Communication methods, related devices, and computer program products for model training parameters

By grouping deployed devices and employing high-speed broadcasting within groups and point-to-point transmission across groups, the communication efficiency and cost issues in distributed model training are resolved, thereby improving training efficiency and scalability.

CN123123048APending Publication Date: 2026-11-17BEIJING BAIDU NETCOM SCI & TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611272345.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In distributed model deployment and training, the communication efficiency and cost between deployment devices become bottlenecks, affecting training results and overhead.

Method used

By grouping deployed devices and using high-speed broadcasting within groups and point-to-point transmission across groups, communication paths are optimized, device bandwidth pressure is alleviated, and communication efficiency is improved.

Benefits of technology

It effectively alleviates the communication pressure between devices, improves the efficiency and scalability of distributed training, and optimizes the parameter synchronization process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123048A_ABST
    Figure CN123123048A_ABST
Patent Text Reader

Abstract

This disclosure provides a communication method, related apparatus, and computer program product for model training parameters, relating to artificial intelligence technologies such as deep learning, distributed training, and data communication. One specific implementation of the method includes: broadcasting an update result generated by a first deployment device for target parameters to at least one second deployment device; controlling the second deployment device to broadcast the update result to a third deployment device in the same deployment device group as the second deployment device after receiving the update result; and broadcasting the update result to a fourth deployment device in the same deployment device group as the first deployment device. The first, second, third, and fourth deployment devices are used to jointly deploy the target model through distributed deployment. This improves the efficiency and scalability of distributed training.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, specifically to artificial intelligence technologies such as deep learning, distributed training, and data communication, and particularly to communication methods, devices, electronic devices, computer-readable storage media, and computer program products for model training parameters. Background Technology

[0002] As technology continues to advance and the demand for data processing continues to grow, the number of parameters and the overall scale of models are expanding. As a result, a single deployment device (e.g., graphics card, graphics processing unit) can no longer independently complete the deployment and training of large-scale models.

[0003] Against this backdrop, distributed deployment and distributed training technologies have emerged. Distributed deployment allows models to be allocated across multiple deployment devices, enabling collaborative training across these devices. For example, ZeRO (Zero Redundancy Optimizer) can be used for distributed training. During ZeRO training, one deployment device can act as a computing device to aggregate gradients and calculate parameter updates, which are then broadcast to other deployment devices participating in the distributed deployment and training.

[0004] In this context, the communication efficiency and capabilities between deployed devices directly impact the training performance and overhead of the model. Therefore, reducing communication costs and improving communication efficiency between deployed devices has become a crucial and increasingly important concern in distributed model deployment and training. Summary of the Invention

[0005] This disclosure provides a communication method for model training parameters, an electronic device, a computer-readable storage medium, and a computer program product.

[0006] In a first aspect, embodiments of this disclosure propose a communication method for model training parameters, applied to a first deployment device, comprising: broadcasting an update result generated by the first deployment device for target parameters to at least one second deployment device; controlling the second deployment device to broadcast the update result to a third deployment device in the same deployment device group as the second deployment device via intra-group broadcast after receiving the update result; and broadcasting the update result to a fourth deployment device in the same deployment device group as the first deployment device, wherein the first deployment device, the second deployment device, the third deployment device, and the fourth deployment device are used to jointly deploy the target model in a distributed deployment manner.

[0007] Secondly, embodiments of this disclosure propose a communication device for model training parameters, applied to a first deployment device, comprising: a first broadcast unit configured to broadcast update results generated by the first deployment device for target parameters to at least one second deployment device; a broadcast control unit configured to control the second deployment device to broadcast the update results to a third deployment device in the same deployment device group as the second deployment device via intra-group broadcast after receiving the update results; and a second broadcast unit configured to broadcast the update results to a fourth deployment device in the same deployment device group as the first deployment device, wherein the first deployment device, the second deployment device, the third deployment device, and the fourth deployment device are used to jointly deploy the target model via distributed deployment.

[0008] Thirdly, embodiments of this disclosure provide an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to implement a communication method for model training parameters as described in any implementation of the first aspect.

[0009] Fourthly, embodiments of this disclosure provide a non-transitory computer-readable storage medium storing computer instructions for enabling a computer to communicate model training parameters as described in any implementation of the first aspect.

