TorchServe linux-aarch64 experimental support (#3071)

* Changes for building TorchServe on linux aarch64

* Changes for building TorchServe on linux aarch64

* Added an example for linux aarch64

* Doc update for linux aarch64

* Doc update for linux aarch64

* Doc update for linux aarch64

* removed torchtext for aarch64

* lint failure

* lint failure

* Build conda binaries

* Build conda binaries

* resolving merge conflicts

* resolving merge conflicts

* update documentation

* review comments

* Updated based on review comments

---------

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

binaries/conda/build_packages.py

@@ -22,7 +22,13 @@
 PACKAGES = ["torchserve", "torch-model-archiver", "torch-workflow-archiver"]
 
 # conda convert supported platforms https://docs.conda.io/projects/conda-build/en/stable/resources/commands/conda-convert.html
-PLATFORMS = ["linux-64", "osx-64", "win-64", "osx-arm64"]  # Add a new platform here
+PLATFORMS = [
+    "linux-64",
+    "osx-64",
+    "win-64",
+    "osx-arm64",
+    "linux-aarch64",
+]  # Add a new platform here
 
 if os.name == "nt":
     # Assumes miniconda is installed in windows

docs/linux_aarch64.md

@@ -0,0 +1,29 @@
+# TorchServe on linux aarch64 - Experimental
+
+TorchServe has been tested to be working on linux aarch64 for some of the examples.
+- Tested this on Amazon Graviton 3 instance(m7g.4x.large)
+
+## Installation
+
+Currently installation from PyPi or installing from source works
+
+```
+python ts_scripts/install_dependencies.py
+pip install torchserve torch-model-archiver torch-workflow-archiver
+```
+
+## Optimizations
+
+You can also enable this optimizations for Graviton 3 to get an improved performance. More details can be found in this [blog](https://pytorch.org/blog/optimized-pytorch-w-graviton/)
+```
+export DNNL_DEFAULT_FPMATH_MODE=BF16
+export LRU_CACHE_CAPACITY=1024
+```
+
+## Example
+
+This [example](https://github.com/pytorch/serve/tree/master/examples/text_to_speech_synthesizer/SpeechT5) on Text to Speech synthesis was verified to be working on Graviton 3
+
+## To Dos
+- CI
+- Regression tests

examples/text_to_speech_synthesizer/SpeechT5/README.md

@@ -0,0 +1,50 @@
+# Text to Speech synthesis with SpeechT5
+
+This is an example showing text to speech synthesis using SpeechT5 model. This has been verified to work on (linux-aarch64) Graviton 3 instance
+
+While  running this model on `linux-aarch64`, you can enable these optimizations
+
+```
+export DNNL_DEFAULT_FPMATH_MODE=BF16
+export LRU_CACHE_CAPACITY=1024
+```
+More details can be found in this [blog](https://pytorch.org/blog/optimized-pytorch-w-graviton/)
+
+
+## Pre-requisites
+```
+chmod +x setup.sh
+./setup.sh
+```
+
+## Download model
+
+This saves the model artifacts to `model_artifacts` directory
+```
+huggingface-cli login
+python download_model.py
+```
+
+## Create model archiver
+
+```
+mkdir model_store
+
+torch-model-archiver --model-name SpeechT5-TTS --version 1.0 --handler text_to_speech_handler.py --config-file model-config.yaml --archive-format no-archive --export-path model_store -f
+
+mv model_artifacts/* model_store/SpeechT5-TTS/
+```
+
+## Start TorchServe
+
+```
+torchserve --start --ncs --model-store model_store --models SpeechT5-TTS
+```
+
+## Send Inference request
+
+```
+curl http://127.0.0.1:8080/predictions/SpeechT5-TTS -T sample_input.txt  -o speech.wav
+```
+
+This generates an audio file `speech.wav` corresponding to the text in `sample_input.txt`

examples/text_to_speech_synthesizer/SpeechT5/download_model.py

@@ -0,0 +1,14 @@
+from datasets import load_dataset
+from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor
+
+processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
+model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts")
+vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
+
+embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
+
+model.save_pretrained(save_directory="model_artifacts/model")
+processor.save_pretrained(save_directory="model_artifacts/processor")
+vocoder.save_pretrained(save_directory="model_artifacts/vocoder")
+embeddings_dataset.save_to_disk("model_artifacts/speaker_embeddings")
+print("Save model artifacts to directory model_artifacts")

examples/text_to_speech_synthesizer/SpeechT5/model-config.yaml

@@ -0,0 +1,8 @@
+minWorkers: 1
+maxWorkers: 1
+handler:
+  model: "model"
+  vocoder: "vocoder"
+  processor: "processor"
+  speaker_embeddings: "speaker_embeddings"
+  output_dir: "/tmp"

examples/text_to_speech_synthesizer/SpeechT5/sample_input.txt

@@ -0,0 +1 @@
+"I love San Francisco"

