Deployment scripts

.gitignore

+ssmt_triton_repo

Dockerfile

+FROM nvcr.io/nvidia/tritonserver:23.06-py3
+WORKDIR /opt/tritonserver
+RUN apt-get update && apt-get install -y python3.10-venv
+ENV VIRTUAL_ENV=/opt/dhruva-mt
+RUN python3 -m venv $VIRTUAL_ENV
+ENV PATH="$VIRTUAL_ENV/bin:$PATH"
+RUN pip install -U ctranslate2 OpenNMT-py==1.2.0 git+https://github.com/vmujadia/tokenizer.git
+CMD ["tritonserver", "--model-repository=/models", "--cache-config=local,size=1048576"]
+EXPOSE 8000

README.md

 # mt-model-deploy-dhruva
 
+## TL;DR
+
+This repo contains code for python backend CTranslate2 based triton models for the SSMT project.
+Prerequisites: `python3.xx-venv`, `nvidia-docker`
+```bash
+git clone http://ssmt.iiit.ac.in/meitygit/ssmt/mt-model-deploy-dhruva.git
+cd mt-model-deploy-dhruva
+sh make_triton_model_repo.sh "https://ssmt.iiit.ac.in/uploads/data_mining/models.zip" "float16"
+docker build -t dhruva/ssmt-model-server:1 .
+nvidia-docker run --gpus=all --rm --shm-size 5g --network=host --name dhruva-ssmt-triton-server -v./ssmt_triton_repo:/models dhruva/ssmt-model-server:1
+```
+
+## What this repo does
+
+* This repo contains the templates and component triton models for the SSMT project.
+* Also contained is a Dockerfile to construct the triton server instance.
+* Given a URL and quantization method (those supported by CTranslate2 i.e. `int8`, `int8_float16`, `int8_bfloat16`, `int16`, `float16` and `bfloat16`) it will download, quantize and construct the SSMT Tritom Repository in `./ssmt_triton_repo`.
+* Dynamic batching and caching is supported and enabled by default.
+* The repository folder can me mounted to the dhruva ssmt triton server on `/models` and can be queried via a client.
+* Sample client code is also given as an ipython notebook.
+* The `model.zip` package needs to contain a folder of `.pt` and `.src` files named `1` through `9` with each file corresponding to the following mapping:
+`{'eng-hin': 1, 'hin-eng': 2, 'eng-tel': 3, 'tel-eng': 4, 'hin-tel': 6, 'tel-hin': 7, 'eng-guj': 8, 'guj-eng': 9}`
+
+## Architecture of the pipeline
+
+The pipeline consists of 4 components, executed in order:
+* Tokenizer - This model is CPU only and tokenizes and applies BPE on the input string.
+* Model Demuxer - This model is CPU only and depending on the language pair requested, queues up an `InferenceRequest` for the appropriate model and returns it as the response.
+* Model - This is a GPU based model and it processes the tokenized text and returns the final form of the translated text to the caller.
+* Pipeline - This is an ensemble model that wraps the above three components together and is the one meant to be exposed to the client.
+The exact specifications of the model inputs and outputs can be looked at in the corresponding `config.pbtxt` files.
+One can construct the triton repo like so:
+```bash
+git clone http://ssmt.iiit.ac.in/meitygit/ssmt/mt-model-deploy-dhruva.git
+cd mt-model-deploy-dhruva
+sh make_triton_model_repo.sh "https://ssmt.iiit.ac.in/uploads/data_mining/models.zip" "float16"
+```
+
+## Starting the triton server
+
+We customize the tritonserver image with the required python packages in a venv and enable the cache in the startup command. After the model repo has beeen built, one can build and run the server like so:
+```bash
+docker build -t dhruva/ssmt-model-server:1 .
+nvidia-docker run --gpus=all --rm --shm-size 5g --network=host --name dhruva-ssmt-triton-server -v./ssmt_triton_repo:/models dhruva/ssmt-model-server:1
+```
+
+## Querying the triton server
+
+We provide a sample ipython notebook that shows how to concurrently request the client for translations.
+Prerequisites: `pip install "tritonclient[all]" tqdm numpy`

