Compression-guided task decomposition for machine learning models

Abstract

Provided are systems and methods that can enhance the efficiency of sequence processing models, such as Large Language Models (LLMs), by intelligently decomposing model prompts into subprompts which are then processed by a number of different processing systems and / or models. In particular, a computing system can decompose the model into subprompts based on the compressibility of each potential subprompt. In this way, compression benefits can be achieved because each subprompt is sent to the appropriate processing system and model based on its compressibility. For example, this selective processing can optimize the use of computational resources by ensuring that data-intensive tasks are handled locally whenever possible, while more complex tasks that require less data transfer can be offloaded to remote, more powerful models.

Description

COMPRESSION-GUIDED TASK DECOMPOSITION FOR MACHINE LEARNINGMODELSFIELD

[0001] The present disclosure relates generally to machine learning processes and machine-learned devices and systems. More particularly, the present disclosure relates to compression-guided task decomposition for improved computational efficiency of machine learning models.BACKGROUND

[0002] Sequence processing models are machine-learned models that can be utilized to handle and analyze sequences of data. These models can be designed to process a variety of data types including text, images, audio, and biochemical sequences, among others. For instance, in the text domain, some example sequence processing models are often referred to as Large Language Models (LLMs). Depending on the application, these models can vary in size from large, computationally expensive models to smaller, more efficient ones. They may also be capable of processing multiple types of data simultaneously, which adds a layer of complexity and versatility to their functionality.

[0003] Recent technological advancements have significantly increased the context sizes that these models can handle, thereby broadening their potential applications. For example, some state-of-the-art models now support context sizes up to 128,000 tokens and show promising results for sizes up to 1 million tokens. This enhancement allows for the inclusion of extensive data sets, such as two-hour-long videos, directly in the prompts for analysis and interaction. However, these large models are typically hosted on servers, necessitating substantial data uploads to leverage their capabilities fully.

[0004] The requirement to upload large volumes of data to servers poses several challenges. Users with poor internet connections may experience slow processing times, making the use of sequence processing models less feasible. Additionally, the cost associated with data transmission can become prohibitive as each megabyte uploaded incurs greater computational expense. Moreover, while the ability to process larger context windows offers exciting new possibilities, it also demands judicious use of these capabilities to avoid unnecessary computational load on the server. Therefore, there is a pressing need for amethod that enables processing of large context windows while also preserving bandwidth and reducing server load.SUMMARY

[0005] A system of one or more computers can be configured to perform particular operations or actions by virtue of having software, firmware, hardware, or a combination of them installed on the system that in operation causes or cause the system to perform the actions. One or more computer programs can be configured to perform particular operations or actions by virtue of including instructions that, when executed by data processing apparatus, cause the apparatus to perform the actions.

[0006] One general aspect includes a computer-implemented method to compression- guided decomposition of model prompts for distributed processing. The computer- implemented method includes obtaining, by a computing system may include one or more computing devices, a model prompt associated with a context window. The method also includes decomposing, by the computing system, the model prompt into a plurality of subprompts, where decomposing the model prompt may include assigning, based at least in part on a compressibility of the subprompt, each of the plurality of subprompts to one of a plurality of different processing systems respectively may include a plurality of different machine-learned models with different computational characteristics. The method also includes causing, by the computing system, each of the plurality of subprompts to be respectively processed by the different processing system to which the subprompt has been mapped to generate a plurality of suboutputs. The method also includes aggregating, by the computing system, the plurality of suboutputs to generate a response to the model prompt. Other embodiments of this aspect include corresponding computer systems, apparatus, and computer programs recorded on one or more computer storage devices, each configured to perform the actions of the methods.

[0007] Example implementations may include one or more of the following features. The computer-implemented method of any preceding claim, where: the model prompt may include a multimodal prompt that may include data associated with at least two different modalities, and decomposing, by the computing system, the model prompt into the plurality of subprompts may includes decomposing, by the computing system, the model prompt into the plurality of subprompts based at least in part on transitions in the data between the at least two different data modalities. The plurality of different processing systems may include: a local processing system may include a first machine-learned model with a relatively lowercomputational cost; and a remote processing system may include a second machine-learned model with a relatively higher computational cost. Decomposing, by the computing system, the model prompt into the plurality of subprompts may include: performing, by the computing system, a decomposition algorithm on the model prompt to generate a plurality of candidate decompositions, where each of the plurality of candidate decompositions may include a different set of candidate subprompts; generating, by the computing system, a plurality of scores respectively for the plurality of candidate decompositions according to one or more scoring criteria; and selecting, by the computing system, one of the candidate decompositions based on the plurality of scores, where the plurality of subprompts are extracted from the selected candidate decomposition. The decomposition algorithm may assign each of the candidate subprompts to one of a plurality of classes respectively associated with the plurality of different processing systems; and the one or more scoring criteria may compare the compressibility of the candidate subprompt to the class into which such candidate subprompt is assigned. Decomposing, by the computing system, the model prompt into the plurality of subprompts may include decomposing, by the computing system, the model prompt into a hierarchical decomposition tree that encodes hierarchical interdependence between two or more of the plurality of subprompts. Decomposing, by the computing system, the model prompt into the hierarchical decomposition tree may include applying a constraint that weights inter-dependent nodes to be mapped to the same processing system. Decomposing, by the computing system, the model prompt into the plurality of subprompts may include evaluating a compressibility of each of a plurality of candidate subprompts. Evaluating the compressibility of at least one of the candidate subprompts may include subsampling data from the candidate subprompt and evaluating a compressibility of the subsampled data. Decomposing, by the computing system, the model prompt into the plurality of subprompts may include evaluating an access control status of each of a plurality of candidate subprompts. Evaluating the compressibility of at least one of the candidate subprompts may include processing at least a portion of the candidate subprompt with a machine-learned compression prediction model to obtain a compressibility prediction as an output of the machine-learned compression prediction model. Evaluating the compressibility of at least one of the candidate subprompts may include evaluating one or more file formats associated with the candidate subprompt. Implementations of the described techniques may include hardware, a method or process, or computer software on a computer-accessible medium.

[0008] One general aspect includes a computing system. The computing system also includes one or more processors. The system also includes one or more non-transitory computer-readable media that collectively store instructions that, when executed by the one or more processors, cause the computing system to perform operations. The operations include obtaining a model prompt associated with a context window. The operations include decomposing the model prompt into a plurality of subprompts, where decomposing the model prompt may include assigning, based at least in part on a compressibility of the subprompt, each of the plurality of subprompts to one of a plurality of different processing systems respectively may include a plurality of different machine-learned models with different computational characteristics. The operations include includes causing each of the plurality of subprompts to be respectively processed by the different processing system to which the subprompt has been mapped to generate a plurality of suboutputs. The operations include aggregating the plurality of suboutputs to generate a response to the model prompt. Other embodiments of this aspect include corresponding computer systems, apparatus, and computer programs recorded on one or more computer storage devices, each configured to perform the actions of the methods.

[0009] Example implementations may include one or more of the following features. The computing system where: the model prompt may include a multimodal prompt that may include data associated with at least two different modalities, and decomposing the model prompt into the plurality of subprompts may include decomposing the model prompt into the plurality of subprompts based at least in part on transitions in the data between the at least two different data modalities. The plurality of different processing systems may include: a local processing system may include a first machine-learned model with a relatively lower computational cost; and a remote processing system may include a second machine-learned model with a relatively higher computational cost. Decomposing the model prompt into the plurality of subprompts may include: performing a decomposition algorithm on the model prompt to generate a plurality of candidate decompositions, where each of the plurality of candidate decompositions may include a different set of candidate subprompts; generating a plurality of scores respectively for the plurality of candidate decompositions according to one or more scoring criteria; and selecting one of the candidate decompositions based on the plurality of scores, where the plurality of subprompts are extracted from the selected candidate decomposition. The decomposition algorithm may assign each of the candidate subprompts to one of a plurality of classes respectively associated with the plurality of different processing systems; and the one or more scoring criteria may compare thecompressibility of the candidate subprompt to the class into which such candidate subprompt is assigned. Decomposing the model prompt into the plurality of subprompts may include decomposing the model prompt into a hierarchical decomposition tree that encodes hierarchical interdependence between two or more of the plurality of subprompts. Decomposing the model prompt into the hierarchical decomposition tree may include applying a constraint that weights inter-dependent nodes to be mapped to the same processing system. Implementations of the described techniques may include hardware, a method or process, or computer software on a computer-accessible medium.

[0010] One general aspect includes one or more non-transitory computer-readable media that collectively store instructions for performing operations. The operations include obtaining a model prompt associated with a context window. The operations include decomposing the model prompt into a plurality of subprompts, where decomposing the model prompt may include assigning, based at least in part on a compressibility of the subprompt, each of the plurality of subprompts to one of a plurality of different processing systems respectively may include a plurality of different machine-learned models with different computational characteristics. The operations include causing each of the plurality of subprompts to be respectively processed by the different processing system to which the subprompt has been mapped to generate a plurality of suboutputs. The operations include aggregating the plurality of suboutputs to generate a response to the model prompt. Other embodiments of this aspect include corresponding computer systems, apparatus, and computer programs recorded on one or more computer storage devices, each configured to perform the actions of the methods.BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 is a block diagram of an example computing system for performing compression-guided task decomposition according to example implementations of aspects of the present disclosure;

[0012] Figure 2 is a block diagram of an example computing system for performing compression-guided task decomposition according to example implementations of aspects of the present disclosure;

[0013] Figure 3 is a flow chart diagram of an example method for performing compression-guided task decomposition according to example implementations of aspects of the present disclosure;

[0014] Figure 4 is a flow chart diagram illustrating an example method for training a machine-learned model according to example implementations of aspects of the present disclosure;

[0015] Figure 5 is a block diagram of an example processing flow for using machine- learned model(s) to process input(s) to generate output(s) according to example implementations of aspects of the present disclosure;

[0016] Figure 6 is a block diagram of an example sequence processing model according to example implementations of aspects of the present disclosure;

[0017] Figure 7 is a block diagram of an example technique for populating an example input sequence for processing by a sequence processing model according to example implementations of aspects of the present disclosure;

[0018] Figure 8 is a block diagram of an example model development platform according to example implementations of aspects of the present disclosure;

[0019] Figure 9 is a block diagram of an example training workflow for training a machine-learned model according to example implementations of aspects of the present disclosure;

[0020] Figure 10 is a block diagram of an inference system for operating one or more machine-learned model(s) to perform inference according to example implementations of aspects of the present disclosure;

[0021] Figure 11 is a block diagram of an example networked computing system according to example implementations of aspects of the present disclosure;

[0022] Figure 12 is a block diagram of an example computing device according to example implementations of aspects of the present disclosure; and

[0023] Figure 13 is a block diagram of an example computing device according to example implementations of aspects of the present disclosure.DETAILED DESCRIPTION

[0024] Example aspects of the present disclosure are directed to systems and methods that can enhance the efficiency of sequence processing models, such as Large Language Models (LLMs), by intelligently decomposing model prompts into subprompts which are then processed by a number of different processing systems and / or models. In particular, a computing system can decompose the model into subprompts based on the compressibility of each potential subprompt. In this way, compression benefits can be achieved because each subprompt is sent to the appropriate processing system and model based on itscompressibility. For example, this selective processing can optimize the use of computational resources by ensuring that data-intensive tasks are handled locally whenever possible, while more complex tasks that require less data transfer can be offloaded to remote, more powerful models. This method can be particularly beneficial when managing large context windows, which are known to be computationally expensive and data-intensive. By applying compression constraints when performing the task decomposition, the system can evaluate and determine the most appropriate processing system for each subprompt, thereby potentially reducing both computational and communication costs. The proposed techniques can also reduce latency. For example, handling large but simple subprompts locally reduces unnecessary latency introduced in the form of data uploading, while handling small but difficult parts on the server reduces latency from a slow model call on the limited, local computational resources.

[0025] Thus, a computing system can obtain a model prompt associated with a context window and can decompose the obtained model prompt into multiple subprompts. This decomposition can be based on the compressibility of each candidate subprompt, allowing each subprompt to be mapped to different processing systems with varying computational characteristics. These systems can range from local models with lower computational costs to remote models that are more computationally intensive but can handle more complex tasks. The subprompts are then processed by the appropriate system, and the results are aggregated to generate a response to the original model prompt.

[0026] More particularly, in some implementations, the proposed technique can begin when a user prompts the sequence processing model (or a related system that serves as a handler or wrapper for the model) with a specific query or request. For example, the user may input a lengthy text document, a series of images, or a combination of audio and video files, which the model needs to analyze or respond to. Depending on the complexity and the data types involved, the system can determine the appropriate method for processing the prompt.

[0027] Once the user has initiated the prompt, the computing system can evaluate the length or “difficulty” of the prompt to decide on the processing strategy. For short or less complex prompts, the system can utilize standard processing techniques that are typically less resource-intensive. For instance, a brief text query might be directly processed by a local, lightweight model without the need for further decomposition. However, for longer or more complex prompts, such as a two-hour video or a large set of interrelated data files, the system may activate a custom decomposition technique. This approach involves breaking down the prompt into smaller, more manageable subprompts, which can then be processed individuallyand / or in parallel, potentially using different models optimized for specific types of data or tasks.

[0028] In some implementations, the decomposition algorithm can parse the context window into subprompts based on various characteristics such as data modality, file start / end, and / or content markers. For instance, the algorithm can parse the context window based on data modality, such as distinguishing between text, image, or audio data, allowing each type to be processed by the most suitable system. Additionally, the algorithm can identify specific file start and end points or use content markers and delimiters to segment the context window effectively. This segmentation can help in isolating different sections of the data that may require distinct processing approaches or have varying levels of importance and compressibility.

[0029] In some implementations, the task decomposition algorithm can leverage heuristics to evaluate and score potential subprompts, categorizing them into predefined classes based on certain criteria. This scoring process can be executed using either a handwritten function or a specialized model tailored to assess the characteristics of each subprompt. For instance, a subprompt containing simple text data might be classified as being appropriate for processing locally, while a subprompt with complex image data might be classified as being appropriate to send to a more powerful remote server for processing.

[0030] As one example, a subprompt that is classified as “high complexity, but low data usage” can be directed to a remote model that specializes in handling complex tasks efficiently. This approach leverages the powerful computational capabilities of the remote model while minimizing data upload requirements.

[0031] As another example, subprompts that are “low complexity, but high data usage” can be processed locally. This strategy can be particularly beneficial in scenarios where data transmission to a remote server is constrained by bandwidth or cost, for example including any cost associated with using an application programming interface (API) to make a call to the remote server-based model. By handling these tasks locally, the system can reduce dependency on network conditions and avoid unnecessary data transfers.

[0032] Additionally, there may be cases where subprompts do not fit neatly into the first two categories. In such situations, subprompts that are “high complexity, but decomposed in a way that enables significant data compression” are also sent to the remote model. This class can include tasks that are challenging but have been decomposed in such a manner that they can be compressed effectively, thus reducing the data load during transmission. This tiered classification system allows the computing system to makeinformed decisions about where and how different subprompts should be processed, balancing the need for computational power and data efficiency.