[0010] Fifthly, embodiments of this disclosure provide a computer program product including a computer program that, when executed by a processor, enables a communication method for model training parameters as described in any implementation of the first aspect.

[0011] The communication method, apparatus, electronic device, computer-readable storage medium, and computer program product for model training parameters provided in this disclosure firstly broadcasts the updated results generated by the first deployment device for the target parameters to at least one second deployment device; then, controls the second deployment device to broadcast the updated results to a third deployment device in the same deployment device group as the second deployment device after receiving the updated results, via group broadcast; finally, broadcasts the updated results to a fourth deployment device in the same deployment device group as the first deployment device, wherein the first deployment device, the second deployment device, the third deployment device, and the fourth deployment device are used to jointly deploy the target model in a distributed deployment manner.

[0012] This disclosure splits global broadcast into cross-group fixed-point transmission and intra-group high-speed broadcast by grouping deployed devices. This enables the acceleration of parameter synchronization by utilizing deployed device groups while alleviating the bandwidth pressure on source devices, thereby improving the efficiency and scalability of distributed training.

[0013] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0014] Other features, objects, and advantages of this disclosure will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is an exemplary system architecture to which this disclosure can be applied; Figure 2 A flowchart illustrating a communication parameter process provided in this embodiment of the disclosure; Figure 3 A flowchart illustrating a process for generating an updated result, provided as an embodiment of this disclosure; Figure 4 A flowchart illustrating the process of setting communication parameters in a specific application scenario, as provided in an embodiment of this disclosure; Figure 5 A structural block diagram of a communication device for model training parameters provided in an embodiment of this disclosure; Figure 6 This is a schematic diagram of the structure of an electronic device for a communication method for executing model training parameters, provided in an embodiment of this disclosure. Detailed Implementation

[0015] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding; these should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description. It should be noted that, unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0016] Furthermore, if the technical solutions disclosed herein involve information that falls under the category of user personal information (for example, target parameters may be trained and determined based on input samples that are samples of user personal information), then in such cases, the acquisition, storage, use, processing, transportation, provision, and disclosure of such information in this disclosure comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0017] Figure 1 An exemplary system architecture 100 is shown, illustrating embodiments of communication methods, apparatuses, electronic devices, and computer-readable storage media to which the model training parameters of this disclosure can be applied.

[0018] like Figure 1 As shown, system architecture 100 may include servers 101, 102, and 103, and network 104. Network 104 serves as the medium for providing communication links between servers 101, 102, and 103. Network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0019] Each server may include at least one deployment device (e.g., graphics card, graphics processor, etc.) to enable model deployment (e.g., deploying the model in a distributed manner) by using multiple deployment devices within a server or by using multiple deployment devices across servers. For example, server 101 may include deployment devices 110, 111...11N (where N is a positive integer).

[0020] In practice, the deployed model can be a "large model" with many model parameters and a large model size, such as a Large Language Model (LLM).

[0021] In some embodiments, architecture 100 may also include "terminal devices" with computing resources, such as smartphones and tablets. For example, in scenarios such as federated deployments, the deployment devices may also be at least partially these "terminal devices".

[0022] In practice, servers 101, 102, and 103 can be either hardware or software. When servers 101, 102, and 103 are hardware, they can be implemented as a distributed server cluster consisting of multiple servers, or as a single server. When servers 101, 102, and 103 are software, they can be implemented as multiple software programs or software modules, or as a single software program or software module; no specific limitations are made here. Similarly, if there is a terminal device, it can also be either software or hardware. For example, when the terminal device is hardware, it can be various electronic devices with a display screen, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0023] In practice, considering that distributed deployment and the maintenance and updating of a specific parameter during training are usually performed by the deployment device, such as deployment device 110, the communication device for model training parameters can generally also be set in deployment device 110.

[0024] It should be understood that Figure 1 The number of networks, servers, and deployment devices shown is merely illustrative. Depending on implementation needs, there can be any number of networks, servers, deployment devices (and terminal devices).

[0025] Please refer to Figure 2 , Figure 2 A flowchart of a communication parameter process provided for an embodiment of this disclosure, including process 200.

[0026] Process 200 specifically includes the following steps: Step 201: Broadcast the updated results generated by the first deployment device for the target parameters to at least one second deployment device; In embodiments of this disclosure, deployment devices used to deploy a target model (e.g., the large model described above) may be grouped to form one or at least two deployment device groups.