make_triton_model_repo.sh

+MODELS_URL=$1
+QUANTIZATION=$2
+wget -O models.zip $MODELS_URL
+unzip models.zip
+python3 -m venv ./ssmt_ct2
+source ./ssmt_ct2/bin/activate
+pip install ctranslate2 "OpenNMT-py==1.2.0"
+cd models
+ct2-opennmt-py-converter --model_path 1.pt --quantization $QUANTIZATION --output_dir ./1_ct2
+ct2-opennmt-py-converter --model_path 2.pt --quantization $QUANTIZATION --output_dir ./2_ct2
+ct2-opennmt-py-converter --model_path 3.pt --quantization $QUANTIZATION --output_dir ./3_ct2
+ct2-opennmt-py-converter --model_path 4.pt --quantization $QUANTIZATION --output_dir ./4_ct2
+ct2-opennmt-py-converter --model_path 6.pt --quantization $QUANTIZATION --output_dir ./6_ct2
+ct2-opennmt-py-converter --model_path 7.pt --quantization $QUANTIZATION --output_dir ./7_ct2
+ct2-opennmt-py-converter --model_path 8.pt --quantization $QUANTIZATION --output_dir ./8_ct2
+ct2-opennmt-py-converter --model_path 9.pt --quantization $QUANTIZATION --output_dir ./9_ct2
+cd ..
+mkdir ssmt_triton_repo
+cd ssmt_triton_repo
+cp -r ../triton_models/ssmt_pipeline .
+cp -r ../triton_models/ssmt_model_demuxer .
+cp -r ../triton_models/ssmt_tokenizer .
+cp -r ../models/*.src ssmt_tokenizer/1/bpe_src
+cp -r ../triton_models/ssmt_template_model_repo ssmt_1_ct2
+cp -r ../models/1_ct2 ssmt_1_ct2/1/translator
+sed -i 's/model_name/ssmt_1_ct2/' ssmt_1_ct2/config.pbtxt
+cp -r ../triton_models/ssmt_template_model_repo ssmt_2_ct2
+cp -r ../models/2_ct2 ssmt_2_ct2/1/translator
+sed -i 's/model_name/ssmt_2_ct2/' ssmt_2_ct2/config.pbtxt
+cp -r ../triton_models/ssmt_template_model_repo ssmt_3_ct2
+cp -r ../models/3_ct2 ssmt_3_ct2/1/translator
+sed -i 's/model_name/ssmt_3_ct2/' ssmt_3_ct2/config.pbtxt
+cp -r ../triton_models/ssmt_template_model_repo ssmt_4_ct2
+cp -r ../models/4_ct2 ssmt_4_ct2/1/translator
+sed -i 's/model_name/ssmt_4_ct2/' ssmt_4_ct2/config.pbtxt
+cp -r ../triton_models/ssmt_template_model_repo ssmt_6_ct2
+cp -r ../models/6_ct2 ssmt_6_ct2/1/translator
+sed -i 's/model_name/ssmt_6_ct2/' ssmt_6_ct2/config.pbtxt
+cp -r ../triton_models/ssmt_template_model_repo ssmt_7_ct2
+cp -r ../models/7_ct2 ssmt_7_ct2/1/translator
+sed -i 's/model_name/ssmt_7_ct2/' ssmt_7_ct2/config.pbtxt
+cp -r ../triton_models/ssmt_template_model_repo ssmt_8_ct2
+cp -r ../models/8_ct2 ssmt_8_ct2/1/translator
+sed -i 's/model_name/ssmt_8_ct2/' ssmt_8_ct2/config.pbtxt
+cp -r ../triton_models/ssmt_template_model_repo ssmt_9_ct2
+cp -r ../models/9_ct2 ssmt_9_ct2/1/translator
+sed -i 's/model_name/ssmt_9_ct2/' ssmt_9_ct2/config.pbtxt
+cd ..
+source deactivate
+rm -rf ssmt_ct2
+rm -f models.zip
+rm -rf models