[0033] In some implementations, the data type of a file can be considered by the heuristics to evaluate its compressibility characteristics. For instance, a file suffix such as “.zip” might indicate that an item is already compressed and may not benefit from further compression efforts. Similarly, other file types like “.mp4” or “.jpeg” might suggest different levels of inherent compressibility due to their respective encoding schemes. This approach allows the computing system to make informed decisions about which compression techniques can be applied, or whether compression should be attempted at all, based on the file type involved. Additionally, this method can be extended to include analysis of metadata associated with the files, which might provide further insights into the compressibility or the priority of the data contained within the file.

[0034] In some implementations, a compression algorithm can be applied to assess the compressibility of portions of the context window in a sequence processing model. This can involve applying the compression algorithm to either some or all of the context window to determine how well the data can be compressed. For example, if the context window includes a large text document, the compression algorithm can evaluate how much this document can be compressed before it is sent to a remote server for further processing. Another example might involve applying the compression algorithm to different types of data within the same context window, such as text and images, to evaluate their individual compressibility.

[0035] In some implementations, a computing system can apply multiple compression algorithms to assess the compressibility of a portion of the context window. This process can be useful for determining the most efficient way to handle data before transmission or processing. For example, the system might compress text data using algorithms like Huffman coding or Lempel-Ziv-Welch (LZW) compression. In the case of image data, techniques such as JPEG or PNG compression can be utilized. The choice of compression algorithm can depend on the data type and the specific requirements for compression efficiency and data integrity.

[0036] In some implementations, subsampling techniques can be used to test the compressibility of data by using only a subset of the data in the context window. This approach can be particularly useful for large datasets where applying a compression algorithm to the entire set would be computationally expensive. For instance, a computing system can randomly select portions of a video file to apply the compression algorithm, thusestimating the overall compressibility of the video without having to process every frame. Similarly, in the case of a large dataset of sensor readings, subsampling can involve selecting periodic entries from the dataset to assess their compressibility before deciding whether to compress and transmit the entire dataset.

[0037] In some implementations, the frame rate of a video can also be considered to evaluate its compressibility. Generally, a higher frame rate might suggest that the video is more compressible. For example, a computing system can analyze the frame rate data of a video file to determine how effectively it can be compressed before transmission or processing. For instance, a video with a frame rate of 60 frames per second might be more amenable to compression algorithms compared to one with a lower frame rate, such as 30 frames per second. Another example might involve the system adjusting the compression ratio based on the frame rate, applying more aggressive compression techniques for higher frame rates to optimize both the quality and the size of the transmitted data.

[0038] In some implementations, a computing system can include and use a relatively smaller machine-learned compressibility estimation model to evaluate the compressibility of a portion of the context window. This model can predict how effectively data can be compressed before it is transmitted or processed. For instance, the model might analyze text, images, or video data to determine if compression can significantly reduce the size without losing critical information. In one example, the machine-learned compressibility estimation model can have been trained on various data types to predict compressibility based on characteristics such as redundancy, data type, and resolution. This functionality is beneficial as it allows the system to make informed decisions about which parts of the data to compress and transmit, optimizing both bandwidth usage and processing time.

[0039] In some implementations, the decomposition algorithm can create a graphical and / or hierarchical representation of subprompts, which may have a tree structure. Each node within this tree can represent a different subprompt, with relationships between nodes indicating dependencies in processing. For example, a parent node might represent a complex task that encompasses several smaller tasks, each represented by a child node. The processing of the parent node’s task may depend on the results from its child nodes. This hierarchical approach can facilitate organized and systematic processing of tasks, ensuring that dependencies are maintained and tasks are completed in an optimal sequence. As one example, this could mean that a computing system first processes the audio of a video to detect speech, and then processes the detected speech segments to extract textual information, with each step corresponding to a node in the decomposition tree.

[0040] In some implementations, the decomposition algorithm can evaluate an access control status of each processing system relative to each candidate subprompt when decomposing the model prompt into multiple subprompts. This evaluation can ensure that each subprompt adheres to specific security and privacy protocols before being processed by an appropriate computing system. By assessing the access control status, the system can determine which subprompts contain sensitive or restricted data and thus require special handling, such as encryption or local processing, to comply with data protection objectives.

[0041] In some implementations, the computing system can assign scores to candidate subprompts based on initial assignments determined by a decomposition algorithm. This scoring can be influenced by various factors such as the complexity of the task, the amount of data involved, and the potential compressibility of the data. For example, a subprompt that involves a complex task with a large amount of easily compressible data might be assigned a higher score if it is assigned to be processed at a remote server. In this case, the higher score indicates that the current assignment is a relatively strong match, as the complex task and high compressibility lend this subtask to remote processing. Conversely, a complex task with a large amount of easily compressible data might be assigned a lower score if it is assigned to be processed locally. In this case, the lower score indicates that the current assignment is a relatively poor match, indicating that the subtask should be processed by a more powerful remote model.

[0042] Furthermore, the computing system can re-weight these scores by applying specific scoring heuristics, enhancing the decision-making process regarding where each subprompt should be processed. This re-weighting can prioritize certain types of decompositions over others, for instance, favoring subprompts that fall into classes 1 and 2 (tasks that are either less complex but data-intensive or very complex but require little data) over class 3 (complex tasks that also require significant data but can be compressed effectively). Such a strategy ensures that the system optimizes both computational resources and data transmission, potentially reducing costs and increasing processing speed. This method can be particularly beneficial in environments where bandwidth is limited or computational resources are costly.

[0043] In addition to basic scoring criteria, the heuristics used in the task decomposition algorithm can also take into account various situational factors that may affect processing. For instance, permissions associated with the data can influence the scoring, as certain subprompts might contain sensitive information that cannot be processed or stored in certain locations. Similarly, current network conditions can be considered, with the algorithmadjusting the relative preference toward local processing for subprompts when bandwidth is limited. This adaptive approach allows the computing system to optimize resource use and maintain performance across varying operational environments.

[0044] The process of decomposition and scoring can be implemented iteratively by a computing system to refine the breakdown of model prompts into subprompts. In some implementations, this iterative process can be executed recursively until certain criteria are met, such as achieving a minimum subprompt size measured in bytes or tokens, or reaching a specified depth within a decomposition tree. For example, a computing system can start by decomposing a large text prompt into smaller textual segments. These segments can then be further decomposed based on complexity or data type until the segments reach a predefined minimum size or until the decomposition achieves a certain level of detail as represented in the hierarchical structure of the decomposition tree.

[0045] Additionally, the scoring of decompositions can be dynamically adjusted during each iteration. For instance, a computing system can apply different scoring heuristics in each iteration based on factors such as the compressibility of subprompts and the computational cost associated with processing each subprompt. This allows the system to prioritize subprompts that are easier to compress and less costly to process, thereby optimizing the overall efficiency of the model. In some implementations, the system might also consider the interdependencies between subprompts, ensuring that related tasks are processed in a manner that minimizes data transfer and maximizes the reuse of computed results.

[0046] In some implementations, the computing system can select the highest scoring decomposition from a set of candidate decompositions based on re-weighted scores that align with predefined compression criteria. This selection process can be implemented using various algorithms that enhance the efficiency and effectiveness of the decomposition. For example, a greedy algorithm can be employed where the system chooses the locally-optimal decomposition at each step of a graphical representation or decomposition tree. This approach ensures that the most suitable subprompt is selected based on immediate available data, potentially speeding up the processing time by avoiding less optimal paths.

[0047] Additionally, in some implementations, the computing system can consider the interdependence between subprompts when selecting the highest scoring decomposition. This consideration can be implemented by introducing constraints that group interdependent nodes, such as those in a parent-child relationship, to be processed by the same processing system or model. For instance, if two subprompts are operationally dependent, the system canensure they are handled by the same local or remote model, thus reducing the latency that might occur if these subprompts were processed on different systems. This method of grouping related subprompts can be particularly beneficial in complex tasks where the processing of one subprompt might directly influence the outcome or the processing requirements of another.

[0048] Thus, each subprompt can be classified into one of several different classes that correspond to various processing options within a distributed processing system. This classification can occur either during the decomposition process or as a post-decomposition task. For instance, a subprompt that requires intensive computational resources but involves minimal data transfer might be classified for processing by a powerful remote model. Conversely, a subprompt that involves a larger volume of data but requires less computational power might be better suited for processing by a local model. This allows the system to optimize both data transfer and processing efficiency by matching subprompts with the most appropriate processing resources.

[0049] For example, in some implementations, a computing system can use metadata or content analysis to determine the class into which a subprompt should be categorized. A subprompt containing high-resolution images, which are large in file size but relatively straightforward in terms of processing, might be directed to local processing to avoid high data transfer costs. On the other hand, a subprompt involving complex natural language processing, which might be computationally intensive but relatively light in data size, could be sent to a more capable remote server.

[0050] In some implementations, a computing system can process subprompts by calling the corresponding model that best fits the characteristics of each subprompt, taking into account the dependencies outlined in the decomposition tree. The system can iteratively solve each subprompt and then integrate the subsolutions to form a complete response to the original prompt. This method can be particularly beneficial in scenarios where different subprompts vary significantly in their processing requirements, such as a mixed-media prompt that includes both text and high-resolution images.

[0051] In some implementations, the dependencies expressed in the decomposition tree can guide the sequence in which subprompts are processed. For example, if a certain subprompt is dependent on the results of another, the system can ensure that the dependent subprompt is processed only after its prerequisite subprompt has been successfully handled. This can be useful for maintaining the logical flow and integrity of the overall task. The system might use a variety of scheduling algorithms to optimize the order of processing, suchas prioritizing subprompts that are prerequisites for multiple other subprompts, or those that are expected to take longer to process. This strategic processing order can help in minimizing overall response time and improving the efficiency of resource utilization.

[0052] The computing system can deliver an aggregate output to the user after processing subprompts through different models based on their complexity and data requirements. This can include aggregating the outputs generated from various subprompts, which have been processed either locally or remotely, to form a coherent response to the initial query or prompt. For example, if a user submits a large video file, the system can decompose the video into segments, analyze the segments that require intensive processing remotely, handle less complex segments locally, and then combine the results to provide the aggregate output to the user.

[0053] The aggregation process can be implemented in several ways depending on the complexity and type of the initial prompt. For instance, a computing system can use a hierarchical aggregation approach where results from child nodes in the decomposition tree are progressively combined until the final solution is synthesized at the root. Alternatively or additionally, parallel processing can be employed, where independent subprompts are processed simultaneously and their results are merged upon completion. This method can significantly reduce the response time, making the system more efficient in handling extensive or complex prompts.

[0054] In some implementations, a computing system can maintain a cached representation of a decomposed context window to enhance efficiency in processing repeated or similar queries. This caching mechanism allows the system to avoid redundant decompositions and computations, thereby saving computational resources and time. For instance, if a user submits a query that only modifies a small portion of a previously submitted context window, the system can decompose only the new or changed portion. The remainder of the context can be handled using the cached decomposition data from the previous query.

[0055] Moreover, in scenarios where certain parts of the context window are frequently used across multiple queries, the outputs or subsolutions for these parts can be stored and reused. For example, if a sequence processing model frequently encounters a specific video segment or text passage in various user prompts, the system can cache the results of processing this segment or passage. Subsequent queries involving this segment or passage can then utilize the cached results, thereby bypassing the need for reprocessing the same data. This approach not only speeds up the response time for the user but alsominimizes the load on the processing servers, making the system more efficient and responsive.

[0056] In some implementations, the disclosed technology can be applied in scenarios where a user prompts the model with a long video. In this example, not all parts of the video may be relevant to the query, but the relevant segments may require sophisticated analysis that is beyond the capabilities of a local model. The computing system can decompose the video into subprompts, identify the segments that are useful for the query, and then compress these segments before sending them to a more powerful remote model for processing. This selective uploading and processing can significantly reduce the bandwidth and computational resources required, making the system more efficient and responsive.

[0057] Another application of the disclosed technology involves handling large images. For instance, when a user uploads a large photo, the computing system can implement a task decomposition strategy by first cropping the image into smaller, relevant sections. Following this, the system can generate specific subquestions related to each cropped section. These cropped images and their corresponding subquestions can then be compressed optimally before being sent to a remote model for further analysis. This method not only conserves bandwidth by reducing the amount of data that needs to be transmitted but also focuses the computational efforts of the remote model on the most pertinent information, thereby enhancing overall processing efficiency.

[0058] One of the technical benefits of the disclosed technology is the potential to decrease communication costs. By decomposing model prompts into subprompts that can be selectively processed either locally or remotely, the volume of data that needs to be transmitted over networks can be significantly reduced. For example, a computing system can evaluate the complexity and data intensity of each subprompt and decide to process simpler, less data-intensive subprompts on local machines. Conversely, only subprompts that are complex and require more powerful computational capabilities are sent to remote servers. This selective processing can be particularly beneficial in scenarios where data transmission costs are high or network bandwidth is limited.

[0059] In some implementations, the computing system can further optimize communication costs by compressing subprompts before they are transmitted to remote processing systems. The system can use various compression algorithms tailored to the specific data types of the subprompts. For instance, video data might be compressed using a different algorithm than text data. Additionally, the system can evaluate the compressibility of each subprompt and choose the most effective compression method for each, therebyminimizing the amount of data that needs to be sent over the network and further reducing communication costs.

[0060] Another example technical benefit of the disclosed method is the decreased computational cost achieved by reducing the frequency at which a large model operates over the entirety of a large context window. In sequence processing models, operations such as attention mechanisms scale with the size of the context, which can be computationally intensive. By decomposing the model prompt into smaller, manageable subprompts, the system can limit the scope of data each model processes at a time, thus conserving computational resources. For example, instead of processing a lengthy document in one go, a computing system can break the document into paragraphs or sections and process each independently. This method not only reduces the load on the server but also allows for more efficient use of computational resources by focusing on smaller, more relevant chunks of data at a time.

[0061] In some implementations, this approach can be further optimized by selectively processing only those subprompts that require the capabilities of a more powerful model. For instance, a local model may handle straightforward queries or data segments, while only complex or critical subprompts are sent to a more robust, remote model. This selective processing ensures that computational efforts are concentrated where they are most needed, further enhancing efficiency. Additionally, by distributing the processing load across multiple models with varying computational powers, the system can balance the workload and prevent any single model from becoming a bottleneck, thereby streamlining the overall processing pipeline.

[0062] Another example technical benefit includes enhancing privacy protection by enabling local processing of certain subproblems that are sensitive or privacy-related. For example, a computing system can implement this concept by identifying subprompts that contain sensitive data and processing these subprompts locally rather than sending them to a remote server. This method can be further refined by using a classification algorithm that automatically detects privacy-sensitive data within the prompt and routes it for local processing. Additionally, the system can use encryption techniques to secure data that must be transmitted, ensuring that even if data must leave the local environment, its confidentiality is preserved.