[0027] For example, the target model could be a generative model that processes user-input instructions (e.g., text information in natural language form) to output corresponding text (e.g., answers to questions, written code), images, audio, and video to the user.

[0028] For deployment device groups, deployment devices can be pre-grouped based on criteria such as whether they belong to the same central processing unit, the same physical host, or the same motherboard, to form one or at least two deployment device groups.

[0029] For example, when grouping based on whether they belong to the same physical host or the same motherboard, each deployment device group can correspond to a physical host separately. That is, each deployment device (e.g., a graphics card) belonging to the same physical host can be considered to belong to the same deployment device group.

[0030] In some embodiments, when grouping deployed devices, the groups can also be based on specific communication conditions and metrics between the deployed devices. For example, communication metrics (e.g., communication rate, packet loss rate, etc.) between each deployed device and other deployed devices can be obtained, and a communication relationship graph can be constructed based on these metrics. Then, the deployed devices can be grouped using clustering. For example, clustering can be used to group deployed devices whose communication rates are greater than or equal to a communication rate threshold, so that in the clustering results, the communication rates between all pairs of deployed devices are greater than or equal to the communication rate threshold.

[0031] Therefore, based on the actual situation between the deployed devices, they can be grouped to more effectively utilize their communication capabilities and resources, thus alleviating the communication pressure on the "first deployed device" (discussed later). In other words, the overall communication pressure on the "first deployed device" can be reduced through inter-group communication and data transfer.

[0032] In some embodiments, if the grouping condition is selected based on the fact that the communication rate between any two deployed devices in the deployed device group is greater than or equal to a communication rate threshold, the specific value of the communication rate threshold can be further referenced to the communication rate of the first deployed device when communicating externally.

[0033] For example, since the first deployed device usually has both communication with the same physical machine and communication across physical machines, and given that the local protocol and bus used for communication with the same physical machine often have strong communication capabilities, the communication rate threshold can be set by referring to the optimal or better communication rate when communicating across physical machines.

[0034] This allows for the transformation of communication between deployed devices and the execution entity that is less efficient in direct communication with the execution entity into more efficient intra-group communication after grouping deployed devices. This enables the use of intra-group communication within deployed device groups to further improve the overall communication rate.

[0035] Next, in this step, the entity responsible for implementing the communication method aimed at training parameters from the model (e.g., Figure 1 After generating an update result for the target parameters, the deployment device 110 shown broadcasts the update result to at least one second deployment device.

[0036] In the embodiments of this disclosure, for ease of distinction, the “executing entity”, that is, the deployment device that provides the update results, can be described as the first deployment device. In other words, the “executing entity” and the “first deployment device” can be used interchangeably.

[0037] Target parameters (e.g., the feedforward network weight matrix of the target model, the projection matrix of the attention head, the gate network weights when the hybrid expert network is used as the target model, specific slices of the word embedding matrix, scaling parameters for layer normalization, etc.) can be pre-assigned to the execution agent so that after completing gradient reduction and determining the global gradient, the execution agent can calculate the updated result for the target parameters based on the global gradient. For example, a portion of the deployment devices can be selected as the first deployment devices, and the target parameters to be processed and calculated by the first deployment device can be assigned to it.

[0038] Accordingly, the first deployment device (i.e. the execution entity) can utilize the global gradient to calculate the update result of the global gradient corresponding to the target parameters it is assigned in this iteration.

[0039] It should be understood that in practice, the endpoint of this "gradient reduction," that is, the deployment device that actually completes the reduction, can be either the executing entity or other deployment devices outside of the executing entity. In such cases, if the global reduction and global gradient are generated by deployment devices outside of the executing entity, the executing entity can obtain the global gradient by communicating with them, in order to generate and form the updated results for the target parameters.

[0040] Accordingly, after generating the update result, the executing entity in this step can broadcast the update result to at least one second deployment device.

[0041] The second deployment device can be understood as a deployment device located in a different deployment device group than the execution subject. Furthermore, each deployment device group can include at least one "second deployment device" relative to the execution subject. For example, a deployment device can be pre-selected from each deployment device group as the second deployment device corresponding to the execution subject.

[0042] Accordingly, through such broadcasting, the executing entity can actually transmit and send the updated results it generates to each deployment device group via the second deployment device.