triton_client.ipynb

+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import numpy as np\n",
+    "from tqdm import tqdm\n",
+    "from random import choice\n",
+    "from tritonclient.utils import *\n",
+    "import tritonclient.http as httpclient\n",
+    "from multiprocessing.pool import ThreadPool"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "model_name = \"ssmt_pipeline\"\n",
+    "shape = [1]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def task(x):\n",
+    "    lang_pair_map = list({'eng-hin': 1, 'hin-eng': 2, 'eng-tel':3, 'tel-eng': 4, 'hin-tel': 6, 'tel-hin': 7, 'eng-guj': 8, 'guj-eng': 9}.keys())\n",
+    "    with httpclient.InferenceServerClient(\"localhost:8000\") as client:\n",
+    "        async_responses = []\n",
+    "        for i in range(10):\n",
+    "            s = 'this is a sentence.'\n",
+    "            source_data = np.array([[s]], dtype='object')\n",
+    "            inputs = [httpclient.InferInput(\"INPUT_TEXT\", source_data.shape, np_to_triton_dtype(source_data.dtype)), httpclient.InferInput(\"INPUT_LANGUAGE_ID\", source_data.shape, np_to_triton_dtype(source_data.dtype)), httpclient.InferInput(\"OUTPUT_LANGUAGE_ID\", source_data.shape, np_to_triton_dtype(source_data.dtype))]\n",
+    "            inputs[0].set_data_from_numpy(np.array([[s]], dtype='object'))\n",
+    "            langpair = choice(lang_pair_map)\n",
+    "            inputs[1].set_data_from_numpy(np.array([[langpair.split('-')[0].strip()]], dtype='object'))\n",
+    "            inputs[2].set_data_from_numpy(np.array([[langpair.split('-')[1].strip()]], dtype='object'))\n",
+    "            outputs = [httpclient.InferRequestedOutput(\"OUTPUT_TEXT\")]\n",
+    "            async_responses.append(client.async_infer(model_name, inputs, request_id=str(1), outputs=outputs))\n",
+    "        for r in async_responses: r.get_result(timeout=10).get_response()\n",
+    "    return 0"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "100%|██████████| 1000/1000 [01:49<00:00,  9.15it/s]\n"
+     ]
+    }
+   ],
+   "source": [
+    "with ThreadPool(100) as pool:\n",
+    "    for output in tqdm(pool.imap_unordered(task, range(1000), chunksize=1), total=1000): pass"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "model_metrics",
+   "language": "python",
+   "name": "model_metrics"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.11.4"
+  },
+  "orig_nbformat": 4
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}

triton_models/ssmt_model_demuxer/1/model.py

+import json
+import asyncio
+import triton_python_backend_utils as pb_utils
+
+class TritonPythonModel:
+    def initialize(self, args):
+        self.model_config = json.loads(args["model_config"])
+        target_config = pb_utils.get_output_config_by_name(self.model_config, "OUTPUT_TEXT")
+        self.target_dtype = pb_utils.triton_string_to_numpy(target_config["data_type"])
+        self.lang_pair_map = {'eng-hin': 1, 'hin-eng': 2, 'eng-tel': 3, 'tel-eng': 4, 'hin-tel': 6, 'tel-hin': 7, 'eng-guj': 8, 'guj-eng': 9}
+
+    async def execute(self, requests):
+        responses = []
+        infer_response_awaits = []
+        for request in requests:
+            language_pair = f"{pb_utils.get_input_tensor_by_name(request, 'INPUT_LANGUAGE_ID').as_numpy()[0, 0].decode('utf-8')}-{pb_utils.get_input_tensor_by_name(request, 'OUTPUT_LANGUAGE_ID').as_numpy()[0, 0].decode('utf-8')}"
+            inference_request = pb_utils.InferenceRequest(model_name=f'ssmt_{self.lang_pair_map[language_pair]}_ct2', requested_output_names=['OUTPUT_TEXT'], inputs=[pb_utils.get_input_tensor_by_name(request, 'INPUT_TEXT_TOKENIZED')])
+            infer_response_awaits.append(inference_request.async_exec())
+        responses = await asyncio.gather(*infer_response_awaits)
+        return responses
+
+    def finalize(self): pass