[0063] Various example implementations are described herein with respect to the accompanying Figures.

[0064] Figure 1 illustrates an example computing system 100 that can implement the disclosed technology. The illustrated system 100 includes a prompt handling system 104 and multiple different processing systems 106, 116, and 126. The prompt handling system 104 can receive a model prompt 102 and cooperatively interact with the different processing systems 106, 116, and 126 to generate an aggregate output 130.

[0065] The model prompt 102 can be any type of data that can be processed by a computing system. For example, the model prompt 102 can include text, audio, images, video, sensor data, and / or any other type of data that can be processed by a computing system. The model prompt 102 can be a prompt received from a user, a prompt received from a different model or system, or a prompt generated by the prompt handling system 104.

[0066] The prompt handling system 104 can be any type of computing system that can process the model prompt 102 and generate an aggregate output 130. The aggregate output 130 can be any type of output that can be generated based on the model prompt 102. For example, the aggregate output 130 can include a response to the model prompt 102, a summary of the model prompt 102, a classification of the model prompt 102, or any other type of output that can be generated based on the model prompt 102.

[0067] Each processing system 106, 116 and 126 can be any type of computing system that can process all or a portion of the model prompt 102. For example, each processing system 106, 116 and 126 can be a local computing system, a remote computing system, or a combination of a local and a remote computing system.

[0068] The processing systems 106, 116, and 126 can optionally be interacted with using APIs. The API can be any type of API that can be used to interact with the processing systems 106, 116, and 126. For example, the API can be a REST API, a SOAP API, a JSON API, or any other type of API that can be used to interact with the processing systems 106, 116, and 126.

[0069] Each processing system 106, 116 and 126 can include one or more different machine-learned models that offer different processing capabilities. For example, the processing system 106 can include machine-learned models 108a, 108b, and 108c, the processing system 116 can include machine-learned models 118a and 118b, and the processing system 126 can include a machine-learned model 128. Each of the different models 108a, 108b, 108c, 118a, 118b, and 128 can be any type of machine-learned model. For example, the models 108a, 108b, 108c, 118a, 118b, and 128 can include neural networks, such as deep neural networks, or other types of machine-learned models.

[0070] Thus, the system 100 illustrated in Figure 1 can process the model prompt 102 using the prompt handling system 104 and the processing systems 106, 116, and 126 to generate the aggregate output 130. In some implementations, the prompt handling system 104 can decompose the model prompt 102 into subprompts. The prompt handling system 104 can then process the subprompts using the processing systems 106, 116, and 126 to generate the aggregate output 130.

[0071] Figure 2 provides a more detailed illustration of an example prompt handling system 204. The prompt handling system 204 can receive a model prompt 202 and interact with one or more processing systems 216 to generate an aggregate output 218.

[0072] The example prompt handling system 204 can include a prompt decomposition system 206 and a subprompt handling / aggregation system 214. The prompt decomposition system can include a decomposition proposal system 208, a decomposition scoring system 210, and a compression evaluation system 212.

[0073] The decomposition proposal system 208 can decompose the model prompt 202 into subprompts. The decomposition proposal system 208 can decompose the model prompt 202 into subprompts using various different types of decomposition algorithm.

[0074] In one example, the decomposition proposal system 208 can perform a recursive decomposition approach. This approach can break down a complex prompt into smaller subtasks by iteratively identifying components within the prompt and addressing them individually. For example, a prompt asking for a summary of a document, analysis of themes, and critique of arguments can be recursively decomposed into subprompts for each different task.

[0075] In another example, the decomposition proposal system 208 can perform a graph decomposition approach. This approach can be used for prompts involving relationships or dependencies between entities or concepts. The prompt can be represented as a graph where nodes represent entities and edges represent relationships. The graph can then be decomposed into subgraphs, each focusing on a subset of entities or a specific type of interaction. This allows for independent analysis of smaller groups or relationships, which can be synthesized to address the original prompt.

[0076] In another example, the decomposition proposal system 208 can perform a hierarchical decomposition. This approach can create a hierarchical tree structure where each node represents a subprompt. Relationships between nodes indicate dependencies in processing, allowing for organized and systematic processing of tasks while maintaining dependencies.

[0077] In another example, the decomposition proposal system 208 can perform a content-based decomposition. This approach can segment the prompt based on content markers, delimiters, or other characteristics within the data. For example, a document can be decomposed into paragraphs, sections, or chapters based on headings or other structural elements.

[0078] In another example, the decomposition proposal system 208 can perform a data modality -based decomposition. This approach can segment the prompt based on different data modalities, such as text, images, audio, or video. Each modality can then be processed by the most suitable system.

[0079] In another example, the decomposition proposal system 208 can perform a file-based decomposition. This approach can decompose the prompt based on individual files within the context window. This allows for separate processing of different files based on their type, size, or other characteristics.

[0080] In another example, the decomposition proposal system 208 can perform a hybrid decomposition. A hybrid approach can combine multiple decomposition techniques to achieve a more comprehensive and efficient breakdown of the prompt. For instance, a combination of recursive and graph decomposition can be used to handle complex prompts with both hierarchical and relational structures.

[0081] The decomposition proposal system 208 can decompose the model prompt 202 into subprompts using any combination of the aforementioned decomposition techniques. The decomposition proposal system 208 can then pass the subprompts to the decomposition scoring system 210.

[0082] The decomposition scoring system 210 can score the subprompts. The decomposition scoring system 210 can score the subprompts using one or more different types of scoring algorithm. For example, the decomposition scoring system 210 can score the subprompts based on the complexity of the subprompts, the data size of the subprompts, the computational requirements of the subprompts, or any other type of scoring algorithm.

[0083] The compression evaluation system 212 can evaluate the compressibility of the subprompts. The compression evaluation system 212 can evaluate the compressibility of the subprompts using any type of compression algorithm. For example, the compression evaluation system 212 can evaluate the compressibility of the subprompts based on the data type of the subprompts, the complexity of the subprompts, the data size of the subprompts, the computational requirements of the subprompts, or any other characteristics of the subprompts.

[0084] In some implementations, the compression evaluation system 212 can evaluate the compressibility of the subprompts based on the data type of the subprompts. For example, the compression evaluation system 212 can evaluate whether the subprompts are text-based, image-based, audio-based, or video-based.

[0085] In some implementations, the compression evaluation system 212 can evaluate the compressibility of the subprompts based on the complexity of the subprompts. For example, the compression evaluation system 212 can evaluate whether the subprompts are simple or complex.

[0086] In some implementations, the compression evaluation system 212 can evaluate the compressibility of the subprompts based on the data size of the subprompts. For example, the compression evaluation system 212 can evaluate whether the subprompts are large or small.

[0087] In some implementations, the compression evaluation system 212 can evaluate the compressibility of the subprompts by performing sampling on the subprompts. For example, the compression evaluation system 212 can perform sampling on the subprompts by randomly sampling the subprompts. The compression evaluation system 212 can then perform compression evaluations on samples taken from the subprompts.

[0088] In some implementations, the compression evaluation system 212 can include a machine-learned model that can be trained to evaluate the compressibility of the subprompts. For example, the compression evaluation system 212 can include a neural network that can be trained to evaluate the compressibility of the subprompts. The neural network can be trained to evaluate the compressibility of the subprompts based on their data type, complexity, data size, computational requirements, and / or any other type of factor. The machine-learned model can be a relatively small model that runs on a local computing system.

[0089] The decomposition scoring system 210 can score the subprompts based on their compressibility, as evaluated by the compression evaluation system 212. The scoring system 210 can assign higher scores to subprompts that are more compressible, as this indicates that they can be transmitted more efficiently. The scoring system 210 can also take into account other factors, such as the complexity of the subprompts, the data size of the subprompts, and the computational requirements of the subprompts. However, the compressibility of the subprompts may be a key factor in determining the score assigned by the scoring system 210.

[0090] In some implementations, the decomposition proposal system 208 can propose a particular processing system 216 to handle each subprompt. The decomposition scoring system 210 can score the proposals by comparing the compressibility of each subprompt with the particular processing system 216 to which it is assigned. For example, the decomposition scoring system 210 can score the proposals by assigning a higher score to subprompts that are more compressible and are assigned to a processing system 216 that is better suited for handling compressed data. This can allow the decomposition scoring system 210 to assign higher scores to proposals that are more compressible and are assigned to a processing system 216 that is better suited for handling compressed data.

[0091] In some implementations, the decomposition proposal system 208 can generate a plurality of candidate decompositions, each comprising a different set of candidate subprompts. The decomposition scoring system 210 can then score each of the candidate decompositions based on one or more scoring criteria. For example, the scoring criteria can include the compressibility of the candidate subprompts, the complexity of the candidate subprompts, the data size of the candidate subprompts, and the computational requirements of the candidate subprompts, for example with reference to the capabilities and / or characteristics of the respective processing system to which each candidate subprompt is assigned. The decomposition scoring system 210 can then select one of the candidate decompositions based on the scores, and the subprompts from the selected candidate decomposition can be used for further processing. This iterative process can be repeated until a satisfactory decomposition is found. For example, the decomposition proposal system 208 can generate a new set of candidate decompositions based on the feedback from the decomposition scoring system 210. This process can continue until the decomposition scoring system 210 identifies a decomposition that meets the desired criteria. The top scoring decomposition proposal can then be selected.

[0092] After, a particular decomposition has been selected, the subprompt handling / aggregation system 214 can process the subprompts of the selected decomposition. For example, the subprompt handling / aggregation system 214 can process the subprompts using some combination of the processing systems 216.

[0093] The subprompt handling / aggregation system 214 can generate the aggregate output 218 based on the responses received from the processing systems 216. The aggregate output 218 can be any type of output that can be generated based on the model prompt 202. For example, the aggregate output 218 can include a response to the model prompt 202, asummary of the model prompt 202, a classification of the model prompt 202, or any other type of output that can be generated based on the model prompt 202.

[0094] One example aggregation approach is to concatenate the suboutputs in the order they were generated. This approach is simple to implement but may not be optimal in all cases.

[0095] Another example aggregation approach is to use a weighted average of the suboutputs. The weights can be based on the scores of the subprompts or on other factors, such as the relative importance of the subprompts. For example, a prompt that asks for a summary of a document can be assigned a higher weight than a prompt that asks for analysis of the document.

[0096] Another example aggregation approach is to use a hierarchical aggregation approach where results from child nodes in the decomposition tree are progressively combined until the final solution is synthesized at the root. Another example aggregation approach is to use a parallel processing approach where independent subprompts are processed simultaneously and their results are merged upon completion.

[0097] Another example aggregation approach is to use a machine-learned model to aggregate the suboutputs. This approach can learn to aggregate the suboutputs in a way that is optimal for the specific task at hand. As one example, the suboutputs can be aggregated using a machine-learned model that has been trained on similar aggregation tasks. For example, the model could be trained on a dataset of suboutputs from different models, along with the corresponding aggregated outputs. The model could then learn to aggregate the suboutputs in a way that is optimal for the specific task at hand.

[0098] As one example, in some implementations, the subprompt handling / aggregation system 214 can include one or machine-learned models that operate to aggregate the responses generated by the processing systems 216 responsive to the subprompts. For example, the system 214 could include a model such as a large language model to aggregate responses. For example, the large language model could be trained on a dataset of suboutputs from different models, along with the corresponding aggregated outputs. The large language model could then learn to aggregate the suboutputs in a way that is optimal for the specific task at hand.

[0099] In some cases, it may be necessary to perform further processing on the suboutputs before aggregating them. For example, if the suboutputs are text, it may be necessary to perform text normalization or other text processing tasks before aggregating them.

[0100] Another example aggregation approach is to use a combination of the aforementioned approaches. For example, the suboutputs can be aggregated using a weighted average of the suboutputs and then further processed using a machine-learned model.

[0101] Figure 3 illustrates a flow chart diagram of an example method 300 to process a model prompt. Although the example method 300 depicts a particular sequence of operations, the sequence may be altered without departing from the scope of the present disclosure. For example, some of the operations depicted may be performed in parallel or in a different sequence that does not materially affect the function of the method 300. In other examples, different components of an example device or system that implements the method 300 may perform functions at substantially the same time or in a specific sequence.

[0102] At 302, the method 300 includes receiving a model prompt. A computing system could implement step 302 by receiving a model prompt from a user interface, a network connection, a file system, a text entry field, or any other source. The model prompt could be received in a variety of formats, such as text, audio, image, video, or a combination thereof.

[0103] At 304, the method 300 includes decomposing the model prompt into subprompts. As one example, computing system can implement step 304 by performing a decomposition algorithm on the model prompt to generate a plurality of candidate decompositions, wherein each of the plurality of candidate decompositions comprises a different set of candidate subprompts. For example, the decomposition algorithm could be a recursive decomposition algorithm, a graph decomposition algorithm, a hierarchical decomposition algorithm, a content-based decomposition algorithm, a data modality-based decomposition algorithm, a file-based decomposition algorithm, or a hybrid decomposition algorithm. The decomposition algorithm could also consider factors such as the complexity of the subprompts, the data size of the subprompts, the computational requirements of the subprompts, and the compressibility of the subprompts.

[0104] At 306, the method 300 includes scoring the subprompts. As one example, a computing system can implement step 306 by performing a scoring algorithm on the subprompts to generate a plurality of scores respectively for the plurality of subprompts according to one or more scoring criteria. For example, the scoring criteria could include the complexity of the subprompts, the data size of the subprompts, the computational requirements of the subprompts, and the compressibility of the subprompts. The scoring algorithm could assign higher scores to subprompts that are more compressible, as this indicates that they can be transmitted more efficiently. The scoring algorithm could also takeinto account other factors, such as the complexity of the subprompts, the data size of the subprompts, and the computational requirements of the subprompts.

[0105] In some implementations, at 306, the method 300 includes scoring the model subprompts by evaluating the compressibility of the subprompts. As one example, a computing system can evaluate the compressibility of each subprompt by applying a compression algorithm to each subprompt. For example, the computing system can use a compression algorithm such as Huffman coding, Lempel-Ziv-Welch (LZW) compression, or a combination thereof. The computing system can then evaluate the compressibility of each subprompt based on the compression ratio achieved by the compression algorithm.

[0106] In another example, a computing system can evaluate the compressibility of each subprompt by using a machine-learned compression prediction model. For example, the computing system can use a neural network that has been trained on a dataset of subprompts and their corresponding compression ratios. The computing system can then use the machine- learned compression prediction model to predict the compression ratio of each subprompt.

[0107] In another example, a computing system can evaluate the compressibility of each subprompt by performing sampling on the subprompts. For example, the computing system can randomly sample the subprompts and then perform compression evaluations on the sampled subprompts. The computing system can then use the compression evaluations of the sampled subprompts to estimate the compressibility of the entire subprompt.

[0108] In another example, a computing system can evaluate the compressibility of each subprompt by using a combination of the aforementioned techniques. For example, the computing system can use a compression algorithm to evaluate the compressibility of each subprompt, and then use a machine-learned compression prediction model to refine the compressibility estimates. The computing system can then use the refined compressibility estimates to select the most appropriate processing system for each subprompt.

