Skip to content

Commit

Permalink
Bump version and update benchmark numbers (#40)
Browse files Browse the repository at this point in the history
  • Loading branch information
guillaumekln authored Nov 6, 2019
1 parent c4a5631 commit 9da12fa
Show file tree
Hide file tree
Showing 2 changed files with 17 additions and 8 deletions.
9 changes: 9 additions & 0 deletions CHANGELOG.md
Original file line number Diff line number Diff line change
Expand Up @@ -4,6 +4,15 @@

### Fixes and improvements

## [v1.2.1](https://github.com/OpenNMT/CTranslate2/releases/tag/v1.2.1) (2019-11-06)

### Fixes and improvements

* Avoid copying decoder states when possible to improve decoding performance (10% to 20% faster)
* Fix execution profiling on GPU (device was not synchronized before measuring the time)
* Include `Mul` operation in profiling report
* Add a Python 3 wheel in Ubuntu Docker images

## [v1.2.0](https://github.com/OpenNMT/CTranslate2/releases/tag/v1.2.0) (2019-10-28)

### New features
Expand Down
16 changes: 8 additions & 8 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@ CTranslate2 is a custom C++ inference engine for [OpenNMT-py](https://github.com

## Key features

* **Fast execution**<br/>The execution aims to be faster than a general purpose deep learning framework: on standard translation tasks, it is [up to 3x faster](#benchmarks) than OpenNMT-py.
* **Fast execution**<br/>The execution aims to be faster than a general purpose deep learning framework: on standard translation tasks, it is [up to 4x faster](#benchmarks) than OpenNMT-py.
* **Model quantization**<br/>Support INT16 quantization on CPU and INT8 quantization (experimental) on CPU and GPU.
* **Parallel translation**<br/>Translations can be run efficiently in parallel without duplicating the model data in memory.
* **Dynamic memory usage**<br/>The memory usage changes dynamically depending on the request size while still meeting performance requirements thanks to caching allocators on both CPU and GPU.
Expand Down Expand Up @@ -302,8 +302,8 @@ Configuration:

| | Tokens/s | BLEU |
| --- | --- | --- |
| CTranslate2 1.0.1 | 3314.65 | 26.69 |
| CTranslate2 1.0.1 (int8) | 2254.05 | 26.79 |
| CTranslate2 1.2.1 | 3917.32 | 26.70 |
| CTranslate2 1.2.1 (int8) | 2519.24 | 26.80 |
| OpenNMT-tf 1.25.0 | 1338.26 | 26.90 |
| OpenNMT-py 0.9.2 | 980.44 | 26.69 |

Expand All @@ -316,11 +316,11 @@ Configuration:

| | Tokens/s | BLEU |
| --- | --- | --- |
| CTranslate2 1.0.1 (int8 + vmap) | 471.43 | 26.59 |
| CTranslate2 1.0.1 (int16 + vmap) | 423.58 | 26.63 |
| CTranslate2 1.0.1 (int8) | 383.91 | 26.84 |
| CTranslate2 1.0.1 (int16) | 346.25 | 26.68 |
| CTranslate2 1.0.1 (float) | 335.34 | 26.69 |
| CTranslate2 1.2.1 (int8 + vmap) | 591.78 | 26.59 |
| CTranslate2 1.2.1 (int16 + vmap) | 514.36 | 26.63 |
| CTranslate2 1.2.1 (int8) | 448.21 | 26.84 |
| CTranslate2 1.2.1 (int16) | 395.46 | 26.68 |
| CTranslate2 1.2.1 (float) | 381.31 | 26.69 |
| OpenNMT-py 0.9.2 | 241.92 | 26.69 |
| OpenNMT-tf 1.25.0 | 119.34 | 26.90 |

Expand Down

0 comments on commit 9da12fa

Please sign in to comment.