triton_models/ssmt_model_demuxer/config.pbtxt

+name: "ssmt_model_demuxer"
+backend: "python"
+max_batch_size: 4096
+
+input [
+  {
+    name: "INPUT_TEXT_TOKENIZED"
+    data_type: TYPE_STRING
+    dims: [ 1 ]
+  }
+]
+input [
+  {
+    name: "INPUT_LANGUAGE_ID"
+    data_type: TYPE_STRING
+    dims: [ 1 ]
+  }
+]
+input [
+  {
+    name: "OUTPUT_LANGUAGE_ID"
+    data_type: TYPE_STRING
+    dims: [ 1 ]
+  }
+]
+
+output [
+  {
+    name: "OUTPUT_TEXT"
+    data_type: TYPE_STRING
+    dims: [ 1 ]
+  }
+]
+
+dynamic_batching {}
+
+instance_group [
+  {
+    count: 1
+    kind: KIND_CPU
+  }
+]
+
+response_cache {
+  enable: true
+}

triton_models/ssmt_pipeline/config.pbtxt

+name: "ssmt_pipeline"
+platform: "ensemble"
+max_batch_size: 4096
+
+input [
+  {
+    name: "INPUT_TEXT"
+    data_type: TYPE_STRING
+    dims: [ 1 ]
+  }
+]
+input [
+  {
+    name: "INPUT_LANGUAGE_ID"
+    data_type: TYPE_STRING
+    dims: [ 1 ]
+  }
+]
+input [
+  {
+    name: "OUTPUT_LANGUAGE_ID"
+    data_type: TYPE_STRING
+    dims: [ 1 ]
+  }
+]
+
+output [
+  {
+    name: "OUTPUT_TEXT"
+    data_type: TYPE_STRING
+    dims: [ 1 ]
+  }
+]
+
+ensemble_scheduling {
+  step [
+    {
+      model_name: "ssmt_tokenizer"
+      model_version: 1
+      input_map {
+        key: "INPUT_TEXT"
+        value: "INPUT_TEXT"
+      }
+      input_map {
+        key: "INPUT_LANGUAGE_ID"
+        value: "INPUT_LANGUAGE_ID"
+      }
+      input_map {
+        key: "OUTPUT_LANGUAGE_ID"
+        value: "OUTPUT_LANGUAGE_ID"
+      }
+      output_map {
+        key: "INPUT_TEXT_TOKENIZED"
+        value: "INPUT_TEXT_TOKENIZED"
+      }
+    },
+    {
+      model_name: "ssmt_model_demuxer"
+      model_version: 1
+      input_map {
+        key: "INPUT_TEXT_TOKENIZED"
+        value: "INPUT_TEXT_TOKENIZED"
+      }
+      input_map {
+        key: "INPUT_LANGUAGE_ID"
+        value: "INPUT_LANGUAGE_ID"
+      }
+      input_map {
+        key: "OUTPUT_LANGUAGE_ID"
+        value: "OUTPUT_LANGUAGE_ID"
+      }
+      output_map {
+        key: "OUTPUT_TEXT"
+        value: "OUTPUT_TEXT"
+      }
+    }
+  ]
+}
+
+response_cache {
+  enable: true
+}

triton_models/ssmt_template_model_repo/1/model.py

+import os
+import json
+import numpy
+from itertools import islice
+from ctranslate2 import Translator
+import triton_python_backend_utils as pb_utils
+
+class TritonPythonModel:
+    def initialize(self, args):
+        current_path = os.path.dirname(os.path.abspath(__file__))
+        self.model_config = json.loads(args["model_config"])
+        self.device_id = int(json.loads(args['model_instance_device_id']))
+        target_config = pb_utils.get_output_config_by_name(self.model_config, "OUTPUT_TEXT")
+        self.target_dtype = pb_utils.triton_string_to_numpy(target_config["data_type"])
+        try: self.translator = Translator(f"{os.path.join(current_path, 'translator')}", device="cuda", intra_threads=1, inter_threads=1, device_index=[self.device_id])
+        except: self.translator = Translator(f"{os.path.join(current_path, 'translator')}", device="cpu", intra_threads=4)
+
+    def execute(self, requests):
+        source_list = [pb_utils.get_input_tensor_by_name(request, "INPUT_TEXT_TOKENIZED") for request in requests]
+        bsize_list = [source.as_numpy().shape[0] for source in source_list]
+        src_sentences = [s[0].decode('utf-8').strip().split(' ') for source in source_list for s in source.as_numpy()]
+        tgt_sentences = [' '.join(result.hypotheses[0]).replace('@@ ', '') for result in self.translator.translate_iterable(src_sentences, max_batch_size=128, max_input_length=100, max_decoding_length=100)]
+        responses = [pb_utils.InferenceResponse(output_tensors=[pb_utils.Tensor("OUTPUT_TEXT", numpy.array([[s]for s in islice(tgt_sentences, bsize)], dtype='object').astype(self.target_dtype))]) for bsize in bsize_list]
+        return responses
+
+    def finalize(self): self.translator.unload_model()