[0109] At 308, the method 300 includes selecting a particular decomposition of the model subprompt. As one example, a computing system can implement step 308 by evaluating the scores of the subprompts and selecting the candidate decomposition that has the highest score. For example, the computing system could select the candidate decomposition that has the highest score among all the candidate decompositions. The computing system could also select the candidate decomposition that has the highest score among a subset of the candidate decompositions. For example, the computing system could select the candidate decomposition that has the highest score among the candidatedecompositions that have subprompts that are more compressible than the subprompts of other candidate decompositions

[0110] At 310, the method 300 includes processing the subprompts. As one example, a computing system can implement step 310 by causing each of the plurality of subprompts to be respectively processed by the different processing system to which the subprompt has been mapped to generate a plurality of suboutputs. For example, the computing system could use a local processing system to process subprompts that are not computationally intensive and can be processed locally, such as subprompts that require less computational power. The computing system could use a remote processing system to process subprompts that require more computational power and can be compressed to reduce communication costs, such as subprompts that are complex and require more processing power. The computing system could also use a combination of local and remote processing systems to process subprompts. For example, the computing system could use a local processing system to process the initial portion of a subprompt and then use a remote processing system to process the remaining portion of the subprompt. The computing system could also use a caching mechanism to store the results of processing subprompts, which could be used to speed up the processing of subsequent subprompts. The computing system could also use a scheduling algorithm to optimize the order in which subprompts are processed. For example, the computing system could prioritize subprompts that are prerequisites for multiple other subprompts, or those that are expected to take longer to process. This strategic processing order could help in minimizing overall response time and improving the efficiency of resource utilization.

[0111] At 312, the method 300 includes generating an aggregate output. As one example, a computing system can implement step 312 by aggregating the plurality of suboutputs to generate a response to the model prompt. For example, the computing system could use a hierarchical aggregation approach where results from child nodes in the decomposition tree are progressively combined until the final solution is synthesized at the root. Alternatively or additionally, parallel processing can be employed, where independent subprompts are processed simultaneously and their results are merged upon completion. This method can significantly reduce the response time, making the system more efficient in handling extensive or complex prompts.

[0112] Figure 4 depicts a flowchart of a method 400 for training one or more machine-learned models according to aspects of the present disclosure. One or more portion(s) of example method 400 can be implemented by a computing system that includes one or more computing devices such as, for example, computing systems described withreference to the other figures. Each respective portion of example method 400 can be performed by any (or any combination) of one or more computing devices. Moreover, one or more portion(s) of example method 400 can be implemented on the hardware components of the device(s) described herein, for example, to train one or more systems or models. Figure 4 depicts elements performed in a particular order for purposes of illustration and discussion. Those of ordinary skill in the art, using the disclosures provided herein, will understand that the elements of any of the methods discussed herein can be adapted, rearranged, expanded, omitted, combined, or modified in various ways without deviating from the scope of the present disclosure. Figure 4 is described with reference to elements / terms described with respect to other systems and figures for exemplary illustrated purposes and is not meant to be limiting. One or more portions of example method 400 can be performed additionally, or alternatively, by other systems.

[0113] At 402, example method 400 can include obtaining a training instance. A set of training data can include a plurality of training instances divided between multiple datasets (e.g., a training dataset, a validation dataset, or testing dataset). A training instance can be labeled or unlabeled. Although referred to in example method 400 as a “training” instance, it is to be understood that runtime inferences can form training instances when a model is trained using an evaluation of the model’s performance on that runtime instance (e.g., online training / learning). Example data types for the training instance and various tasks associated therewith are described throughout the present disclosure.

[0114] At 404, example method 400 can include processing, using one or more machine-learned models, the training instance to generate an output. The output can be directly obtained from the one or more machine-learned models or can be a downstream result of a chain of processing operations that includes an output of the one or more machine- learned models.

[0115] At 406, example method 400 can include receiving an evaluation signal associated with the output. The evaluation signal can be obtained using a loss function. Various determinations of loss can be used, such as mean squared error, likelihood loss, cross entropy loss, hinge loss, contrastive loss, or various other loss functions. The evaluation signal can be computed using known ground-truth labels (e.g., supervised learning), predicted or estimated labels (e.g., semi- or self-supervised learning), or without labels (e.g., unsupervised learning). The evaluation signal can be a reward (e.g., for reinforcement learning). The reward can be computed using a machine-learned reward model configured togenerate rewards based on output(s) received. The reward can be computed using feedback data describing human feedback on the output(s).

[0116] At 408, example method 400 can include updating the machine-learned model using the evaluation signal. For example, values for parameters of the machine-learned model(s) can be learned, in some embodiments, using various training or learning techniques, such as, for example, backwards propagation. For example, the evaluation signal can be backpropagated from the output (or another source of the evaluation signal) through the machine-learned model(s) to update one or more parameters of the model(s) (e.g., based on a gradient of the evaluation signal with respect to the parameter value(s)). For example, system(s) containing one or more machine-learned models can be trained in an end-to-end manner. Gradient descent techniques can be used to iteratively update the parameters over a number of training iterations. In some implementations, performing backwards propagation of errors can include performing truncated backpropagation through time. Example method 400 can include implementing a number of generalization techniques (e.g., weight decays, dropouts, etc.) to improve the generalization capability of the models being trained.

[0117] In some implementations, example method 400 can be implemented for training a machine-learned model from an initialized state to a fully trained state (e.g., when the model exhibits a desired performance profile, such as based on accuracy, precision, recall, etc.).

[0118] In some implementations, example method 400 can be implemented for particular stages of a training procedure. For instance, in some implementations, example method 400 can be implemented for pre-training a machine-learned model. Pre-training can include, for instance, large-scale training over potentially noisy data to achieve a broad base of performance levels across a variety of tasks / data types.

[0119] In some implementations, example method 400 can be implemented for finetuning a machine-learned model. Fine-tuning can include, for instance, smaller-scale training on higher-quality (e.g., labeled, curated, etc.) data. Fine-tuning can affect all or a portion of the parameters of a machine-learned model. For example, various portions of the machine- learned model can be “frozen” for certain training stages. For example, parameters associated with an embedding space can be “frozen” during fine-tuning (e.g., to retain information learned from a broader domain(s) than present in the fine-tuning dataset(s)). In some implementations, example method 400 uses adapter modules. Adapters can be small trainable layers that are inserted between pre-existing layers of a pre-trained model. During the fine-tuning process, the original parameters of the pre-trained model are typically frozen, and only the parameters of the adapters are updated.

[0120] In some implementations, example method 400 can be implemented to execute parameter-efficient fine-tuning methods, such as Layerwise Optimization of Residuals (LoRA). LoRA can refine pre-trained models with minimal adjustments to the original parameters. This can be achieved by introducing trainable low-rank matrices that modify the behavior of the pre-trained weights without directly altering them. In some implementations, during fine-tuning, only these auxiliary matrices are updated, which significantly reduces the number of parameters that are trained.

[0121] An example fine-tuning approach includes reinforcement learning. Reinforcement learning can be based on user feedback on model performance during use.

[0122] Figure 5 is a block diagram of an example processing flow for using machine- learned model(s) 1 to process input(s) 2 to generate output(s) 3.

[0123] Machine-learned model(s) 1 can be or include one or multiple machine- learned models or model components. Example machine-learned models can include neural networks (e.g., deep neural networks). Example machine-learned models can include nonlinear models or linear models. Example machine-learned models can use other architectures in lieu of or in addition to neural networks. Example machine-learned models can include decision tree based models, support vector machines, hidden Markov models, Bayesian networks, linear regression models, k-means clustering models, etc.

[0124] Machine-learned model(s) 1 can be or include, or otherwise be representative of any one or more of the machine-learned models described above with respect to the preceding figures. For example, machine-learned model(s) 1 can be or include, or otherwise be representative of any one or more of any of the models mentioned herein, etc. Although various features, variations, and implementations described below are described with respect to machine-learned model(s) 1, it is to be understood that such features, variations, and implementations are to be understood as described with respect to each of any of the models mentioned herein etc., any other machine-learned component described herein.

[0125] Example neural networks can include feed-forward neural networks, recurrent neural networks (RNNs), including long short-term memory (LSTM) based recurrent neural networks, convolutional neural networks (CNNs), diffusion models, generative-adversarial networks, or other forms of neural networks. Example neural networks can be deep neural networks. Some example machine-learned models can leverage an attention mechanism suchas self-attention. For example, some example machine-learned models can include multiheaded self-attention models.

[0126] Machine-learned model(s) 1 can include a single or multiple instances of the same model configured to operate on data from input(s) 2. Machine-learned model(s) 1 can include multiple different models or multiple different model portions configured to operate on data from input(s) 2.

[0127] Machine-learned model(s) 1 can include an ensemble of different models that can cooperatively interact to process data from input(s) 2. For example, a model ensemble can include multiple models that have different attributes (e.g., different architectures, trained with different recipes, etc.). The ensemble can output an overall output based on the individual outputs of the constituent models. In this manner, for instance, the diverse constituent models can work together to provide system-level robustness by effectively aggregating over individual strengths and weaknesses of any given model. The respective individual outputs can be combined in a weighted combination, using a voting or routing mechanism, or a learned output layer (e.g., one or more feedforward or fully-connected layers).

[0128] Machine-learned model(s) 1 can employ a mixture-of-experts structure. See, e.g., Zhou et al., Mixture -of -Experts with Expert Choice Routing, ARXIV:2202.09368V2 (Oct. 14, 2022). For example, different portions of a model can learn (explicitly or implicitly) different expertise areas, with pathways through the model being selected by a learned routing mechanism that engages the appropriate expert for a given input (e.g., a given portion of an input, such as on a per-token basis). For example, a feedforward network can be sparsely activated for a given portion of an input based on an output of a routing mechanism that processes the portion of the input. In this manner, for instance, the group of activated weights can form an “expert” that is selected by the router. On each forward pass, only a subset of the total model weights may be engaged, thereby decreasing a quantity of operations performed for processing a given input compared to a densely activated model. In this manner, for instance, the expressive and interpretive power of a high-parameter-count model can be achieved with more compute-efficient forward passes.

[0129] Input(s) 2 can generally include or otherwise represent various types of data. Input(s) 2 can include one type or many different types of data. Output(s) 3 can be data of the same type(s) or of different types of data as compared to input(s) 2. Output(s) 3 can include one type or many different types of data.

[0130] Example data types for input(s) 2 or output(s) 3 include natural language text data, software code data (e.g., source code, object code, machine code, or any other form of computer-readable instructions or programming languages), machine code data (e.g., binary code, assembly code, or other forms of machine-readable instructions that can be executed directly by a computer’s central processing unit), assembly code data (e.g., low-level programming languages that use symbolic representations of machine code instructions to program a processing unit), genetic data or other chemical or biochemical data, image data, audio data, audiovisual data, haptic data, biometric data, medical data, financial data, statistical data, geographical data, astronomical data, historical data, sensor data generally (e.g., digital or analog values, such as voltage or other absolute or relative level measurement values from a real or artificial input, such as from an audio sensor, light sensor, displacement sensor, etc.), and the like. Data can be raw or processed and can be in any format or schema.

[0131] In multimodal inputs 2 or outputs 3, example combinations of data types include image data and audio data, image data and natural language data, natural language data and software code data, image data and biometric data, sensor data and medical data, etc. It is to be understood that any combination of data types in an input 2 or an output 3 can be present.

[0132] An example input 2 can include one or multiple data types, such as the example data types noted above. An example output 3 can include one or multiple data types, such as the example data types noted above. The data type(s) of input 2 can be the same as or different from the data type(s) of output 3. It is to be understood that the example data types noted above are provided for illustrative purposes only. Data types contemplated within the scope of the present disclosure are not limited to those examples noted above.

[0133] Figure 6 is a block diagram of an example implementation of an example machine-learned model configured to process sequences of information. For instance, an example implementation of machine-learned model(s) 1 can include machine-learned sequence processing model(s) 4. An example system can pass input(s) 2 to sequence processing model(s) 4. Sequence processing model(s) 4 can include one or more machine- learned components. Sequence processing model(s) 4 can process the data from input(s) 2 to obtain an input sequence 5. Input sequence 5 can include one or more input elements 5-1, 5- 2, . . . , 5-A , etc. obtained from input(s) 2. Sequence processing model 4 can process input sequence 5 using prediction layer(s) 6 to generate an output sequence 7. Output sequence 7 can include one or more output elements 7-1, 7-2, . . . , 7 -A, etc. generated based on input sequence 5. The system can generate output(s) 3 based on output sequence 7.

[0134] Sequence processing model(s) 4 can include one or multiple machine-learned model components configured to ingest, generate, or otherwise reason over sequences of information. For example, some example sequence processing models in the text domain are referred to as “Large Language Models,” or LLMs. See, e.g., PaLM 2 Technical Report, GOOGLE, https: / / ai.google / static / documents / palm2techreport.pdf (n.d.). Other example sequence processing models can operate in other domains, such as image domains, see, e.g., Dosovitskiy et al., An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, ARXIV:2010.11929v2 (Jun. 3, 2021), audio domains, see, e.g., Agostinelli et al., MusicLM: Generating Music From Text, ARXIV:2301.1 1325V1 (Jan. 26, 2023), biochemical domains, see, e.g., Jumper et al., Highly accurate protein structure prediction with AlphaFold, 596 Nature 583 (Aug. 26, 2021), by way of example. Sequence processing model(s) 4 can process one or multiple types of data simultaneously. Sequence processing model(s) 4 can include relatively large models (e.g., more parameters, computationally expensive, etc.), relatively small models (e.g., fewer parameters, computationally lightweight, etc.), or both.

[0135] In general, sequence processing model(s) 4 can obtain input sequence 5 using data from input(s) 2. For instance, input sequence 5 can include a representation of data from input(s) 2 in a format understood by sequence processing model(s) 4. One or more machine- learned components of sequence processing model(s) 4 can ingest the data from input(s) 2, parse the data into pieces compatible with the processing architectures of sequence processing model(s) 4 (e.g., via “tokenization”), and project the pieces into an input space associated with prediction layer(s) 6 (e.g., via “embedding”).

[0136] Sequence processing model(s) 4 can ingest the data from input(s) 2 and parse the data into a sequence of elements to obtain input sequence 5. For example, a portion of input data from input(s) 2 can be broken down into pieces that collectively represent the content of the portion of the input data. The pieces can provide the elements of the sequence.

[0137] Elements 5-1, 5-2, . . . , 5-M can represent, in some cases, building blocks for capturing or expressing meaningful information in a particular data domain. For instance, the elements can describe “atomic units” across one or more domains. For example, for textual input source(s), the elements can correspond to groups of one or more words or sub-word components, such as sets of one or more characters.