examples/text_to_speech_synthesizer/SpeechT5/setup.sh

@@ -0,0 +1,6 @@
+#!/bin/bash
+
+# Needed for soundfile
+sudo apt install libsndfile1 -y
+
+pip install --upgrade transformers sentencepiece datasets[audio] soundfile

examples/text_to_speech_synthesizer/SpeechT5/text_to_speech_handler.py

@@ -0,0 +1,68 @@
+import logging
+import os
+import uuid
+
+import soundfile as sf
+import torch
+from datasets import load_from_disk
+from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor
+
+from ts.torch_handler.base_handler import BaseHandler
+
+logger = logging.getLogger(__name__)
+
+
+class SpeechT5_TTS(BaseHandler):
+    def __init__(self):
+        self.model = None
+        self.processor = None
+        self.vocoder = None
+        self.speaker_embeddings = None
+        self.output_dir = "/tmp"
+
+    def initialize(self, ctx):
+        properties = ctx.system_properties
+        model_dir = properties.get("model_dir")
+
+        processor = ctx.model_yaml_config["handler"]["processor"]
+        model = ctx.model_yaml_config["handler"]["model"]
+        vocoder = ctx.model_yaml_config["handler"]["vocoder"]
+        embeddings_dataset = ctx.model_yaml_config["handler"]["speaker_embeddings"]
+        self.output_dir = ctx.model_yaml_config["handler"]["output_dir"]
+
+        self.processor = SpeechT5Processor.from_pretrained(processor)
+        self.model = SpeechT5ForTextToSpeech.from_pretrained(model)
+        self.vocoder = SpeechT5HifiGan.from_pretrained(vocoder)
+
+        # load xvector containing speaker's voice characteristics from a dataset
+        embeddings_dataset = load_from_disk(embeddings_dataset)
+        self.speaker_embeddings = torch.tensor(
+            embeddings_dataset[7306]["xvector"]
+        ).unsqueeze(0)
+
+    def preprocess(self, requests):
+        assert len(requests) == 1, "This is currently supported with batch_size=1"
+        req_data = requests[0]
+
+        input_data = req_data.get("data") or req_data.get("body")
+
+        if isinstance(input_data, (bytes, bytearray)):
+            input_data = input_data.decode("utf-8")
+
+        inputs = self.processor(text=input_data, return_tensors="pt")
+
+        return inputs
+
+    def inference(self, inputs):
+        output = self.model.generate_speech(
+            inputs["input_ids"], self.speaker_embeddings, vocoder=self.vocoder
+        )
+        return output
+
+    def postprocess(self, inference_output):
+        path = self.output_dir + "/{}.wav".format(uuid.uuid4().hex)
+        sf.write(path, inference_output.numpy(), samplerate=16000)
+        with open(path, "rb") as output:
+            data = output.read()
+        os.remove(path)
+        return [data]

requirements/developer.txt

@@ -15,7 +15,7 @@ pre-commit==3.3.2
 twine==4.0.2
 mypy==1.3.0
 torchpippy==0.1.1
-intel_extension_for_pytorch==2.2.0; sys_platform != 'win32' and sys_platform != 'darwin'
+intel_extension_for_pytorch==2.2.0; sys_platform != 'win32' and sys_platform != 'darwin' and platform_machine != 'aarch64'
 onnxruntime==1.17.1
 googleapis-common-protos
 onnx==1.16.0

requirements/torch_linux_aarch64.txt

@@ -0,0 +1,6 @@
+#pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cpu
+--extra-index-url https://download.pytorch.org/whl/cpu
+-r torch_common.txt
+torch==2.2.1; sys_platform == 'linux' and platform_machine == 'aarch64'
+torchvision==0.17.1; sys_platform == 'linux' and platform_machine == 'aarch64'
+torchaudio==2.2.1; sys_platform == 'linux' and platform_machine == 'aarch64'

ts_scripts/install_dependencies.py

@@ -118,9 +118,14 @@ def install_torch_packages(self, cuda_version):
                 f"{sys.executable} -m pip install -U -r {torch_neuronx_requirements_file}"
             )
         else:
-            os.system(
-                f"{sys.executable} -m pip install -U -r requirements/torch_{platform.system().lower()}.txt"
-            )
+            if platform.machine() == "aarch64":
+                os.system(
+                    f"{sys.executable} -m pip install -U -r requirements/torch_{platform.system().lower()}_{platform.machine()}.txt"
+                )
+            else:
+                os.system(
+                    f"{sys.executable} -m pip install -U -r requirements/torch_{platform.system().lower()}.txt"
+                )
 
     def install_python_packages(self, cuda_version, requirements_file_path, nightly):
         check = "where" if platform.system() == "Windows" else "which"

ts_scripts/spellcheck_conf/wordlist.txt

@@ -1216,6 +1216,10 @@ libomp
 rpath
 venv
 TorchInductor
+Graviton
+aarch
+linux
+SpeechT
 Pytests
 deviceType
 XGBoost