[0043] Furthermore, it should be understood that whether it is the first deployment device, the second deployment device, or the third and fourth deployment devices that will be discussed in detail later, they can be understood as a collective deployment of the target model (such as the LLM model mentioned above) through a distributed deployment method.

[0044] In some embodiments, if the second deployment device in each deployment device group relative to the executing entity is determined in advance (e.g., by pre-specified commands or determination rules), then in such cases, the executing entity may also read the first communication indicators between the first deployment device and each of the second deployment devices before actually broadcasting.

[0045] If the executing entity can determine based on the first communication indicator that it can currently communicate normally with each of the second deployed devices, that is, the communication status between the executing entity and each of the second deployed devices is in a normal communication state, then the executing entity can respond by broadcasting to these second deployed devices.

[0046] If the executing entity determines, based on the first communication indicator, that there is a blocked second deployment device in the second deployment device that is in a blocked communication state with the first deployment device, then the executing entity can respond by reading the second communication indicator between the first deployment device and each deployment device in the target deployment device group where the blocked second deployment device is located.

[0047] Then, based on the second communication metric, the executing entity determines the updated second deployment device from the deployment devices in the target deployment device group to replace the blocking second deployment device. For example, the "deployment device" with the strongest and best communication capability indicated by the second communication metric is selected as the updated second deployment device, and it is used to replace the blocking second deployment device, serving as the "representative" of the target deployment device group to obtain the update result.

[0048] This is to avoid broadcast failures or delays caused by single-point communication failures, improve communication reliability and system fault tolerance, and ensure the smooth execution of distributed tasks and overall efficiency.

[0049] In some embodiments, if the correspondence between the executing entity and the second deployment device is not predetermined, the executing entity may first determine the second deployment device corresponding to the first deployment device in each deployment device group based on the third communication index between each deployment device in the deployment device group and the first deployment device. For example, similarly as described above, the executing entity may select the deployment device with the optimal communication rate and strongest communication capability indicated by the third communication index in each deployment device group as the "second deployment device".

[0050] Then, after identifying each second deployment device, the executing entity broadcasts the updated results generated by the first deployment device for the target parameters to these second deployment devices.

[0051] Therefore, the executing entity can dynamically select the deployment device with the best current communication capability to perform cross-group communication based on this real-time communication indicator, thereby effectively improving the global communication efficiency of update results.

[0052] In some embodiments, if the number of deployment devices included in each deployment device group is the same, for example, if the grouping is based on whether they belong to the same physical machine, and the same deployment devices are set in each physical machine, then in such a case, the execution entity can use the same logical numbering rules to assign corresponding logical numbers to the deployment devices in each deployment device group.

[0053] For example, the implementing entity uses a 0-N numbering method to number each deployed device in each deployed device group.

[0054] Then, when determining the corresponding second deployment device, the executing entity can choose to identify the deployment device with the same logical number (described as the second logical number for convenience) in each deployment device group as the second deployment device, based on the first logical number of the first deployment device. That is, the second deployment devices identified in each deployment device group are assigned the same second logical number as the first logical number in their respective deployment device groups.

[0055] For example, if the logical number of the executing entity is "0", the executing entity can use the deployment device with logical number "0" in each deployment device group as the second deployment device.

[0056] The implementing entity can then broadcast the "update results" to these second-deployed devices as discussed above, which will not be repeated here.

[0057] Therefore, by aligning logical numbers, not only can the selection process of cross-group communication nodes be simplified and the efficiency of determining the second deployment device as the "representative" be improved, but the correspondence of the same logical numbers can also make the communication topology between groups more regular and symmetrical, avoiding communication hotspots or resource imbalances caused by uneven node selection, thereby improving the coordination efficiency of distributed tasks and the maintainability of the overall system. For example, when assigning corresponding execution entities to each "parameter" in the future, the global load can be mapped based on the load situation within a single deployment device group, thereby making it easier to assess and schedule the global load.

[0058] Step 202: Control the second deployment device to broadcast the update result to the third deployment device in the same deployment device group as the second deployment device after receiving the update result. In the embodiments of this disclosure, after the executing entity provides the update result to the second deployment device based on step 201 described above, it can further control the second deployment device to broadcast the update result to other deployment devices (referred to as the third deployment device for ease of description) within the same deployment device group as the second deployment device upon receiving the update result. Thus, hierarchical global communication of the update result is achieved using the second deployment device; that is, global communication and transmission of the update result are achieved by first performing inter-group communication and then intra-group broadcasting.