[0138] For example, elements 5-1, 5-2, . . . , 5-M can represent tokens obtained using a tokenizer. For instance, a tokenizer can process a given portion of an input source and output a series of tokens (e.g., corresponding to input elements 5-1, 5-2, . . . , 5-M) that represent the portion of the input source. Various approaches to tokenization can be used. Forinstance, textual input source(s) can be tokenized using a byte-pair encoding (BPE) technique. See, e.g., Kudo et al., SentencePiece: A simple and language independent subword tokenizer and detokenizer for Neural Text Processing, PROCEEDINGS OF THE 2018 CONFERENCE ON EMPIRICAL METHODS IN NATURAL LANGUAGE PROCESSING (System Demonstrations), pages 66-71 (October 31-November 4, 2018), https: / / aclanthology.org / D18-2012.pdf. Image-based input source(s) can be tokenized by extracting and serializing patches from an image.

[0139] In general, arbitrary data types can be serialized and processed into input sequence 5. It is to be understood that element(s) 5-1, 5-2, . . . , 5-M depicted in Figure 6 can be the tokens or can be the embedded representations thereof.

[0140] Prediction layer(s) 6 can predict one or more output elements 7-1, 7-2, . . . , 7- N based on the input elements. Prediction layer(s) 6 can include one or more machine-learned model architectures, such as one or more layers of learned parameters that manipulate and transform the input(s) to extract higher-order meaning from, and relationships between, input element(s) 5-1, 5-2, . . . , 5-M. In this manner, for instance, example prediction layer(s) 6 can predict new output element(s) in view of the context provided by input sequence 5.

[0141] Prediction layer(s) 6 can evaluate associations between portions of input sequence 5 and a particular output element. These associations can inform a prediction of the likelihood that a particular output follows the input context. For example, consider the textual snippet, “The carpenter’s toolbox was small and heavy. It was full of .” Example prediction layer(s) 6 can identify that “It” refers back to “toolbox” by determining a relationship between the respective embeddings. Example prediction layer(s) 6 can also link “It” to the attributes of the toolbox, such as “small” and “heavy.” Based on these associations, prediction layer(s) 6 can, for instance, assign a higher probability to the word “nails” than to the word “sawdust.”

[0142] A transformer is an example architecture that can be used in prediction layer(s) 4. See, e.g., Vaswani et al., Attention Is All You Need, ARXIV: 1706.03762v7 (Aug. 2, 2023). A transformer is an example of a machine-learned model architecture that uses an attention mechanism to compute associations between items within a context window. The context window can include a sequence that contains input sequence 5 and potentially one or more output element(s) 7-1, 7-2, . . . , 7-N. A transformer block can include one or more attention layer(s) and one or more post-attention layer(s) (e.g., feedforward layer(s), such as a multi-layer perceptron).

[0143] Prediction layer(s) 6 can include other machine-learned model architectures in addition to or in lieu of transformer-based architectures. For example, recurrent neural networks (RNNs) and long short-term memory (LSTM) models can also be used, as well as convolutional neural networks (CNNs). In general, prediction layer(s) 6 can leverage various kinds of artificial neural networks that can understand or generate sequences of information.

[0144] Output sequence 7 can include or otherwise represent the same or different data types as input sequence 5. For instance, input sequence 5 can represent textual data, and output sequence 7 can represent textual data. Input sequence 5 can represent image, audio, or audiovisual data, and output sequence 7 can represent textual data (e.g., describing the image, audio, or audiovisual data). It is to be understood that prediction layer(s) 6, and any other interstitial model components of sequence processing model(s) 4, can be configured to receive a variety of data types in input sequence(s) 5 and output a variety of data types in output sequence(s) 7.

[0145] Output sequence 7 can have various relationships to input sequence 5. Output sequence 7 can be a continuation of input sequence 5. Output sequence 7 can be complementary to input sequence 5. Output sequence 7 can translate, transform, augment, or otherwise modify input sequence 5. Output sequence 7 can answer, evaluate, confirm, or otherwise respond to input sequence 5. Output sequence 7 can implement (or describe instructions for implementing) an instruction provided via input sequence 5.

[0146] Output sequence 7 can be generated autoregressively. For instance, for some applications, an output of one or more prediction layer(s) 6 can be passed through one or more output layers (e.g., softmax layer) to obtain a probability distribution over an output vocabulary (e.g., a textual or symbolic vocabulary) conditioned on a set of input elements in a context window. In this manner, for instance, output sequence 7 can be autoregressively generated by sampling a likely next output element, adding that element to the context window, and re-generating the probability distribution based on the updated context window, and sampling a likely next output element, and so forth.

[0147] Output sequence 7 can also be generated non-autoregressively. For instance, multiple output elements of output sequence 7 can be predicted together without explicit sequential conditioning on each other. See, e.g., Saharia et al., Non-Autoregressive Machine Translation with Latent Alignments, ARXIV:2004.07437V3 (NOV. 16, 2020).

[0148] Output sequence 7 can include one or multiple portions or elements. In an example content generation configuration, output sequence 7 can include multiple elements corresponding to multiple portions of a generated output sequence (e.g., a textual sentence,values of a discretized waveform, computer code, etc.). In an example classification configuration, output sequence 7 can include a single element associated with a classification output. For instance, an output “vocabulary” can include a set of classes into which an input sequence is to be classified. For instance, a vision transformer block can pass latent state information to a multilayer perceptron that outputs a likely class value associated with an input image.

[0149] Figure 7 is a block diagram of an example technique for populating an example input sequence 8. Input sequence 8 can include various functional elements that form part of the model infrastructure, such as an element 8-0 obtained from a task indicator 9 that signals to any model(s) that process input sequence 8 that a particular task is being performed (e.g., to help adapt a performance of the model(s) to that particular task). Input sequence 8 can include various data elements from different data modalities. For instance, an input modality 10-1 can include one modality of data. A data-to- sequence model 11-1 can process data from input modality 10-1 to project the data into a format compatible with input sequence 8 (e.g., one or more vectors dimensioned according to the dimensions of input sequence 8) to obtain elements 8-1, 8-2, 8-3. Another input modality 10-2 can include a different modality of data. A data-to-sequence model 11-2 can project data from input modality 10-2 into a format compatible with input sequence 8 to obtain elements 8-4, 8-5, 8- 6. Another input modality 10-3 can include yet another different modality of data. A data-to- sequence model 11-3 can project data from input modality 10-3 into a format compatible with input sequence 8 to obtain elements 8-7, 8-8, 8-9.

[0150] Input sequence 8 can be the same as or different from input sequence 5. Input sequence 8 can be a multimodal input sequence that contains elements that represent data from different modalities using a common dimensional representation. For instance, an embedding space can have P dimensions. Input sequence 8 can be configured to contain a plurality of elements that have / Jdimensions. In this manner, for instance, example implementations can facilitate information extraction and reasoning across diverse data modalities by projecting data into elements in the same embedding space for comparison, combination, or other computations therebetween.

[0151] For example, elements 8-0, . . . , 8-9 can indicate particular locations within a multidimensional embedding space. Some elements can map to a set of discrete locations in the embedding space. For instance, elements that correspond to discrete members of a predetermined vocabulary of tokens can map to discrete locations in the embedding space that are associated with those tokens. Other elements can be continuously distributed acrossthe embedding space. For instance, some data types can be broken down into continuously defined portions (e.g., image patches) that can be described using continuously distributed locations within the embedding space.

[0152] In some implementations, the expressive power of the embedding space may not be limited to meanings associated with any particular set of tokens or other building blocks. For example, a continuous embedding space can encode a spectrum of high-order information. An individual piece of information (e.g., a token) can map to a particular point in that space: for instance, a token for the word “dog” can be projected to an embedded value that points to a particular location in the embedding space associated with canine-related information. Similarly, an image patch of an image of a dog on grass can also be projected into the embedding space. In some implementations, the projection of the image of the dog can be similar to the projection of the word “dog” while also having similarity to a projection of the word “grass,” while potentially being different from both. In some implementations, the projection of the image patch may not exactly align with any single projection of a single word. In some implementations, the projection of the image patch can align with a combination of the projections of the words “dog” and “grass.” In this manner, for instance, a high-order embedding space can encode information that can be independent of data modalities in which the information is expressed.

[0153] Task indicator 9 can include a model or model component configured to identify a task being performed and inject, into input sequence 8, an input value represented by element 8-0 that signals which task is being performed. For instance, the input value can be provided as a data type associated with an input modality and projected along with that input modality (e.g., the input value can be a textual task label that is embedded along with other textual data in the input; the input value can be a pixel-based representation of a task that is embedded along with other image data in the input; etc.). The input value can be provided as a data type that differs from or is at least independent from other input(s). For instance, the input value represented by element 8-0 can be a learned within a continuous embedding space.

[0154] Input modalities 10-1, 10-2, and 10-3 can be associated with various different data types (e.g., as described above with respect to input(s) 2 and output(s) 3).

[0155] Data-to-sequence models 11-1, 11-2, and 11-3 can be the same or different from each other. Data-to-sequence models 11-1, 11-2, and 11-3 can be adapted to each respective input modality 10-1, 10-2, and 10-3. For example, a textual data-to-sequence model can subdivide a portion of input text and project the subdivisions into element(s) ininput sequence 8 (e.g., elements 8-1, 8-2, 8-3, etc.). An image data-to- sequence model can subdivide an input image and project the subdivisions into element(s) in input sequence 8 (e.g., elements 8-4, 8-5, 8-6, etc.). An arbitrary datatype data-to-sequence model can subdivide an input of that arbitrary datatype and project the subdivisions into element(s) in input sequence 8 (e.g., elements 8-7, 8-8, 8-9, etc.).

[0156] Data-to-sequence models 11-1, 11-2, and 11-3 can form part of machine- learned sequence processing model(s) 4. Data-to-sequence models 11-1, 11-2, and 11-3 can be jointly trained with or trained independently from machine-learned sequence processing model(s) 4. Data-to-sequence models 11-1, 11-2, and 11-3 can be trained end-to-end with machine-learned sequence processing model(s) 4.

[0157] Figure 8 is a block diagram of an example model development platform 12 that can facilitate creation, adaptation, and refinement of example machine-learned models (e.g., machine-learned model(s) 1, sequence processing model(s) 4, etc.). Model development platform 12 can provide a number of different toolkits that developer systems can employ in the development of new or adapted machine-learned models.

[0158] Model development platform 12 can provide one or more model libraries 13 containing building blocks for new models. Model libraries 13 can include one or more pretrained foundational models 13-1, which can provide a backbone of processing power across various tasks. Model libraries 13 can include one or more pre-trained expert models 13-2, which can be focused on performance in particular domains of expertise. Model libraries 13 can include various model primitives 13-3, which can provide low-level architectures or components (optionally pre-trained), which can be assembled in various arrangements as desired. Model primitives 13-3 can include a library of pre-trained adapters or LoRA modules that can adapt a baseline foundational model to align its outputs with a desired performance profile, augment model capabilities (e.g., to adapt to a different input modality, etc.), and the like.

[0159] Model development platform 12 can receive selections of various model components 14. Model development platform 12 can pass selected model components 14 to a workbench 15 that combines selected model components 14 into a development model 16.

[0160] Workbench 15 can facilitate further refinement and adaptation of development model 16 by leveraging a number of different toolkits integrated with model development platform 12. For example, workbench 15 can facilitate alignment of the development model 16 with a desired performance profile on various tasks using a model alignment toolkit 17.

[0161] Model alignment toolkit 17 can provide a number of tools for causing development model 16 to generate outputs aligned with desired behavioral characteristics. Alignment can include increasing an accuracy, precision, recall, etc. of model outputs. Alignment can include enforcing output styles, schema, or other preferential characteristics of model outputs. Alignment can be general or domain-specific. For instance, a pre-trained foundational model 13-1 can begin with an initial level of performance across multiple domains. Alignment of the pre-trained foundational model 13-1 can include improving a performance in a particular domain of information or tasks (e.g., even at the expense of performance in another domain of information or tasks).

[0162] Model alignment toolkit 17 can integrate one or more dataset(s) 17-1 for aligning development model 16. Curated dataset(s) 17-1 can include labeled or unlabeled training data. Dataset(s) 17-1 can be obtained from public domain datasets. Dataset(s) 17-1 can be obtained from private datasets associated with one or more developer system(s) for the alignment of bespoke machine-learned model(s) customized for private use-cases.

[0163] Pre-training pipelines 17-2 can include a machine-learned model training workflow configured to update development model 16 over large-scale, potentially noisy datasets. For example, pre-training can leverage unsupervised learning techniques (e.g., denoising, etc.) to process large numbers of training instances to update model parameters from an initialized state and achieve a desired baseline performance. Pre-training pipelines 17-2 can leverage unlabeled datasets in dataset(s) 17-1 to perform pre-training. Workbench 15 can implement a pre-training pipeline 17-2 to pre-train development model 16.

[0164] Fine-tuning pipelines 17-3 can include a machine-learned model training workflow configured to refine the model parameters of development model 16 with higher- quality data. Fine-tuning pipelines 17-3 can update development model 16 by conducting supervised training with labeled dataset(s) in dataset(s) 17-1. Fine-tuning pipelines 17-3 can update development model 16 by conducting reinforcement learning using reward signals from user feedback signals. Workbench 15 can implement a fine-tuning pipeline 17-3 to finetune development model 16.

[0165] Prompt libraries 17-4 can include sets of inputs configured to induce behavior aligned with desired performance criteria. Prompt libraries 17-4 can include few-shot prompts (e.g., inputs providing examples of desired model outputs for prepending to a desired runtime query), chain-of-thought prompts (e.g., inputs providing step-by-step reasoning within the exemplars to facilitate thorough reasoning by the model), and the like.

[0166] Example prompts can be retrieved from an available repository of prompt libraries 17-4. Example prompts can be contributed by one or more developer systems using workbench 15.

[0167] In some implementations, pre-trained or fine-tuned models can achieve satisfactory performance without exemplars in the inputs. For instance, zero-shot prompts can include inputs that lack exemplars. Zero-shot prompts can be within a domain within a training dataset or outside of the training domain(s).

[0168] Prompt libraries 17-4 can include one or more prompt engineering tools.Prompt engineering tools can provide workflows for retrieving or learning optimized prompt values. Prompt engineering tools can facilitate directly learning prompt values (e.g., input element values) based one or more training iterations. Workbench 15 can implement prompt engineering tools in development model 16.

[0169] Prompt libraries 17-4 can include pipelines for prompt generation. For example, inputs can be generated using development model 16 itself or other machine- learned models. In this manner, for instance, a first model can process information about a task and output a input for a second model to process in order to perform a step of the task. The second model can be the same as or different from the first model. Workbench 15 can implement prompt generation pipelines in development model 16.

[0170] Prompt libraries 17-4 can include pipelines for context injection. For instance, a performance of development model 16 on a particular task can improve if provided with additional context for performing the task. Prompt libraries 17-4 can include software components configured to identify desired context, retrieve the context from an external source (e.g., a database, a sensor, etc.), and add the context to the input prompt. Workbench 15 can implement context injection pipelines in development model 16.

