Concurrent machine learning model prediction on edge devices
In some examples, a method includes receiving a request to load a first machine learning model on an edge device, wherein the request includes an external identifier of the first machine learning model. The method further includes loading at least one instance of the first machine learning model onto the edge device, wherein the first machine learning model is loaded into a model pool having at least a second machine learning model using the same external identifier as the first machine learning model.
1 . A computer-implemented method comprising:
receiving a request to load a first machine learning (ML) model into a ML model pool of an edge device, wherein the request includes an external identifier of the first ML model that identifies the ML model pool;
in response to the request, the edge device executing edge compute code using one or more processors to load at least one instance of the first ML model into the ML model pool of the edge device, wherein the ML model pool has at least a second ML model using the external identifier and each of the first and second ML models have different internal aliases; and
a runtime orchestrator executed by the one or more processors of the edge device handling translation of prediction requests to a model layer to underlying runtime application programming interfaces (APIs).
2 . The computer-implemented method of claim 1 , wherein the first ML model is an update of the second ML model wherein only weights of the ML models are different.
3 . The computer-implemented method of claim 1 , wherein the request includes one or more of: an identification of a model handle, an identification of a model alias, an indication of a location of a model and/or its runtime, an indication of input and/or output buffer sizes, an indication of a desired execution environment, and an indication of a number of instances to load.
4 . A computer-implemented method comprising:
receiving a request to load a first machine learning (ML) model on an edge device, wherein the request includes an external identifier of the first ML model;
in response to the request, the edge device executing edge compute code using one or more processors to load at least one instance of the first ML model onto the edge device, wherein the first ML model is loaded into a ML model pool having at least a second ML model using the external identifier and each of the first and second ML models have different internal aliases; and
a runtime orchestrator executed by the one or more processors of the edge device handling translation of prediction requests to a model layer to underlying runtime application programming interfaces (APIs).
5 . The computer-implemented method of claim 4 , wherein the first ML model is an update of the second ML model wherein only weights of the ML models are different.
6 . The computer-implemented method of claim 4 , wherein the first ML model has a different architecture than the second ML model.
7 . The computer-implemented method of claim 4 , further comprising:
receiving data from at least one application to use in a prediction using the first ML model according to a prediction request, wherein the prediction request identifies the external identifier and not the internal alias of the first ML model; and
performing inference using the first ML model.
8 . The computer-implemented method of claim 4 , further comprising:
receiving data from at least one application to use in a prediction using the first ML model according to a prediction request, wherein the prediction request identifies the external identifier and not the internal alias of the first ML model;
pre-processing the data to fit the first ML model; and
performing inference using the first ML model.
9 . The computer-implemented method of claim 4 , further comprising unloading the second ML model upon loading the first ML model.
10 . The computer-implemented method of claim 4 , wherein the request includes one or more of: an identification of a model handle, an identification of a model alias, an indication of a location of a model and/or its runtime, an indication of input and/or output buffer sizes, an indication of a desired execution environment, and an indication of a number of instances to load.
11 . The computer-implemented method of claim 4 , wherein return status codes for the request include at least one of load is successful, an unknown error has occurred, an internal error has occurred, the first ML model does not exist, the first ML model already exists, memory is not available to load the first ML model, or the first ML model is not compiled for the machine.
12 . The computer-implemented method of claim 4 , further comprising batching tensor data to be used by the first ML model during inferencing.
13 . The computer-implemented method of claim 4 , further comprising performing inference using the first ML model and the second ML model.
14 . The computer-implemented method of claim 4 , wherein the request to load the first ML model on the edge device is received by a provider network to invoke loading on the edge device.
15 . A system comprising:
a first one or more electronic devices to implement an edge device service in a multi-tenant provider network; and
a second one or more electronic devices to implement edge compute software on an edge device, the edge compute software including instructions that upon execution by one or more processors cause the edge device to:
receive a request from the edge device service to load a first machine learning (ML) model on the edge device, wherein the request includes an external identifier of the first ML model;
in response to the request, execute the edge compute software to load at least one instance of the first ML model onto the edge device, wherein the first ML model is loaded into a ML model pool having at least a second ML model using the external identifier and each of the first and second ML models have different internal aliases; and
a runtime orchestrator executed by the one or more processors of the edge device handling translation of prediction requests to a model layer to underlying runtime application programming interfaces (APIs).
16 . The system of claim 15 , wherein the request includes one or more of: an identification of a model handle, an identification of a model alias, an indication of a location of a model and/or its runtime, an indication of input and/or output buffer sizes, an indication of a desired execution environment, and an indication of a number of instances to load.
17 . The system of claim 15 , wherein the first ML model is an update of the second ML model wherein only weights of the ML models are different.
18 . The system of claim 15 , wherein the first ML model has a different architecture than the second ML model.
19 . The system of claim 15 , wherein return status codes for the request include at least one of load is successful, an unknown error has occurred, an internal error has occurred, the first ML model does not exist, the first ML model already exists, memory is not available to load the first ML model, or the first ML model is not compiled for the machine.
20 . The system of claim 15 , wherein the edge compute software includes further instructions that upon execution by the one or more processors further cause the edge device to perform inference using the first ML model and the second ML model.