[0059] Step 203: Broadcast the update result to the fourth deployment device, which is in the same deployment device group as the first deployment device.

[0060] In the embodiments of this disclosure, the executing entity can also broadcast the update results to other deployment devices in the deployment device group (referred to as the fourth deployment device for ease of description) within its own deployment device group, thereby completing the propagation of the update results at the group level and ultimately achieving global communication of the update results.

[0061] It should be understood that, in different embodiments, there is no strict execution order restriction between steps 202 and 203, and they can be flexibly arranged according to actual needs. For example, the executing entity can execute step 203 before step 202, or it can execute step 202 in parallel with step 203 to improve communication efficiency and reduce waiting delays.

[0062] In some embodiments, step 203 may even be executed before step 201, enabling the executing entity to prioritize or synchronize with the process of providing update results to other deployed device groups, thereby completing the transmission of update results within its deployed device group and further reducing the overall time consumption of global communication.

[0063] The communication method for model training parameters provided in this disclosure first broadcasts the updated results generated by the first deployment device for the target parameters to at least one second deployment device. Then, the second deployment device, after receiving the updated results, broadcasts them to a third deployment device within the same deployment device group as the second deployment device via intra-group broadcast. Finally, the updated results are broadcast to a fourth deployment device within the same deployment device group as the first deployment device. The first, second, third, and fourth deployment devices are used to jointly deploy the target model in a distributed deployment manner. Thus, by grouping the deployment devices, global broadcasting is split into cross-group fixed-point transmission and intra-group high-speed broadcasting. This allows for the acceleration of parameter synchronization using deployment device groups while alleviating bandwidth pressure on the source device, thereby improving the efficiency and scalability of distributed training.

[0064] In some embodiments, the executing entity can also similarly utilize the hierarchical communication mechanism within and between groups described above to perform the global gradient reduction process, thereby improving communication efficiency during the reduction process and saving communication resources consumed during the reduction process. For ease of understanding, please refer to [the relevant documentation / reference]. Figure 3 Let's have a discussion.

[0065] Figure 3 A flowchart of a process for generating an updated result provided in an embodiment of this disclosure includes process 300. For example, process 300 may be executed by an execution entity prior to process 200.

[0066] Step 301: Control the third deployment device to transmit the generated updated sub-gradient for the target parameters to the second deployment device in the same deployment device group as the third deployment device; Specifically, in this step, the executing entity can send a control command (referred to as the "first control command" for convenience) to the third deployment device in each deployment device group. The first control command instructs and controls the third deployment device to complete the calculation or generation of the updated sub-gradient for the target parameter locally, and then transmits the updated sub-gradient to the second deployment device in the same deployment device group.

[0067] In practice, the third deployment device can generate updated sub-gradients based on its locally stored training data or model parameters, according to preset gradient calculation rules.

[0068] After completing and generating the updated sub-gradient, the third deployment device can send the updated sub-gradient to the second deployment device in the same deployment device group through the intra-group communication link, using point-to-point transmission or intra-group aggregation transmission.

[0069] In some embodiments, the selection of the second deployment device is often based on its communication capabilities with the execution entity, and its gradient aggregation capabilities may not be the strongest. Therefore, in different embodiments, the execution entity can take into account communication capabilities and processing capabilities (e.g., after generating their respective capability scores, calculate the total score based on predetermined weights) to determine whether to use the second deployment device to update the aggregation of sub-gradients or to utilize other deployment devices in the deployment device group.

[0070] Next, taking the case of using a second deployment device for aggregation as an example, the execution entity can continue to execute step 302 after step 301.

[0071] Step 302: Control the second deployment device to form the intra-group gradient of the target parameters on the second deployment device based on the transmitted updated sub-gradient; Specifically, similar to the above, the executing entity can send a control command to the second deployment device (for ease of description, it can be described as a second control command) to instruct the second deployment device to perform an aggregation operation on the received update sub-gradients and the update sub-gradients formed locally by the second deployment device after receiving update sub-gradients transmitted from each third deployment device in the same deployment device group, so as to form the intra-group gradient of the target parameter.