[0171] Although various training examples described herein with respect to model development platform 12 refer to “pre-training” and “fine-tuning,” it is to be understood that model alignment toolkit 17 can generally support a wide variety of training techniques adapted for training a wide variety of machine-learned models. Example training techniques can correspond to the example training method 400 described above.

[0172] Model development platform 12 can include a model plugin toolkit 18. Model plugin toolkit 18 can include a variety of tools configured for augmenting the functionality of a machine-learned model by integrating the machine-learned model with other systems, devices, and software components. For instance, a machine-learned model can use tools to increase performance quality where appropriate. For instance, deterministic tasks can beoffloaded to dedicated tools in lieu of probabilistically performing the task with an increased risk of error. For instance, instead of autoregressively predicting the solution to a system of equations, a machine-learned model can recognize a tool to call for obtaining the solution and pass the system of equations to the appropriate tool. The tool can be a traditional system of equations solver that can operate deterministically to resolve the system of equations. The output of the tool can be returned in response to the original query. In this manner, tool use can allow some example models to focus on the strengths of machine-learned models — e.g., understanding an intent in an unstructured request for a task — while augmenting the performance of the model by offloading certain tasks to a more focused tool for rote application of deterministic algorithms to a well-defined problem.

[0173] Model plugin toolkit 18 can include validation tools 18-1. Validation tools 18- 1 can include tools that can parse and confirm output(s) of a machine-learned model.Validation tools 18-1 can include engineered heuristics that establish certain thresholds applied to model outputs. For example, validation tools 18-1 can ground the outputs of machine-learned models to structured data sources (e.g., to mitigate “hallucinations”).

[0174] Model plugin toolkit 18 can include tooling packages 18-2 for implementing one or more tools that can include scripts or other executable code that can be executed alongside development model 16. Tooling packages 18-2 can include one or more inputs configured to cause machine-learned model(s) to implement the tools (e.g., few-shot prompts that induce a model to output tool calls in the proper syntax, etc.). Tooling packages 18-2 can include, for instance, fine-tuning training data for training a model to use a tool.

[0175] Model plugin toolkit 18 can include interfaces for calling external application programming interfaces (APIs) 18-3. For instance, in addition to or in lieu of implementing tool calls or tool code directly with development model 16, development model 16 can be aligned to output instructions that initiate API calls to send or obtain data via external systems.

[0176] Model plugin toolkit 18 can integrate with prompt libraries 17-4 to build a catalog of available tools for use with development model 16. For instance, a model can receive, in an input, a catalog of available tools, and the model can generate an output that selects a tool from the available tools and initiates a tool call for using the tool.

[0177] Model development platform 12 can include a computational optimization toolkit 19 for optimizing a computational performance of development model 16. For instance, tools for model compression 19-1 can allow development model 16 to be reduced in size while maintaining a desired level of performance. For instance, model compression 19-1can include quantization workflows, weight pruning and sparsification techniques, etc. Tools for hardware acceleration 19-2 can facilitate the configuration of the model storage and execution formats to operate optimally on different hardware resources. For instance, hardware acceleration 19-2 can include tools for optimally sharding models for distributed processing over multiple processing units for increased bandwidth, lower unified memory requirements, etc. Tools for distillation 19-3 can provide for the training of lighter-weight models based on the knowledge encoded in development model 16. For instance, development model 16 can be a highly performant, large machine-learned model optimized using model development platform 12. To obtain a lightweight model for running in resource-constrained environments, a smaller model can be a “student model” that learns to imitate development model 16 as a “teacher model.” In this manner, for instance, the investment in learning the parameters and configurations of development model 16 can be efficiently transferred to a smaller model for more efficient inference.

[0178] Workbench 15 can implement one, multiple, or none of the toolkits implemented in model development platform 12. Workbench 15 can output an output model 20 based on development model 16. Output model 20 can be a deployment version of development model 16. Output model 20 can be a development or training checkpoint of development model 16. Output model 20 can be a distilled, compressed, or otherwise optimized version of development model 16.

[0179] Figure 9 is a block diagram of an example training flow for training a machine-learned development model 16. One or more portion(s) of the example training flow can be implemented by a computing system that includes one or more computing devices such as, for example, computing systems described with reference to the other figures. Each respective portion of the example training flow can be performed by any (or any combination) of one or more computing devices. Moreover, one or more portion(s) of the example training flow can be implemented on the hardware components of the device(s) described herein, for example, to train one or more systems or models. FIG. 9 depicts elements performed in a particular order for purposes of illustration and discussion. Those of ordinary skill in the art, using the disclosures provided herein, will understand that the elements of any of the methods discussed herein can be adapted, rearranged, expanded, omitted, combined, or modified in various ways without deviating from the scope of the present disclosure. FIG. 9 is described with reference to elements / terms described with respect to other systems and figures for exemplary illustrated purposes and is not meant to belimiting. One or more portions of the example training flow can be performed additionally, or alternatively, by other systems.

[0180] Initially, development model 16 can persist in an initial state as an initialized model 21. Development model 16 can be initialized with weight values. Initial weight values can be random or based on an initialization schema. Initial weight values can be based on prior pre-training for the same or for a different model.

[0181] Initialized model 21 can undergo pre-training in a pre-training stage 22. Pretraining stage 22 can be implemented using one or more pre-training pipelines 17-2 over data from dataset(s) 17-1. Pre-training can be omitted, for example, if initialized model 21 is already pre-trained (e.g., development model 16 contains, is, or is based on a pre-trained foundational model or an expert model).

[0182] Pre-trained model 23 can then be a new version of development model 16, which can persist as development model 16 or as a new development model. Pre-trained model 23 can be the initial state if development model 16 was already pre-trained. Pre-trained model 23 can undergo fine-tuning in a fine-tuning stage 24. Fine-tuning stage 24 can be implemented using one or more fine-tuning pipelines 17-3 over data from dataset(s) 17-1. Fine-tuning can be omitted, for example, if a pre-trained model as satisfactory performance, if the model was already fine-tuned, or if other tuning approaches are preferred.

[0183] Fine-tuned model 29 can then be a new version of development model 16, which can persist as development model 16 or as a new development model. Fine-tuned model 29 can be the initial state if development model 16 was already fine-tuned. Fine-tuned model 29 can undergo refinement with user feedback 26. For instance, refinement with user feedback 26 can include reinforcement learning, optionally based on human feedback from human users of fine-tuned model 25. As reinforcement learning can be a form of fine-tuning, it is to be understood that fine-tuning stage 24 can subsume the stage for refining with user feedback 26. Refinement with user feedback 26 can produce a refined model 27. Refined model 27 can be output to downstream system(s) 28 for deployment or further development.

[0184] In some implementations, computational optimization operations can be applied before, during, or after each stage. For instance, initialized model 21 can undergo computational optimization 29-1 (e.g., using computational optimization toolkit 19) before pre-training stage 22. Pre-trained model 23 can undergo computational optimization 29-2 (e.g., using computational optimization toolkit 19) before fine-tuning stage 24. Fine-tuned model 25 can undergo computational optimization 29-3 (e.g., using computational optimization toolkit 19) before refinement with user feedback 26. Refined model 27 canundergo computational optimization 29-4 (e.g., using computational optimization toolkit 19) before output to downstream system(s) 28. Computational optimization(s) 29-1, . . . , 29-4 can all be the same, all be different, or include at least some different optimization techniques.

[0185] Figure 10 is a block diagram of an inference system for operating one or more machine-learned model(s) 1 to perform inference (e.g., for training, for deployment, etc.). A model host 31 can receive machine-learned model(s) 1. Model host 31 can host one or more model instance(s) 31-1, which can be one or multiple instances of one or multiple models. Model host 31 can host model instance(s) 31-1 using available compute resources 31-2 associated with model host 31.

[0186] Model host 31 can perform inference on behalf of one or more client(s) 32. Client(s) 32 can transmit an input request 33 to model host 31. Using input request 33, model host 31 can obtain input(s) 2 for input to machine-learned model(s) 1. Machine-learned model(s) 1 can process input(s) 2 to generate output(s) 3. Using output(s) 3, model host 31 can return an output payload 34 for responding to input request 33 from client(s) 32. Output payload 34 can include or be based on output(s) 3.

[0187] Model host 31 can leverage various other resources and tools to augment the inference task. For instance, model host 31 can communicate with tool interfaces 35 to facilitate tool use by model instance(s) 31-1. Tool interfaces 35 can include local or remote APIs. Tool interfaces 35 can include integrated scripts or other software functionality. Model host 31 can engage online learning interface(s) 36 to facilitate ongoing improvements to machine-learned model(s) 1. For instance, online learning interface(s) 36 can be used within reinforcement learning loops to retrieve user feedback on inferences served by model host 31. Model host 31 can access runtime data source(s) 37 for augmenting input(s) 2 with additional contextual information. For instance, runtime data source(s) 37 can include a knowledge graph 37-1 that facilitates structured information retrieval for information associated with input request(s) 33 (e.g., a search engine service). Runtime data source(s) 37 can include public or private, external or local database(s) 37-2 that can store information associated with input request(s) 33 for augmenting input(s) 2. Runtime data source(s) 37 can include account data 37-3 which can be retrieved in association with a user account corresponding to a client 32 for customizing the behavior of model host 31 accordingly.

[0188] Model host 31 can be implemented by one or multiple computing devices or systems. Client(s) 2 can be implemented by one or multiple computing devices or systems, which can include computing devices or systems shared with model host 31.

[0189] For example, model host 31 can operate on a server system that provides a machine-learning service to client device(s) that operate client(s) 32 (e.g., over a local or wide-area network). Client device(s) can be end-user devices used by individuals. Client device(s) can be server systems that operate client(s) 32 to provide various functionality as a service to downstream end-user devices.

[0190] In some implementations, model host 31 can operate on a same device or system as client(s) 32. Model host 31 can be a machine-learning service that runs on-device to provide machine-learning functionality to one or multiple applications operating on a client device, which can include an application implementing client(s) 32. Model host 31 can be a part of a same application as client(s) 32. For instance, model host 31 can be a subroutine or method implemented by one part of an application, and client(s) 32 can be another subroutine or method that engages model host 31 to perform inference functions within the application. It is to be understood that model host 31 and client(s) 32 can have various different configurations.

[0191] Model instance(s) 31-1 can include one or more machine-learned models that are available for performing inference. Model instance(s) 31-1 can include weights or other model components that are stored on in persistent storage, temporarily cached, or loaded into high-speed memory. Model instance(s) 31-1 can include multiple instance(s) of the same model (e.g., for parallel execution of more requests on the same model). Model instance(s) 31-1 can include instance(s) of different model(s). Model instance(s) 31-1 can include cached intermediate states of active or inactive model(s) used to accelerate inference of those models. For instance, an inference session with a particular model may generate significant amounts of computational results that can be re-used for future inference runs (e.g., using a KV cache for transformer-based models). These computational results can be saved in association with that inference session so that session can be executed more efficiently when resumed.

[0192] Compute resource(s) 31-2 can include one or more processors (central processing units, graphical processing units, tensor processing units, machine-learning accelerators, etc.) connected to one or more memory devices. Compute resource(s) 31-2 can include a dynamic pool of available resources shared with other processes. Compute resource(s) 31-2 can include memory devices large enough to fit an entire model instance in a single memory instance. Compute resource(s) 31-2 can also shard model instance(s) across multiple memory devices (e.g., using data parallelization or tensor parallelization, etc.). Thiscan be done to increase parallelization or to execute a large model using multiple memory devices which individually might not be able to fit the entire model into memory.

[0193] Input request 33 can include data for input(s) 2. Model host 31 can process input request 33 to obtain input(s) 2. Input(s) 2 can be obtained directly from input request 33 or can be retrieved using input request 33. Input request 33 can be submitted to model host 31 via an API.

[0194] Model host 31 can perform inference over batches of input requests 33 in parallel. For instance, a model instance 31-1 can be configured with an input structure that has a batch dimension. Separate input(s) 2 can be distributed across the batch dimension (e.g., rows of an array). The separate input(s) 2 can include completely different contexts. The separate input(s) 2 can be multiple inference steps of the same task. The separate input(s) 2 can be staggered in an input structure, such that any given inference cycle can be operating on different portions of the respective input(s) 2. In this manner, for instance, model host 31 can perform inference on the batch in parallel, such that output(s) 3 can also contain the batch dimension and return the inference results for the batched input(s) 2 in parallel. In this manner, for instance, batches of input request(s) 33 can be processed in parallel for higher throughput of output payload(s) 34.

[0195] Output payload 34 can include or be based on output(s) 3 from machine- learned model(s) 1. Model host 31 can process output(s) 3 to obtain output payload 34. This can include chaining multiple rounds of inference (e.g., iteratively, recursively, across the same model(s) or different model(s)) to arrive at a final output for a task to be returned in output payload 34. Output payload 34 can be transmitted to client(s) 32 via an API.

[0196] Online learning interface(s) 36 can facilitate reinforcement learning of machine-learned model(s) 1. Online learning interface(s) 36 can facilitate reinforcement learning with human feedback (RLHF). Online learning interface(s) 36 can facilitate federated learning of machine-learned model(s) 1.

[0197] Model host 31 can access a library of pre-trained adapters or LoRA modules that can adapt a baseline model to align its outputs with a desired performance profile, augment model capabilities (e.g., to adapt to a different input modality, etc.), and the like. For instance, model host 31 can receive an input request to load a customized model, and model host 31 can retrieve one or more components to adapt a baseline model to the custom profile. Model host 31 can determine that a particular functionality is needed for a particular task (e.g., based on an output of a model that preprocesses an input) and retrieve a pre-trained component accordingly.

[0198] Model host 31 can execute machine-learned model(s) 1 to perform inference for various tasks using various types of data. For example, various different input(s) 2 and output(s) 3 can be used for various different tasks. In some implementations, input(s) 2 can be or otherwise represent image data. Machine-learned model(s) 1 can process the image data to generate an output. As an example, machine-learned model(s) 1 can process the image data to generate an image recognition output (e.g., a recognition of the image data, a latent embedding of the image data, an encoded representation of the image data, a hash of the image data, etc.). As another example, machine-learned model(s) 1 can process the image data to generate an image segmentation output. As another example, machine-learned model(s) 1 can process the image data to generate an image classification output. As another example, machine-learned model(s) 1 can process the image data to generate an image data modification output (e.g., an alteration of the image data, etc.). As another example, machine- learned model(s) 1 can process the image data to generate an encoded image data output (e.g., an encoded and / or compressed representation of the image data, etc.). As another example, machine-learned model(s) 1 can process the image data to generate an upscaled image data output. As another example, machine-learned model(s) 1 can process the image data to generate a prediction output.