triton_models/ssmt_template_model_repo/config.pbtxt

+name: "model_name"
+backend: "python"
+max_batch_size: 512
+
+input [
+  {
+    name: "INPUT_TEXT_TOKENIZED"
+    data_type: TYPE_STRING
+    dims: [ 1 ]
+  }
+]
+output [
+  {
+    name: "OUTPUT_TEXT"
+    data_type: TYPE_STRING
+    dims: [ 1 ]
+  }
+]
+
+dynamic_batching {}
+
+instance_group [
+  {
+    count: 1
+    kind: KIND_GPU
+  }
+]
+
+response_cache {
+  enable: true
+}

triton_models/ssmt_tokenizer/1/model.py

+import os
+import json
+import numpy
+from .apply_bpe import BPE
+from ilstokenizer import tokenizer
+import triton_python_backend_utils as pb_utils
+
+class TritonPythonModel:
+    def initialize(self, args):
+        current_path = os.path.dirname(os.path.abspath(__file__))
+        self.model_config = json.loads(args["model_config"])
+        target_config = pb_utils.get_output_config_by_name(self.model_config, "INPUT_TEXT_TOKENIZED")
+        self.target_dtype = pb_utils.triton_string_to_numpy(target_config["data_type"])
+        self.lang_pair_map = {'eng-hin': 1, 'hin-eng': 2, 'eng-tel': 3, 'tel-eng': 4, 'hin-tel': 6, 'tel-hin': 7, 'eng-guj': 8, 'guj-eng': 9}
+        self.bpes = {lang_pair: BPE(open(os.path.join(current_path, f'bpe_src/{model_id}.src'), encoding='utf-8')) for lang_pair, model_id in self.lang_pair_map.items()}
+
+    def execute(self, requests):
+        source_gen = ((pb_utils.get_input_tensor_by_name(request, "INPUT_TEXT"), pb_utils.get_input_tensor_by_name(request, "INPUT_LANGUAGE_ID"), pb_utils.get_input_tensor_by_name(request, "OUTPUT_LANGUAGE_ID")) for request in requests)
+        tokenized_gen = (self.bpes[f"{input_language_id.as_numpy()[0, 0].decode('utf-8')}-{output_language_id.as_numpy()[0, 0].decode('utf-8')}"].segment(tokenizer.tokenize(input_text.as_numpy()[0, 0].decode('utf-8'))).strip() for input_text, input_language_id, output_language_id in source_gen)
+        responses = [pb_utils.InferenceResponse(output_tensors=[pb_utils.Tensor("INPUT_TEXT_TOKENIZED", numpy.array([[tokenized_sent]], dtype=self.target_dtype))]) for tokenized_sent in tokenized_gen]
+        return responses
+
+    def finalize(self): pass

triton_models/ssmt_tokenizer/config.pbtxt

+name: "ssmt_tokenizer"
+backend: "python"
+max_batch_size: 4096
+
+input [
+  {
+    name: "INPUT_TEXT"
+    data_type: TYPE_STRING
+    dims: [ 1 ]
+  }
+]
+input [
+  {
+    name: "INPUT_LANGUAGE_ID"
+    data_type: TYPE_STRING
+    dims: [ 1 ]
+  }
+]
+input [
+  {
+    name: "OUTPUT_LANGUAGE_ID"
+    data_type: TYPE_STRING
+    dims: [ 1 ]
+  }
+]
+
+output [
+  {
+    name: "INPUT_TEXT_TOKENIZED"
+    data_type: TYPE_STRING
+    dims: [ 1 ]
+  }
+]
+
+dynamic_batching {}
+
+instance_group [
+  {
+    count: 8
+    kind: KIND_CPU
+  }
+]
+
+response_cache {
+  enable: true
+}