Added sentence transformers example with TensorRT and Triton Ensemble (…

…#3615)

* Added  sentence transformers example with TensorRT and Triton Ensemble

* Notebook changes to pass CI build

* Grammar fixes and installing torch for CI build

* Installing torch to pass CI build

Co-authored-by: atqy <[email protected]>

inference/nlp/realtime/huggingface/sentence-transformers-triton-ensemble/README.md

@@ -0,0 +1,38 @@
+# NVIDIA Triton Inference Server on SageMaker - Hugging Face Sentence Transformers
+
+## Introduction
+
+[HuggingFace Sentence Transformers](https://huggingface.co/sentence-transformers) is a Machine Learning (ML) framework and set of pre-trained models to
+extract embeddings from sentence, text, and image. The models in this group can also be used with the default methods exposed through the [Transformers](https://www.google.com/search?q=transofrmers+githbu&rlz=1C5GCEM_enES937ES938&oq=transofrmers+githbu&aqs=chrome..69i57.3022j0j7&sourceid=chrome&ie=UTF-8) library.
+
+[NVIDIA Triton Inference Server](https://github.com/triton-inference-server/server/) is a high-performance ML model server, which enables the deployment of ML models in an easy, scalable, and cost-effective way. It also exposes many easy-to-use optimization features to make the most of the underlying hardware, in particular NVIDIA GPU's.
+
+In this example, we walk through how you can:
+* Create an Amazon SageMaker Studio image based on the official [NVIDIA PyTorch](https://catalog.ngc.nvidia.com/orgs/nvidia/containers/pytorch) image, which includes the necessary dependencies to optimize your model
+* Optimize a pre-trained HuggingFace Sentence Transformers model with NVIDIA TensorRT to enable high-performance inference
+* Create a Triton Model Ensemble, which will allow you to run in sequence a pre-processing step (input tokenization), model inference and post-processing, where sentence embeddings are computed from the raw token embeddings 
+
+This example is meant to serve as a basis for use-cases in which you need to run your own code before and/or after your model, allowing you to optimize the bulk of the computation (the model) using tools such as TensorRT.
+
+<img src="images/triton-ensemble.png" alt="Triton Model Ensamble" />
+
+#### ! Important: The example provided can be tested also by using Amazon SageMaker Notebook Instances
+
+### Prerequisites
+
+1. Required NVIDIA NGC Account. Follow the instruction https://docs.nvidia.com/ngc/ngc-catalog-user-guide/index.html#registering-activating-ngc-account
+
+##  Step 1: Clone this repository
+
+##  Step 2: Build Studio image
+
+In this example, we provide a [Dokerfile](./studio-image/image_tensorrt/Dockerfile) example to build a custom image for SageMaker Studio.
+
+To build the image, push it and make it available in your Amazon SageMaker Studio environment, edit [sagemaker-studio-config](./studio-image/studio-domain-config.json) by replacing `$DOMAIN_ID` with your Studio domain ID.
+
+We also provide automation scripts in order to [build and push](./studio-image/build_image.sh) your docker image to an ECR repository
+and [create](./studio-image/create_studio_image.sh) or [update](./studio-image/update_studio_image.sh) an Amazon SageMaker Image. Please follow the instructions in the [README](./studio-image/README.md) for additional info on the usage of this script.
+
+## Step 3: Compile model, create an Amazon SageMaker Real-Time Endpoint with NVIDIA Triton Inference Server
+
+Clone this repository into your Amazon SageMaker Studio environment and execute the cells in the [notebook](./examples/triton_sentence_embeddings.ipynb)

inference/nlp/realtime/huggingface/sentence-transformers-triton-ensemble/examples/ensemble_hf/bert-trt/config.pbtxt

@@ -0,0 +1,32 @@
+name: "bert-trt"
+platform: "tensorrt_plan"
+max_batch_size: 16
+input [
+  {
+    name: "token_ids"
+    data_type: TYPE_INT32
+    dims: [128]
+  },
+  {
+    name: "attn_mask"
+    data_type: TYPE_INT32
+    dims: [128]
+  }
+]
+output [
+  {
+    name: "output"
+    data_type: TYPE_FP32
+    dims: [128, 384]
+  },
+  {
+    name: "854"
+    data_type: TYPE_FP32
+    dims: [384]
+  }
+]
+instance_group [
+    {
+      kind: KIND_GPU
+    }
+  ]

inference/nlp/realtime/huggingface/sentence-transformers-triton-ensemble/examples/ensemble_hf/ensemble/1/README.md

@@ -0,0 +1 @@
+Do not delete me!