[0199] In some implementations, the task is a computer vision task. In some cases, input(s) 2 includes pixel data for one or more images and the task is an image processing task. For example, the image processing task can be image classification, where the output is a set of scores, each score corresponding to a different object class and representing the likelihood that the one or more images depict an object belonging to the object class. The image processing task may be object detection, where the image processing output identifies one or more regions in the one or more images and, for each region, a likelihood that region depicts an object of interest. As another example, the image processing task can be image segmentation, where the image processing output defines, for each pixel in the one or more images, a respective likelihood for each category in a predetermined set of categories. For example, the set of categories can be foreground and background. As another example, the set of categories can be object classes. As another example, the image processing task can be depth estimation, where the image processing output defines, for each pixel in the one or more images, a respective depth value. As another example, the image processing task can be motion estimation, where the network input includes multiple images, and the image processing output defines, for each pixel of one of the input images, a motion of the scene depicted at the pixel between the images in the network input.

[0200] In some implementations, input(s) 2 can be or otherwise represent natural language data. Machine-learned model(s) 1 can process the natural language data to generate an output. As an example, machine-learned model(s) 1 can process the natural language data to generate a language encoding output. As another example, machine-learned model(s) 1 can process the natural language data to generate a latent text embedding output. As another example, machine-learned model(s) 1 can process the natural language data to generate a translation output. As another example, machine-learned model(s) 1 can process the natural language data to generate a classification output. As another example, machine-learned model(s) 1 can process the natural language data to generate a textual segmentation output. As another example, machine-learned model(s) 1 can process the natural language data to generate a semantic intent output. As another example, machine-learned model(s) 1 can process the natural language data to generate an upscaled text or natural language output (e.g., text or natural language data that is higher quality than the input text or natural language, etc.). As another example, machine-learned model(s) 1 can process the natural language data to generate a prediction output (e.g., one or more predicted next portions of natural language content).

[0201] In some implementations, input(s) 2 can be or otherwise represent speech data (e.g., data describing spoken natural language, such as audio data, textual data, etc.). Machine-learned model(s) 1 can process the speech data to generate an output. As an example, machine-learned model(s) 1 can process the speech data to generate a speech recognition output. As another example, machine-learned model(s) 1 can process the speech data to generate a speech translation output. As another example, machine-learned model(s) 1 can process the speech data to generate a latent embedding output. As another example, machine-learned model(s) 1 can process the speech data to generate an encoded speech output (e.g., an encoded and / or compressed representation of the speech data, etc.). As another example, machine-learned model(s) 1 can process the speech data to generate an upscaled speech output (e.g., speech data that is higher quality than the input speech data, etc.). As another example, machine-learned model(s) 1 can process the speech data to generate a textual representation output (e.g., a textual representation of the input speech data, etc.). As another example, machine-learned model(s) 1 can process the speech data to generate a prediction output.

[0202] In some implementations, input(s) 2 can be or otherwise represent latent encoding data (e.g., a latent space representation of an input, etc.). Machine-learned model(s) 1 can process the latent encoding data to generate an output. As an example, machine-learnedmodel(s) 1 can process the latent encoding data to generate a recognition output. As another example, machine-learned model(s) 1 can process the latent encoding data to generate a reconstruction output. As another example, machine-learned model(s) 1 can process the latent encoding data to generate a search output. As another example, machine-learned model(s) 1 can process the latent encoding data to generate a reclustering output. As another example, machine-learned model(s) 1 can process the latent encoding data to generate a prediction output.

[0203] In some implementations, input(s) 2 can be or otherwise represent statistical data. Statistical data can be, represent, or otherwise include data computed and / or calculated from some other data source. Machine-learned model(s) 1 can process the statistical data to generate an output. As an example, machine-learned model(s) 1 can process the statistical data to generate a recognition output. As another example, machine-learned model(s) 1 can process the statistical data to generate a prediction output. As another example, machine- learned model(s) 1 can process the statistical data to generate a classification output. As another example, machine-learned model(s) 1 can process the statistical data to generate a segmentation output. As another example, machine-learned model(s) 1 can process the statistical data to generate a visualization output. As another example, machine-learned model(s) 1 can process the statistical data to generate a diagnostic output.

[0204] In some implementations, input(s) 2 can be or otherwise represent sensor data. Machine-learned model(s) 1 can process the sensor data to generate an output. As an example, machine-learned model(s) 1 can process the sensor data to generate a recognition output. As another example, machine-learned model(s) 1 can process the sensor data to generate a prediction output. As another example, machine-learned model(s) 1 can process the sensor data to generate a classification output. As another example, machine-learned model(s) 1 can process the sensor data to generate a segmentation output. As another example, machine-learned model(s) 1 can process the sensor data to generate a visualization output. As another example, machine-learned model(s) 1 can process the sensor data to generate a diagnostic output. As another example, machine-learned model(s) 1 can process the sensor data to generate a detection output.

[0205] In some implementations, machine-learned model(s) 1 can be configured to perform a task that includes encoding input data for reliable and / or efficient transmission or storage (and / or corresponding decoding). For example, the task may be an audio compression task. The input may include audio data and the output may comprise compressed audio data. In another example, the input includes visual data (e.g. one or more images or videos), theoutput comprises compressed visual data, and the task is a visual data compression task. In another example, the task may comprise generating an embedding for input data (e.g. input audio or visual data). In some cases, the input includes audio data representing a spoken utterance and the task is a speech recognition task. The output may comprise a text output which is mapped to the spoken utterance. In some cases, the task comprises encrypting or decrypting input data. In some cases, the task comprises a microprocessor performance task, such as branch prediction or memory address translation.

[0206] In some implementations, the task is a generative task, and machine-learned model(s) 1 can be configured to output content generated in view of input(s) 2. For instance, input(s) 2 can be or otherwise represent data of one or more modalities that encodes context for generating additional content.

[0207] In some implementations, the task can be a text completion task. Machine- learned model(s) 1 can be configured to process input(s) 2 that represent textual data and to generate output(s) 3 that represent additional textual data that completes a textual sequence that includes input(s) 2. For instance, machine-learned model(s) 1 can be configured to generate output(s) 3 to complete a sentence, paragraph, or portion of text that follows from a portion of text represented by input(s) 2.

[0208] In some implementations, the task can be an instruction following task. Machine-learned model(s) 1 can be configured to process input(s) 2 that represent instructions to perform a function and to generate output(s) 3 that advance a goal of satisfying the instruction function (e.g., at least a step of a multi-step procedure to perform the function). Output(s) 3 can represent data of the same or of a different modality as input(s) 2. For instance, input(s) 2 can represent textual data (e.g., natural language instructions for a task to be performed) and machine-learned model(s) 1 can process input(s) 2 to generate output(s) 3 that represent textual data responsive to the instructions (e.g., natural language responses, programming language responses, machine language responses, etc.). Input(s) 2 can represent image data (e.g., image-based instructions for a task to be performed, optionally accompanied by textual instructions) and machine-learned model(s) 1 can process input(s) 2 to generate output(s) 3 that represent textual data responsive to the instructions (e.g., natural language responses, programming language responses, machine language responses, etc.). One or more output(s) 3 can be iteratively or recursively generated to sequentially process and accomplish steps toward accomplishing the requested functionality. For instance, an initial output can be executed by an external system or be processed by machine-learnedmodel(s) 1 to complete an initial step of performing a function. Multiple steps can be performed, with a final output being obtained that is responsive to the initial instructions.

[0209] In some implementations, the task can be a question answering task. Machine- learned model(s) 1 can be configured to process input(s) 2 that represent a question to answer and to generate output(s) 3 that advance a goal of returning an answer to the question (e.g., at least a step of a multi-step procedure to perform the function). Output(s) 3 can represent data of the same or of a different modality as input(s) 2. For instance, input(s) 2 can represent textual data (e.g., natural language instructions for a task to be performed) and machine- learned model(s) 1 can process input(s) 2 to generate output(s) 3 that represent textual data responsive to the question (e.g., natural language responses, programming language responses, machine language responses, etc.). Input(s) 2 can represent image data (e.g., image-based instructions for a task to be performed, optionally accompanied by textual instructions) and machine-learned model(s) 1 can process input(s) 2 to generate output(s) 3 that represent textual data responsive to the question (e.g., natural language responses, programming language responses, machine language responses, etc.). One or more output(s) 3 can be iteratively or recursively generated to sequentially process and accomplish steps toward answering the question. For instance, an initial output can be executed by an external system or be processed by machine-learned model(s) 1 to complete an initial step of obtaining an answer to the question (e.g., querying a database, performing a computation, executing a script, etc.). Multiple steps can be performed, with a final output being obtained that is responsive to the question.

[0210] In some implementations, the task can be an image generation task. Machine- learned model(s) 1 can be configured to process input(s) 2 that represent context regarding a desired portion of image content. The context can include text data, image data, audio data, etc. Machine-learned model(s) 1 can be configured to generate output(s) 3 that represent image data that depicts imagery related to the context. For instance, machine-learned model(s) 1 can be configured to generate pixel data of an image. Values for channel(s) associated with the pixels in the pixel data can be selected based on the context (e.g., based on a probability determined based on the context).

[0211] In some implementations, the task can be an audio generation task. Machine- learned model(s) 1 can be configured to process input(s) 2 that represent context regarding a desired portion of audio content. The context can include text data, image data, audio data, etc. Machine-learned model(s) 1 can be configured to generate output(s) 3 that represent audio data related to the context. For instance, machine-learned model(s) 1 can be configuredto generate waveform data in the form of an image (e.g., a spectrogram). Values for channel(s) associated with pixels of the image can be selected based on the context. Machine- learned model(s) 1 can be configured to generate waveform data in the form of a sequence of discrete samples of a continuous waveform. Values of the sequence can be selected based on the context (e.g., based on a probability determined based on the context).

[0212] In some implementations, the task can be a data generation task. Machine- learned model(s) 1 can be configured to process input(s) 2 that represent context regarding a desired portion of data (e.g., data from various data domains, such as sensor data, image data, multimodal data, statistical data, etc.). The desired data can be, for instance, synthetic data for training other machine-learned models. The context can include arbitrary data type(s). Machine-learned model(s) 1 can be configured to generate output(s) 3 that represent data that aligns with the desired data. For instance, machine-learned model(s) 1 can be configured to generate data values for populating a dataset. Values for the data object(s) can be selected based on the context (e.g., based on a probability determined based on the context).

[0213] Figure 11 is a block diagram of an example networked computing system that can perform aspects of example implementations of the present disclosure. The system can include a number of computing devices and systems that are communicatively coupled over a network 49. An example computing device 50 is described to provide an example of a computing device that can perform any aspect of the present disclosure (e.g., implementing model host 31, client(s) 32, or both). An example server computing system 60 is described as an example of a server computing system that can perform any aspect of the present disclosure (e.g., implementing model host 31, client(s) 32, or both). Computing device 50 and server computing system(s) 60 can cooperatively interact (e.g., over network 49) to perform any aspect of the present disclosure (e.g., implementing model host 31, client(s) 32, or both). Model development platform system 70 is an example system that can host or serve model development platform(s) 12 for development of machine-learned models. Third-party system(s) 80 are example system(s) with which any of computing device 50, server computing system(s) 60, or model development platform system(s) 70 can interact in the performance of various aspects of the present disclosure (e.g., engaging third-party tools, accessing third-party databases or other resources, etc.).

[0214] Network 49 can be any type of communications network, such as a local area network (e.g., intranet), wide area network (e.g., Internet), or some combination thereof and can include any number of wired or wireless links. In general, communication over network 49 can be carried via any type of wired or wireless connection, using a wide variety ofcommunication protocols (e.g., TCP / IP, HTTP, SMTP, FTP), encodings or formats (e.g., HTML, XML), or protection schemes (e.g., VPN, secure HTTP, SSL). Network 49 can also be implemented via a system bus. For instance, one or more devices or systems of Figure 11 can be co-located with, contained by, or otherwise integrated into one or more other devices or systems.

[0215] Computing device 50 can be any type of computing device, such as, for example, a personal computing device (e.g., laptop or desktop), a mobile computing device (e.g., smartphone or tablet), a gaming console or controller, a wearable computing device, an embedded computing device, a server computing device, a virtual machine operating on a host device, or any other type of computing device. Computing device 50 can be a client computing device. Computing device 50 can be an end-user computing device. Computing device 50 can be a computing device of a service provided that provides a service to an end user (who may use another computing device to interact with computing device 50).

[0216] Computing device 50 can include one or more processors 51 and a memory 52. Processor(s) 51 can be any suitable processing device (e.g., a processor core, a microprocessor, an ASIC, an FPGA, a controller, a microcontroller, etc.) and can be one processor or a plurality of processors that are operatively connected. Memory 52 can include one or more non-transitory computer-readable storage media, such as HBM, RAM, ROM, EEPROM, EPROM, flash memory devices, magnetic disks, etc., and combinations thereof. Memory 52 can store data 53 and instructions 54 which can be executed by processor(s) 51 to cause computing device 50 to perform operations. The operations can implement any one or multiple features described herein. The operations can implement example methods and techniques described herein.

[0217] Computing device 50 can also include one or more input components that receive user input. For example, a user input component can be a touch-sensitive component (e.g., a touch-sensitive display screen or a touch pad) that is sensitive to the touch of a user input object (e.g., a finger or a stylus). The touch-sensitive component can serve to implement a virtual keyboard. Other example user input components include a microphone, camera, LIDAR, a physical keyboard or other buttons, or other means by which a user can provide user input.

[0218] Computing device 50 can store or include one or more machine-learned models 55. Machine-learned models 55 can include one or more machine-learned model(s) 1, such as a sequence processing model 4. Machine-learned models 55 can include one or multiple model instance(s) 31-1. Machine-learned model(s) 55 can be received from servercomputing system(s) 60, model development platform system 70, third party system(s) 80 (e.g., an application distribution platform), or developed locally on computing device 50. Machine-learned model(s) 55 can be loaded into memory 52 and used or otherwise implemented by processor(s) 51. Computing device 50 can implement multiple parallel instances of machine-learned model(s) 55.

[0219] Server computing system(s) 60 can include one or more processors 61 and a memory 62. Processor(s) 61 can be any suitable processing device (e.g., a processor core, a microprocessor, an ASIC, an FPGA, a controller, a microcontroller, etc.) and can be one processor or a plurality of processors that are operatively connected. Memory 62 can include one or more non-transitory computer-readable storage media, such as HBM, RAM, ROM, EEPROM, EPROM, flash memory devices, magnetic disks, etc., and combinations thereof. Memory 62 can store data 63 and instructions 64 which can be executed by processor(s) 61 to cause server computing system(s) 60 to perform operations. The operations can implement any one or multiple features described herein. The operations can implement example methods and techniques described herein.

[0220] In some implementations, server computing system 60 includes or is otherwise implemented by one or multiple server computing devices. In instances in which server computing system 60 includes multiple server computing devices, such server computing devices can operate according to sequential computing architectures, parallel computing architectures, or some combination thereof.

[0221] Server computing system 60 can store or otherwise include one or more machine-learned models 65. Machine-learned model(s) 65 can be the same as or different from machine-learned model(s) 55. Machine-learned models 65 can include one or more machine-learned model(s) 1, such as a sequence processing model 4. Machine-learned models 65 can include one or multiple model instance(s) 31-1. Machine-learned model(s) 65 can be received from computing device 50, model development platform system 70, third party system(s) 80, or developed locally on server computing system(s) 60. Machine-learned model(s) 65 can be loaded into memory 62 and used or otherwise implemented by processor(s) 61. Server computing system(s) 60 can implement multiple parallel instances of machine-learned model(s) 65.

