Machine learning training device
A machine learning training device having a clustering device configured to group machine learning training requests into a consolidated training block. The clustering device includes an isomorphism engine and an energy estimator configured to construct the consolidated training block based upon an amount of energy that will be saved by training the machine learning training requests together, thereby reducing an overall energy consumption by machine learning processes.
1. A machine learning training device having at least one processor and memory and comprising:
a queuing device that receives a plurality of machine learning training requests, wherein each machine learning training request of the plurality of machine learning training requests includes a number of iterations used to complete training of a machine learning model;
a clustering device that consolidates two or more machine learning training requests selected from the plurality of machine learning training requests received by the queuing device to generate a consolidated training block to be processed as a single machine learning training process, the clustering device further comprising:
an isomorphism engine that implements a comparison process between the two or more machine learning training requests, wherein the comparison includes executing a subgraph isomorphism algorithm that compares formats, data types, and computer architecture formats of the two or more machine learning training requests to determine compatibility for consolidation;
an energy estimator that calculates an estimate of an amount of energy saved by clustering the two or more machine learning training requests into the consolidated training block, wherein the amount of energy saved is based at least in part on a number of iterations associated with each machine learning training request, and wherein the clustering device consolidates the two or more machine learning training requests to generate the consolidated training block responsive to determining that the amount of energy saved by clustering the two or more machine learning training requests into the consolidated training block is greater than at least a threshold amount of energy savings; and
a verification engine that generates feedback to iteratively improve subsequent consolidating of two or more machine learning training requests;
a training engine, in communication with an array of graphics processing units, that executes one or more machine learning processes on the consolidated training block by distributing training of the machine learning model based on the consolidated training block across the array of graphics processing units and outputs a set of model weights converted from 32-bit floating-point data types to 16-bit floating-point data types as shared model weights configured to be shared between the two or more machine learning training requests of the consolidated training block; and
a notification engine that outputs a status of the training of the machine learning model based on the consolidated training block.
2. The machine learning training device of claim 1 , wherein the clustering device analyzes attributes of a selected machine learning training request of the plurality of machine learning training requests received and stored by the queuing device, including:
identifying a data type of the selected machine learning training request;
identifying a target type of the selected machine learning training request; and
identifying architectural characteristics of the selected machine learning training request.
3. The machine learning training device of claim 1 , wherein the clustering device consolidates the two or more machine learning training requests to generate the consolidated training block based upon a priority and a data sensitivity of the two or more machine learning training requests.
4. The machine learning training device of claim 1 , wherein the isomorphism engine outputs a compatibility score.
5. The machine learning training device of claim 4 , wherein the compatibility score is configured to indicate whether the two or more machine learning training requests can be clustered to generate the consolidated training block.
6. The machine learning training device of claim 1 , further comprising a quantization engine that reduces an amount of memory used to store the shared model weights.
7. The machine learning training device of claim 1 , wherein the energy estimator estimates an amount of carbon dioxide (CO2) saved by clustering the two or more machine learning training requests to generate the consolidated training block.
8. The machine learning training device of claim 1 , wherein the verification engine validates an effectiveness of the clustering of the consolidated training block and outputs a validation score to the clustering device.
9. A machine learning training device comprising:
a processor;
a non-transitory computer-readable medium comprising computer-executable instructions that, when executed by the processor, are configured to:
receive a plurality of machine learning training requests, wherein each machine learning training request of the plurality of machine learning training requests includes a number of iterations used to complete training of a machine learning model;
execute, by an isomorphism engine, a subgraph isomorphism algorithm comparing formats, data types, and computer architecture formats of two or more machine learning training requests of the plurality of machine learning training requests to determine compatibility for consolidation;
estimate an amount of energy saved by consolidating the two or more machine learning training requests, wherein the amount of energy saved is based at least in part on a number of iterations associated with each machine learning training request;
determine that the amount of energy saved at least meets a threshold;
responsive to determining that the amount of energy saved at least meets the threshold, consolidate the two or more machine learning training requests to generate a consolidated training block;
generate, based on the consolidated two or more machine learning training requests, feedback;
iteratively update evaluation of subsequent consolidating of two or more machine learning training requests based on the feedback;
execute one or more machine learning processes on the consolidated training block by distributing training of the machine learning model based on the consolidated training block across an array of graphics processing units, and output a set of model weights converted from 32-bit floating-point data types to 16-bit floating-point data types as shared model weights configured to be shared between the two or more machine learning training requests of the consolidated training block; and
output a status of the training of the machine learning model based on the consolidated training block.
10. The machine learning training device of claim 9 , wherein the computer-executable instructions, when executed by the processor, are further configured to analyze attributes of a selected machine learning training request of the plurality of machine learning training requests, including:
identifying a data type of the selected machine learning training request;
identifying a target type of the selected machine learning training request; and
identifying architectural characteristics of the selected machine learning training request.
11. The machine learning training device of claim 9 , wherein the computer-executable instructions, when executed by the processor, are further configured to consolidate the two or more machine learning training requests to generate the consolidated training block based upon a priority and a data sensitivity of the two or more machine learning training requests.
12. The machine learning training device of claim 9 , wherein the isomorphism engine outputs a compatibility score.
13. The machine learning training device of claim 12 , wherein the compatibility score is configured to indicate whether the two or more machine learning training requests can be clustered to generate the consolidated training block.
14. The machine learning training device of claim 9 , wherein the computer-executable instructions, when executed by the processor, are further configured to utilize discrete quantization to reduce an amount of memory used to store the shared model weights.
15. The machine learning training device of claim 9 , wherein the computer-executable instructions, when executed by the processor, are further configured to estimate an amount of CO2 saved by clustering the two or more machine learning training requests to generate the consolidated training block.
16. The machine learning training device of claim 9 , wherein the computer-executable instructions, when executed by the processor, are further configured to output a validation score to a clustering device.
17. A method for reducing energy consumption for training machine learning models, comprising:
receiving a plurality of machine learning training requests, wherein each machine learning training request of the plurality of machine learning training requests includes a number of iterations used to complete training of a machine learning model;
executing, by an isomorphism engine, a subgraph isomorphism algorithm comparing formats, data types, and computer architecture formats of two or more machine learning training requests of the plurality of machine learning training requests to determine compatibility for consolidation;
estimating an amount of energy saved by consolidating the two or more machine learning training requests, wherein the amount of energy saved is based at least in part on a number of iterations associated with each machine learning training request;
determining that the amount of energy saved at least meets a threshold;
responsive to determining that the amount of energy saved at least meets the threshold,
consolidating the two or more machine learning training requests to generate a consolidated training block; generating, based on the consolidated two or more machine learning training requests, feedback;
iteratively updating evaluation of subsequent consolidating of two or more machine learning training requests based on the feedback;
executing one or more machine learning processes on the consolidated training block by distributing training of the machine learning model based on the consolidated training block across an array of graphics processing units, and outputting a set of model weights converted from 32-bit floating-point data types to 16-bit floating-point data types as shared model weights configured to be shared between the two or more machine learning training requests of the consolidated training block; and
outputting a status of the training of the machine learning model based on the consolidated training block.
18. The method of claim 17 , further comprising:
analyzing attributes of a selected machine learning training request of the plurality of machine learning training requests, including:
identifying a data type of the selected machine learning training request;
identifying a target type of the selected machine learning training request; and
identifying architectural characteristics of the selected machine learning training request.