README.md Update

README.md

@@ -1,4 +1,3 @@
-
 ---
 tags:
 - ctranslate2
@@ -16,12 +15,13 @@ license: apache-2.0
 
 CTranslate2 implements a custom runtime that applies many performance optimization techniques such as weights quantization, layers fusion, batch reordering, etc., to accelerate and reduce the memory usage of Transformer models on CPU and GPU.
 
-CTranslate2 is SOTA and is one of the most performant ways of hosting translation models at scale. Current supported models include:
+CTranslate2 is one of the most performant ways of hosting translation models at scale. Current supported models include:
 - Encoder-decoder models: Transformer base/big, M2M-100, NLLB, BART, mBART, Pegasus, T5, Whisper
 - Decoder-only models: GPT-2, GPT-J, GPT-NeoX, OPT, BLOOM, MPT, Llama, Mistral, Gemma, CodeGen, GPTBigCode, Falcon
 - Encoder-only models: BERT, DistilBERT, XLM-RoBERTa
 
-Speed up inference times by about **2x-8x** using **int8** inference in C++. CTranslate2 is SOTA for hosting translation models at scale.
+The project is production-oriented and comes with backward compatibility guarantees, but it also includes experimental features related to model compression and inference acceleration.
+
 # CTranslate2 Benchmarks
 Please note that the results presented below are only valid for the configuration used during this benchmark: absolute and relative performance may change with different settings. Tested against `newstest2014` (En -> De) dataset.
 
@@ -51,12 +51,16 @@ Please note that the results presented below are only valid for the configuratio
 **Source to benchmark information can be found [here](https://github.com/OpenNMT/CTranslate2).**<br />
 **Original model BLEU scores can be found [here](https://huggingface.co/Helsinki-NLP/opus-mt-de-gil).**
 
+## Internal Benchmarks
+Internal testing on our end showed **inference times reduced by 6x-10x** on average compared the vanilla checkpoints using the *transformers* library. A **slight reduction on BLEU scores (~5%)** was also identified in comparison to the vanilla checkpoints with a few exceptions. This is likely due to several factors, one being the quantization applied. Further testing is needed from our end to better assess the reduction in translation quality. The command used to compile the vanilla checkpoint into a CTranslate2 model can be found below. Modifying this command can yield differing balances between inferencing performance and translation quality.
+
+
 # CTranslate2 Installation
 ```bash
 pip install hf-hub-ctranslate2>=1.0.0 ctranslate2>=3.13.0
 ```
 ### ct2-transformers-converter Command Used:
-```
+```bash
 ct2-transformers-converter --model Helsinki-NLP/opus-mt-de-gil --output_dir ./ctranslate2/opus-mt-de-gil-ctranslate2 --force --copy_files README.md generation_config.json tokenizer_config.json vocab.json source.spm .gitattributes target.spm --quantization float16
 ```
 # CTranslate2 Converted Checkpoint Information:
@@ -65,24 +69,29 @@ ct2-transformers-converter --model Helsinki-NLP/opus-mt-de-gil --output_dir ./ct
 - [hf-hub-ctranslate2](https://github.com/michaelfeil/hf-hub-ctranslate2)
 
 **Compute Type:**
-- `compute_type=int8_float16` for `device="cuda"` 
+- `compute_type=int8_float16` for `device="cuda"`
 - `compute_type=int8`  for `device="cpu"`
 
 # Sample Code - ctranslate2
+#### Clone the repository to the working directory or wherever you wish to store the model artifacts. ####
+```bash
+git clone https://huggingface.co/gaudi/opus-mt-de-gil-ctranslate2
+```
+#### Take the python code below and update the 'model_dir' variable to the location of the cloned repository. ####
 ```python
 from ctranslate2 import Translator
 import transformers
 
-model_name = "gaudi/opus-mt-de-gil-ctranslate2"
+model_dir = "./opus-mt-de-gil-ctranslate2" # Path to model directory.
 translator = Translator(
-            model_path=model_name,
-            device="cuda",
-            inter_threads=1,
-            intra_threads=4,
-            compute_type="int8_float16",
+            model_path=model_dir,
+            device="cuda", # cpu, cuda, or auto.
+            inter_threads=1, # Maximum number of parallel translations.
+            intra_threads=4, # Number of OpenMP threads per translator.
+            compute_type="int8_float16", # int8 for cpu or int8_float16 for cuda.
 )
 
-tokenizer = transformers.AutoTokenizer.from_pretrained(model_name)
+tokenizer = transformers.AutoTokenizer.from_pretrained(model_dir)
 
 source = tokenizer.convert_ids_to_tokens(tokenizer.encode("XXXXXX, XXX XX XXXXXX."))
 results = translator.translate_batch([source])
@@ -98,9 +107,9 @@ from transformers import AutoTokenizer
 
 model_name = "gaudi/opus-mt-de-gil-ctranslate2"
 model = TranslatorCT2fromHfHub(
-        model_name_or_path=model_name, 
+        model_name_or_path=model_name,
         device="cuda",
-        compute_type="int8_float16" # load in int8 on CUDA,
+        compute_type="int8_float16",
         tokenizer=AutoTokenizer.from_pretrained(model_name)
 )
 outputs = model.generate(