[0222] In an example configuration, machine-learned models 65 can be included in or otherwise stored and implemented by server computing system 60 to establish a client-server relationship with computing device 50 for serving model inferences. For instance, server computing system(s) 60 can implement model host 31 on behalf of client(s) 32 on computingdevice 50. For instance, machine-learned models 65 can be implemented by server computing system 60 as a portion of a web service (e.g., remote machine-learned model hosting service, such as an online interface for performing machine-learned model operations over a network on server computing system(s) 60). For instance, server computing system(s) 60 can communicate with computing device 50 over a local intranet or internet connection. For instance, computing device 50 can be a workstation or endpoint in communication with server computing system(s) 60, with implementation of machine-learned models 65 being managed by server computing system(s) 60 to remotely perform inference (e.g., for runtime or training operations), with output(s) returned (e.g., cast, streamed, etc.) to computing device 50. Machine-learned models 65 can work cooperatively or interoperatively with machine- learned models 55 on computing device 50 to perform various tasks.

[0223] Model development platform system(s) 70 can include one or more processors 71 and a memory 72. Processor(s) 71 can be any suitable processing device (e.g., a processor core, a microprocessor, an ASIC, an FPGA, a controller, a microcontroller, etc.) and can be one processor or a plurality of processors that are operatively connected. Memory 72 can include one or more non-transitory computer-readable storage media, such as HBM, RAM, ROM, EEPROM, EPROM, flash memory devices, magnetic disks, etc., and combinations thereof. Memory 72 can store data 73 and instructions 74 which can be executed by processor(s) 71 to cause model development platform system(s) 70 to perform operations. The operations can implement any one or multiple features described herein. The operations can implement example methods and techniques described herein. Example operations include the functionality described herein with respect to model development platform 12. This and other functionality can be implemented by developer tool(s) 75.

[0224] Third-party system(s) 80 can include one or more processors 81 and a memory 82. Processor(s) 81 can be any suitable processing device (e.g., a processor core, a microprocessor, an ASIC, an FPGA, a controller, a microcontroller, etc.) and can be one processor or a plurality of processors that are operatively connected. Memory 82 can include one or more non-transitory computer-readable storage media, such as HBM, RAM, ROM, EEPROM, EPROM, flash memory devices, magnetic disks, etc., and combinations thereof. Memory 82 can store data 83 and instructions 84 which can be executed by processor(s) 81 to cause third-party system(s) 80 to perform operations. The operations can implement any one or multiple features described herein. The operations can implement example methods and techniques described herein. Example operations include the functionality described herein with respect to tools and other external resources called when training or performinginference with machine-learned model(s) 1, 4, 16, 20, 55, 65, etc. (e.g., third-party resource(s) 85).

[0225] Figure 1 lillustrates one example arrangement of computing systems that can be used to implement the present disclosure. Other computing system configurations can be used as well. For example, in some implementations, one or both of computing system 50 or server computing system(s) 60 can implement all or a portion of the operations of model development platform system 70. For example, computing system 50 or server computing system(s) 60 can implement developer tool(s) 75 (or extensions thereof) to develop, update / train, or refine machine-learned models 1, 4, 16, 20, 55, 65, etc. using one or more techniques described herein with respect to model alignment toolkit 17. In this manner, for instance, computing system 50 or server computing system(s) 60 can develop, update / train, or refine machine-learned models based on local datasets (e.g., for model personalization / customization, as permitted by user data preference selections).

[0226] Figure 12 is a block diagram of an example computing device 98 that performs according to example embodiments of the present disclosure. Computing device 98 can be a user computing device or a server computing device (e.g., computing device 50, server computing system(s) 60, etc.). Computing device 98 can implement model host 31. For instance, computing device 98 can include a number of applications (e.g., applications 1 through N). Each application can contain its own machine learning library and machine- learned model(s). For example, each application can include a machine-learned model. Example applications include a text messaging application, an email application, a dictation application, a virtual keyboard application, a browser application, etc. As illustrated in Figure 12, each application can communicate with a number of other components of the computing device, such as, for example, one or more sensors, a context manager, a device state component, or additional components. In some implementations, each application can communicate with each device component using an API (e.g., a public API). In some implementations, the API used by each application is specific to that application.

[0227] Figure 13 is a block diagram of an example computing device 99 that performs according to example embodiments of the present disclosure. Computing device 99 can be the same as or different from computing device 98. Computing device 99 can be a user computing device or a server computing device (e.g., computing device 50, server computing system(s) 60, etc.). Computing device 98 can implement model host 31. For instance, computing device 99 can include a number of applications (e.g., applications 1 through N). Each application can be in communication with a central intelligence layer. Exampleapplications include a text messaging application, an email application, a dictation application, a virtual keyboard application, a browser application, etc. In some implementations, each application can communicate with the central intelligence layer (and model(s) stored therein) using an API (e.g., a common API across all applications).

[0228] The central intelligence layer can include a number of machine-learned models. For example, as illustrated in Figure 13, a respective machine-learned model can be provided for each application and managed by the central intelligence layer. In other implementations, two or more applications can share a single machine-learned model. For example, in some implementations, the central intelligence layer can provide a single model for all of the applications. In some implementations, the central intelligence layer is included within or otherwise implemented by an operating system of computing device 99.

[0229] The central intelligence layer can communicate with a central device data layer. The central device data layer can be a centralized repository of data for computing device 99. As illustrated in Figure 13, the central device data layer can communicate with a number of other components of the computing device, such as, for example, one or more sensors, a context manager, a device state component, or additional components. In some implementations, the central device data layer can communicate with each device component using an API (e.g., a private API).

[0230] The technology discussed herein makes reference to servers, databases, software applications, and other computer-based systems, as well as actions taken and information sent to and from such systems. The inherent flexibility of computer-based systems allows for a great variety of possible configurations, combinations, and divisions of tasks and functionality between and among components. For instance, processes discussed herein can be implemented using a single device or component or multiple devices or components working in combination. Databases and applications can be implemented on a single system or distributed across multiple systems. Distributed components can operate sequentially or in parallel.

[0231] While the present subject matter has been described in detail with respect to various specific example embodiments thereof, each example is provided by way of explanation, not limitation of the disclosure. Those skilled in the art, upon attaining an understanding of the foregoing, can readily produce alterations to, variations of, and equivalents to such embodiments. Accordingly, the subject disclosure does not preclude inclusion of such modifications, variations or additions to the present subject matter as would be readily apparent to one of ordinary skill in the art. For instance, features illustrated ordescribed as part of one embodiment can be used with another embodiment to yield a still further embodiment. Thus, it is intended that the present disclosure cover such alterations, variations, and equivalents.

[0232] Aspects of the disclosure have been described in terms of illustrative embodiments thereof. Any and all features in the following claims can be combined or rearranged in any way possible, including combinations of claims not explicitly enumerated in combination together, as the example claim dependencies listed herein should not be read as limiting the scope of possible combinations of features disclosed herein. Accordingly, the scope of the present disclosure is by way of example rather than by way of limitation, and the subject disclosure does not preclude inclusion of such modifications, variations or additions to the present subject matter as would be readily apparent to one of ordinary skill in the art. Moreover, terms are described herein using lists of example elements joined by conjunctions such as “and,” “or,” “but,” etc. It should be understood that such conjunctions are provided for explanatory purposes only. Clauses and other sequences of items joined by a particular conjunction such as “or,” for example, can refer to “and / or,” “at least one of’, “any combination of’ example elements listed therein, etc. Terms such as “based on” should be understood as “based at least in part on.”

[0233] The term “can” should be understood as referring to a possibility of a feature in various implementations and not as prescribing an ability that is necessarily present in every implementation. For example, the phrase “X can perform Y” should be understood as indicating that, in various implementations, X has the potential to be configured to perform Y, and not as indicating that in every instance X must always be able to perform Y. It should be understood that, in various implementations, X might be unable to perform Y and remain within the scope of the present disclosure.

[0234] The term “may” should be understood as referring to a possibility of a feature in various implementations and not as prescribing an ability that is necessarily present in every implementation. For example, the phrase “X may perform Y” should be understood as indicating that, in various implementations, X has the potential to be configured to perform Y, and not as indicating that in every instance X must always be able to perform Y. It should be understood that, in various implementations, X might be unable to perform Y and remain within the scope of the present disclosure.

Claims

WHAT IS CLAIMED IS:

1. A computer-implemented method to compression-guided decomposition of model prompts for distributed processing, the method comprising: obtaining, by a computing system comprising one or more computing devices, a model prompt associated with a context window; decomposing, by the computing system, the model prompt into a plurality of subprompts, wherein decomposing the model prompt comprises assigning, based at least in part on a compressibility of the subprompt, each of the plurality of subprompts to one of a plurality of different processing systems respectively comprising a plurality of different machine-learned models with different computational characteristics; causing, by the computing system, each of the plurality of subprompts to be respectively processed by the different processing system to which the subprompt has been mapped to generate a plurality of suboutputs; and aggregating, by the computing system, the plurality of suboutputs to generate a response to the model prompt.

2. The computer-implemented method of any preceding claim, wherein: the model prompt comprises a multimodal prompt that comprises data associated with at least two different modalities, and decomposing, by the computing system, the model prompt into the plurality of subprompts comprises decomposing, by the computing system, the model prompt into the plurality of subprompts based at least in part on transitions in the data between the at least two different data modalities.

3. The computer-implemented method of any preceding claim, wherein the plurality of different processing systems comprise: a local processing system comprising a first machine-learned model with a relatively lower computational cost; and a remote processing system comprising a second machine-learned model with a relatively higher computational cost.

4. The computer-implemented method of any preceding claim, wherein decomposing, by the computing system, the model prompt into the plurality of subprompts comprises: performing, by the computing system, a decomposition algorithm on the model prompt to generate a plurality of candidate decompositions, wherein each of the plurality of candidate decompositions comprises a different set of candidate subprompts; generating, by the computing system, a plurality of scores respectively for the plurality of candidate decompositions according to one or more scoring criteria; and selecting, by the computing system, one of the candidate decompositions based on the plurality of scores, wherein the plurality of subprompts are extracted from the selected candidate decomposition.

5. The computer-implemented method of claim 4, wherein: the decomposition algorithm assigns each of the candidate subprompts to one of a plurality of classes respectively associated with the plurality of different processing systems; and the one or more scoring criteria compare the compressibility of the candidate subprompt to the class into which such candidate subprompt is assigned.

6. The computer-implemented method of any preceding claim, wherein decomposing, by the computing system, the model prompt into the plurality of subprompts comprises decomposing, by the computing system, the model prompt into a hierarchical decomposition tree that encodes hierarchical interdependence between two or more of the plurality of subprompts.

7. The computer-implemented method of claim 6, wherein decomposing, by the computing system, the model prompt into the hierarchical decomposition tree comprises applying a constraint that weights inter-dependent nodes to be mapped to the same processing system.

8. The computer-implemented method of any preceding claim, wherein decomposing, by the computing system, the model prompt into the plurality of subprompts comprises evaluating a compressibility of each of a plurality of candidate subprompts.

9. The computer-implemented method of claim 8, wherein evaluating the compressibility of at least one of the candidate subprompts comprises subsampling data from the candidate subprompt and evaluating a compressibility of the subsampled data.

10. The computer-implemented method of claim 8, wherein evaluating the compressibility of at least one of the candidate subprompts comprises processing at least a portion of the candidate subprompt with a machine-learned compression prediction model to obtain a compressibility prediction as an output of the machine-learned compression prediction model.

11. The computer-implemented method of claim 8, wherein evaluating the compressibility of at least one of the candidate subprompts comprises evaluating one or more file formats associated with the candidate subprompt.

12. The computer-implemented method of any preceding claim, wherein decomposing, by the computing system, the model prompt into the plurality of subprompts comprises evaluating an access control status of each of a plurality of candidate subprompts.

13. A computing system, comprising: one or more processors; and one or more non-transitory computer-readable media that collectively store instructions that, when executed by the one or more processors, cause the computing system to perform operations, the operations comprising: obtaining a model prompt associated with a context window; decomposing the model prompt into a plurality of subprompts, wherein decomposing the model prompt comprises assigning, based at least in part on a compressibility of the subprompt, each of the plurality of subprompts to one of a plurality of different processing systems respectively comprising a plurality of different machine-learned models with different computational characteristics; causing each of the plurality of subprompts to be respectively processed by the different processing system to which the subprompt has been mapped to generate a plurality of sub outputs; and aggregating the plurality of suboutputs to generate a response to the model prompt.

14. The computing system of claim 13, wherein: the model prompt comprises a multimodal prompt that comprises data associated with at least two different modalities, and decomposing the model prompt into the plurality of subprompts comprises decomposing the model prompt into the plurality of subprompts based at least in part on transitions in the data between the at least two different data modalities.

15. The computing system of claim 13, wherein the plurality of different processing systems comprise: a local processing system comprising a first machine-learned model with a relatively lower computational cost; and a remote processing system comprising a second machine-learned model with a relatively higher computational cost.

16. The computing system of claim 13, wherein decomposing the model prompt into the plurality of subprompts comprises: performing a decomposition algorithm on the model prompt to generate a plurality of candidate decompositions, wherein each of the plurality of candidate decompositions comprises a different set of candidate subprompts; generating a plurality of scores respectively for the plurality of candidate decompositions according to one or more scoring criteria; and selecting one of the candidate decompositions based on the plurality of scores, wherein the plurality of subprompts are extracted from the selected candidate decomposition.

17. The computing system of claim 16, wherein: the decomposition algorithm assigns each of the candidate subprompts to one of a plurality of classes respectively associated with the plurality of different processing systems; and the one or more scoring criteria compare the compressibility of the candidate subprompt to the class into which such candidate subprompt is assigned.

18. The computing system of claim 13, wherein decomposing the model prompt into the plurality of subprompts comprises decomposing the model prompt into a hierarchicaldecomposition tree that encodes hierarchical interdependence between two or more of the plurality of subprompts.

19. The computing system of claim 18, wherein decomposing the model prompt into the hierarchical decomposition tree comprises applying a constraint that weights interdependent nodes to be mapped to the same processing system.

20. One or more non-transitory computer-readable media that collectively store instructions that, when executed by one or more computing devices, cause the one or more computing devices to perform operations, the operations comprising: obtaining a model prompt associated with a context window; decomposing the model prompt into a plurality of subprompts, wherein decomposing the model prompt comprises assigning, based at least in part on a compressibility of the subprompt, each of the plurality of subprompts to one of a plurality of different processing systems respectively comprising a plurality of different machine-learned models with different computational characteristics; causing each of the plurality of subprompts to be respectively processed by the different processing system to which the subprompt has been mapped to generate a plurality of sub outputs; and aggregating the plurality of suboutputs to generate a response to the model prompt.

Images