization and may slightly affect the performance. |