[0072] In practice, aggregation operations can be implemented using gradient accumulation, weighted averaging, or other preset aggregation methods. For example, the second deployment device can allocate a buffer in its local memory or video memory to temporarily store each received update sub-gradient, and trigger aggregation calculation when preset conditions are met. The preset conditions are, for example, receiving all update sub-gradients from the third deployment device, reaching a preset waiting time threshold, or receiving an aggregation trigger command issued by the execution entity.

[0073] After the aggregation is completed, the second deployment device obtains the intra-group gradient, which characterizes the overall gradient update direction and magnitude for the target parameter within the deployment device group.

[0074] Step 303: Control the second deployment device to provide the group gradient to the first deployment device; Specifically, similar to the above, the executing entity can choose to send a control command to the second deployment device (which can be described as a third control command for ease of description) to instruct the second deployment device to transmit the intra-group gradient to the executing entity (first deployment device) through the cross-group communication link after completing the formation of the intra-group gradient.

[0075] In some embodiments, the second deployment device can encapsulate the gradients within a group into a data packet of a preset format, and include the deployment device group identifier to which the gradients belong, the identifier of the target parameter, and the corresponding iteration round information, so that the execution entity can identify and summarize them.

[0076] In practice, each secondary deployment device can provide its own intra-group gradient to the execution entity according to a preset synchronous or asynchronous transmission strategy to adapt to different scenario requirements. For example, the execution entity can control all secondary deployment devices to provide intra-group gradients at once, or it can set rules to allow the secondary deployment devices to determine the timing of sending and providing intra-group gradients themselves.

[0077] Step 304: Based on the intra-group gradient, form the global gradient of the target parameter; Specifically, after receiving the intra-group gradients from each deployed device group, the executing entity can perform a global aggregation operation on the received intra-group gradients to form a global gradient of the target parameters.

[0078] For example, the executing entity can achieve global gradient aggregation by accumulating the gradients within each group and then calculating the mean, or by weighting the aggregation based on factors such as the amount of data and the number of devices in each group.

[0079] Step 305: Generate updated results for the target parameters based on the global gradient.

[0080] Specifically, after obtaining the global gradient based on step 304 above, the execution entity can generate an update result for the target parameter based on the global gradient and the current value of the target parameter, according to a preset parameter update rule.

[0081] In practice, parameter update rules can be determined based on, for example, stochastic gradient descent, optimizers, or other parameter update algorithms suitable for distributed training.

[0082] In some embodiments, the update result may be either the updated value of the target parameter or the amount of update of the target parameter relative to the current value.

[0083] Subsequently, the executing entity can store the generated update results locally and, in accordance with the process discussed in, for example, process 200 above, transmit the update results layer by layer to each deployment device in each deployment device group (i.e., the second deployment device and the third deployment device) through a layered communication mechanism that combines inter-group communication and intra-group broadcasting, so as to complete the synchronous update of the target parameters globally.

[0084] In some embodiments, for the sake of data timeliness, the executing entity may choose to control the third deployment device to transmit the updated sub-gradient for the target parameter to the second deployment device in the same deployment device group, and to control the second deployment device to provide the group gradient to the first deployment device, to control the third deployment device and the second deployment device to perform a transmission communication once each time a transmittable data block (e.g., updated sub-gradient, group gradient) is formed.

[0085] In some embodiments, the executing entity may also choose to use a volume threshold to manage the sending behavior of the second and / or third deployed devices to avoid excessive communication overhead caused by frequent data block sending and communication operations, thus avoiding wasted communication overhead.

[0086] Taking the process of controlling the second deployment device to provide the gradient within the group to the execution subject as an example, the execution subject can control the second deployment device to first open a buffer in the local memory or video memory of the second deployment device to temporarily store the gradient data within the group, and control the second deployment device to store the updated sub-gradients transmitted from each third deployment device in the same deployment device group into the buffer first, instead of immediately sending them to the execution subject.

[0087] Accordingly, as the third deployment device continuously transmits updated sub-gradients, the second deployment device can first aggregate them, form data blocks corresponding to the gradients within the group, and store them in the buffer.

[0088] Then, the executing entity can control the second deployed device to continuously detect the total volume of the data blocks of the formed intra-group gradients, and in response to the total volume of the data blocks of the formed intra-group gradients being greater than or equal to the volume threshold, transmit these collected and formed intra-group gradients to the executing entity.