inference/nlp/realtime/huggingface/sentence-transformers-triton-ensemble/examples/ensemble_hf/ensemble/config.pbtxt

@@ -0,0 +1,70 @@
+name: "ensemble"
+platform: "ensemble"
+max_batch_size: 16
+input [
+  {
+    name: "INPUT0"
+    data_type: TYPE_STRING
+    dims: [ 1 ]
+  }
+]
+output [
+  {
+    name: "finaloutput"
+    data_type: TYPE_FP32
+    dims: [384]
+  }
+]
+ensemble_scheduling {
+  step [
+    {
+      model_name: "preprocess"
+      model_version: -1
+      input_map {
+        key: "INPUT0"
+        value: "INPUT0"
+      }
+      output_map {
+        key: "OUTPUT0"
+        value: "token_ids"
+      }
+      output_map {
+        key: "OUTPUT1"
+        value: "attn_mask"
+      }
+    },
+    {
+      model_name: "bert-trt"
+      model_version: -1
+      input_map {
+        key: "token_ids"
+        value: "token_ids"
+      }
+      input_map {
+        key: "attn_mask"
+        value: "attn_mask"
+      }
+      output_map {
+        key: "output"
+        value: "output"
+      }
+    },
+    {
+      model_name: "postprocess"
+      model_version: -1
+      input_map {
+        key: "TOKEN_EMBEDS_POST"
+        value: "output"
+      }
+      input_map {
+        key: "ATTENTION_POST"
+        value: "attn_mask"
+      }
+      output_map {
+        key: "SENT_EMBED"
+        value: "finaloutput"
+      }
+
+    }
+  ]
+}

inference/nlp/realtime/huggingface/sentence-transformers-triton-ensemble/examples/ensemble_hf/postprocess/1/model.py

@@ -0,0 +1,78 @@
+import json
+import logging
+import numpy as np
+import subprocess
+import sys
+import os
+
+import triton_python_backend_utils as pb_utils
+
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+class TritonPythonModel:
+    """This model loops through different dtypes to make sure that
+    serialize_byte_tensor works correctly in the Python backend.
+    """
+
+    def __mean_pooling(self, token_embeddings, attention_mask):
+        logger.info("token_embeddings: {}".format(token_embeddings))
+        logger.info("attention_mask: {}".format(attention_mask))
+
+        input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
+        return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)
+
+    def initialize(self, args):
+        self.model_dir = args['model_repository']
+        subprocess.check_call([sys.executable, "-m", "pip", "install", '-r', f'{self.model_dir}/requirements.txt'])
+        global torch
+        import torch 
+
+        self.device_id = args['model_instance_device_id']
+        self.model_config = model_config = json.loads(args['model_config'])
+        self.device = torch.device(f'cuda:{self.device_id}') if torch.cuda.is_available() else torch.device('cpu')
+
+        output0_config = pb_utils.get_output_config_by_name(
+            model_config, "SENT_EMBED")
+
+        self.output0_dtype = pb_utils.triton_string_to_numpy(
+            output0_config["data_type"])
+
+    def execute(self, requests):
+
+        responses = []
+
+        for request in requests:
+            tok_embeds = pb_utils.get_input_tensor_by_name(request, "TOKEN_EMBEDS_POST")
+            attn_mask = pb_utils.get_input_tensor_by_name(request, "ATTENTION_POST")
+
+            tok_embeds = tok_embeds.as_numpy()
+
+            logger.info("tok_embeds: {}".format(tok_embeds))
+            logger.info("tok_embeds shape: {}".format(tok_embeds.shape))
+
+            tok_embeds = torch.tensor(tok_embeds,device=self.device)
+
+            logger.info("tok_embeds_tensor: {}".format(tok_embeds))
+
+            attn_mask = attn_mask.as_numpy()
+
+            logger.info("attn_mask: {}".format(attn_mask))
+            logger.info("attn_mask shape: {}".format(attn_mask.shape))
+
+            attn_mask = torch.tensor(attn_mask,device=self.device)
+
+            logger.info("attn_mask_tensor: {}".format(attn_mask))
+
+            sentence_embeddings = self.__mean_pooling(tok_embeds, attn_mask)
+            sentence_embeddings = torch.nn.functional.normalize(sentence_embeddings, p=2, dim=1)
+
+            out_0 = np.array(sentence_embeddings.cpu(),dtype=self.output0_dtype)
+            logger.info("out_0: {}".format(out_0))
+
+            out_tensor_0 = pb_utils.Tensor("SENT_EMBED", out_0)
+            logger.info("out_tensor_0: {}".format(out_tensor_0))
+
+            responses.append(pb_utils.InferenceResponse([out_tensor_0]))
+
+        return responses

inference/nlp/realtime/huggingface/sentence-transformers-triton-ensemble/examples/ensemble_hf/postprocess/config.pbtxt

@@ -0,0 +1,26 @@
+name: "postprocess"
+backend: "python"
+max_batch_size: 16
+
+input [
+  {
+    name: "TOKEN_EMBEDS_POST"
+    data_type: TYPE_FP32
+    dims: [128, 384]
+    
+  },
+   {
+    name: "ATTENTION_POST"
+    data_type: TYPE_INT32
+    dims: [128]
+  }
+]
+output [
+  {
+    name: "SENT_EMBED"
+    data_type: TYPE_FP32
+    dims: [ 384 ]
+  }
+]
+
+instance_group [{ kind: KIND_GPU }]

inference/nlp/realtime/huggingface/sentence-transformers-triton-ensemble/examples/ensemble_hf/postprocess/requirements.txt

@@ -0,0 +1 @@
+torch

inference/nlp/realtime/huggingface/sentence-transformers-triton-ensemble/examples/ensemble_hf/preprocess/1/model.py

@@ -0,0 +1,74 @@
+import json
+import logging
+import numpy as np 
+import subprocess
+import sys
+
+import triton_python_backend_utils as pb_utils
+
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+class TritonPythonModel:
+    """This model loops through different dtypes to make sure that
+    serialize_byte_tensor works correctly in the Python backend.
+    """
+
+    def initialize(self, args):
+        self.model_dir = args['model_repository']
+        subprocess.check_call([sys.executable, "-m", "pip", "install", '-r', f'{self.model_dir}/requirements.txt'])
+        global transformers
+        import transformers
+
+        self.tokenizer = transformers.AutoTokenizer.from_pretrained('sentence-transformers/all-MiniLM-L6-v2')
+        self.model_config = model_config = json.loads(args['model_config'])
+
+        output0_config = pb_utils.get_output_config_by_name(
+            model_config, "OUTPUT0")
+        output1_config = pb_utils.get_output_config_by_name(
+            model_config, "OUTPUT1")
+
+        self.output0_dtype = pb_utils.triton_string_to_numpy(
+            output0_config['data_type'])
+        self.output1_dtype = pb_utils.triton_string_to_numpy(
+            output0_config['data_type'])
+
+    def execute(self, requests):
+
+        file = open("logs.txt", "w")
+
+        responses = []
+        for request in requests:
+            logger.info("Request: {}".format(request))
+
+            in_0 = pb_utils.get_input_tensor_by_name(request, "INPUT0")
+            in_0 = in_0.as_numpy()
+
+            logger.info("in_0: {}".format(in_0))
+
+            tok_batch = []
+
+            for i in range(in_0.shape[0]):                
+                decoded_object = in_0[i,0].decode()
+
+                logger.info("decoded_object: {}".format(decoded_object))
+
+                tok_batch.append(decoded_object)
+
+            logger.info("tok_batch: {}".format(tok_batch))
+
+            tok_sent = self.tokenizer(tok_batch,
+                                      padding='max_length',
+                                      max_length=128,
+                                     )
+
+
+            logger.info("Tokens: {}".format(tok_sent))
+
+            out_0 = np.array(tok_sent['input_ids'],dtype=self.output0_dtype)
+            out_1 = np.array(tok_sent['attention_mask'],dtype=self.output1_dtype)
+            out_tensor_0 = pb_utils.Tensor("OUTPUT0", out_0)
+            out_tensor_1 = pb_utils.Tensor("OUTPUT1", out_1)
+
+            responses.append(pb_utils.InferenceResponse([out_tensor_0,out_tensor_1]))
+        return responses

inference/nlp/realtime/huggingface/sentence-transformers-triton-ensemble/examples/ensemble_hf/preprocess/config.pbtxt

@@ -0,0 +1,28 @@
+name: "preprocess"
+backend: "python"
+max_batch_size: 16
+
+input [
+  {
+    name: "INPUT0"
+    data_type: TYPE_STRING
+    dims: [ 1 ]
+    
+  }
+]
+output [
+  {
+    name: "OUTPUT0"
+    data_type: TYPE_INT32
+    dims: [ 128 ]
+  },
+  
+  {
+    name: "OUTPUT1"
+    data_type: TYPE_INT32
+    dims: [ 128 ]
+  }
+  
+]
+
+instance_group [{ kind: KIND_CPU }]

inference/nlp/realtime/huggingface/sentence-transformers-triton-ensemble/examples/ensemble_hf/preprocess/requirements.txt

@@ -0,0 +1 @@
+transformers