[0089] In practice, the volume threshold can be preset or dynamically adjusted based on factors such as the actual communication environment, network bandwidth, hardware resources, and historical transmission efficiency. For example, in scenarios with high communication bandwidth and low transmission latency, a smaller volume threshold can be set to reduce waiting time and improve the timeliness of gradient reduction. Conversely, in scenarios with limited communication bandwidth or severe network congestion, a larger volume threshold can be set to reduce the number of cross-machine communications and lower communication overhead.

[0090] Therefore, by using a volume threshold as a trigger condition, the second deployment device can batch merge multiple gradient data blocks within a group and then uniformly provide them to the first deployment device, thereby effectively reducing the communication startup overhead and latency accumulation caused by frequent small data volume communication and improving the communication efficiency of global gradient reduction.

[0091] In some embodiments, the condition for triggering such communication and providing the action can also be the number of data blocks.

[0092] Similarly, taking the process of controlling the second deployment device to provide the intra-group gradient to the first deployment device as an example, the executing entity can control the second deployment device to record the number of data blocks of the generated intra-group gradient when receiving and aggregating the updated sub-gradient transmitted by the third deployment device.

[0093] If the number of data blocks for the intra-group gradient already formed on the second deployment device is greater than or equal to a threshold, the second deployment device can respond by triggering a communication, as discussed above, to provide the currently formed and cached intra-group gradients to the first deployment device. Thus, by introducing a block number threshold, communication behavior can be managed from the perspective of communication frequency, allowing for more efficient use of communication resources.

[0094] In some embodiments, the executing entity may control and manage the communication actions and behaviors based on both quantity thresholds and volume thresholds. For example, the executing entity may control the second deployed device to trigger transmission only when either the volume condition or the quantity condition is met, or to trigger transmission only when both conditions are met simultaneously.

[0095] Accordingly, by introducing a block number threshold, a decision can be made on whether to trigger a transmission action when the size of consecutive data blocks is small, thus avoiding excessive transmission delays caused by continuous waiting for the volume to accumulate. This allows for both the advantages of batch transmission and the timeliness of gradient reduction.

[0096] For example, when there are many parameters and the update sub-gradient of each parameter is small, if the judgment is based solely on the volume threshold, the buffer may fail to meet the triggering condition for a long time. In this case, the decision-making method based on the number of blocks can trigger the transmission more promptly, thereby avoiding the communication channel from being idle for a long time.

[0097] In some embodiments, while controlling the second deployment device to provide intra-group gradients to the first deployment device, the executing entity may also choose to control the second deployment device to periodically provide intra-group gradients to the first deployment device.

[0098] Specifically, the executing entity can control the second deployment device to periodically provide the currently formed intra-group gradient in its buffer to the first deployment device according to a preset time period.

[0099] Typically, this time period can be set based on factors such as training iteration step size, network status, and system load, and can be dynamically adjusted during training.

[0100] Therefore, by periodically triggering transmission, the second deployment device can regularly and reliably send the currently accumulated intra-group gradients to the execution subject, so as to ensure that the global gradient reduction process is stable and predictable.

[0101] Correspondingly, periodic transmission can also be combined with the aforementioned volume and quantity thresholds. For example, periodic transmission can be used as a safety net mechanism, forcing transmission when the data volume has not reached the threshold but the waiting time exceeds a preset period, thereby achieving a balance between communication efficiency and timeliness. In this way, the second deployment device can provide intra-group gradients to the executing entity at an orderly and controllable pace, improving the stability and predictability of the overall communication process.

[0102] Accordingly, as discussed above, during the transmission and transfer of updated sub-gradients from the third deployment device to the second deployment device, the transmission process can be controlled individually or jointly by using volume thresholds (the volume threshold here can be the same as or different from the volume threshold corresponding to the total volume of the data blocks of the intra-group gradient mentioned above), block number, and periodicity, so as to achieve a similar technical purpose as discussed above, which will not be repeated here.

[0103] Similarly, in some embodiments, the executing entity may use volume thresholds, block counts, or periodicity to control the broadcasting process (e.g., triggering broadcast actions) individually or in combination during the broadcasting process (e.g., broadcasting update results to the second deployment device) to achieve similar technical objectives as discussed above, which will not be repeated here.

[0104] To enhance understanding, this disclosure also provides an implementation scheme for a specific application scenario, which can be referred to accordingly. Figure 4 Please provide an explanation. Figure 4 A flowchart of a process for implementing communication parameters in a specific scenario according to an embodiment of this disclosure, including process 400.

[0105] Furthermore, for ease of understanding of process 400, reference can also be made to the exemplary architecture 100 to illustrate process 400. For example, process 400 may include servers 101, 102, and 103, and each server may include three deployment devices. For example, server 101 includes deployment devices 110, 111, and 112, server 102 includes deployment devices 410, 411, and 412, and server 103 includes deployment devices 420, 421, and 422.

[0106] Accordingly, in process 400, each server can be understood as a deployment device group, that is, deployment devices belonging to the same server are in or form a deployment device group.

[0107] It should be understood that the number of servers and the number of devices deployed on the servers mentioned above are merely examples for ease of understanding and are not intended to limit the specific number of servers or deployed devices.

[0108] Next, in process 400, the "target parameter" can be managed, updated, and maintained by deployment device 110. Accordingly, after generating an update result 405 for the target parameter, deployment device 110 can execute S401 to broadcast the generated update result 405 for the target parameter to deployment devices determined as the second deployment device (e.g., those determined based on logical numbers, which, when deployment device 110 is the aforementioned "first deployment device," can act as the second deployment device to its corresponding deployment devices 410 and 420). This achieves "inter-group transmission and communication" of the update result 405.

[0109] Then, after broadcasting the update result 405, deployment device 110 can control deployment devices 410 and 420 to execute S402, so that after receiving the update result 405, they can broadcast the update result 405 to the third deployment device in the same deployment device group via intra-group broadcast. For example, deployment device 410 can broadcast the update result 405 to deployment devices 411 and 412, which are third deployment devices in the same deployment device group; deployment device 420 can broadcast the update result 405 to deployment devices 421 and 422, which are third deployment devices in the same deployment device group.

[0110] Additionally, as discussed above, deployment device 110 may execute S403 before S401, or in parallel with S401, or in parallel with S402, or after S402, to broadcast update result 405 to deployment devices 111 and 112, which are the fourth deployment devices in the same deployment device group as deployment device 110, so as to fully realize the "intra-group transmission and communication" of update result 405.

[0111] Further reference Figure 5 As an implementation of the methods shown in the above figures, this disclosure provides an embodiment of a communication device for model training parameters, which is similar to... Figure 2 Corresponding to the method embodiment shown, the device can be specifically applied to a first deployment device, such as the deployment device 110 described above.

[0112] like Figure 5 As shown, the communication device 500 for model training parameters in this embodiment may include: a first broadcast unit 501, a broadcast control unit 502, and a second broadcast unit 503. The first broadcast unit 501 is configured to broadcast the update results generated by the first deployment device for the target parameters to at least one second deployment device. The broadcast control unit 502 is configured to control the second deployment device to broadcast the update results to a third deployment device in the same deployment device group as the second deployment device after receiving the update results, via intra-group broadcasting. The second broadcast unit 503 is configured to broadcast the update results to a fourth deployment device in the same deployment device group as the first deployment device. The first, second, third, and fourth deployment devices are used to jointly deploy the target model in a distributed deployment manner.

[0113] In this embodiment, the specific processing of the first broadcast unit 501, the broadcast control unit 502, and the second broadcast unit 503 in the communication device 500 for model training parameters, and their resulting technical effects, can be found in reference to [reference needed]. Figure 2The relevant descriptions of steps 201-203 in the corresponding embodiments will not be repeated here.

[0114] In some optional implementations of this embodiment, the apparatus 500 further includes: an update sub-gradient transmission control unit configured to control the third deployment device to transmit the generated update sub-gradient for the target parameter to a second deployment device in the same deployment device group as the third deployment device; an intra-group gradient forming control unit configured to control the second deployment device to form an intra-group gradient of the target parameter in the second deployment device based on the transmitted update sub-gradient; an intra-group gradient providing control unit configured to control the second deployment device to provide the intra-group gradient to the first deployment device; a global gradient forming unit configured to form a global gradient of the target parameter based on the intra-group gradient; and an update result generation unit configured to generate an update result for the target parameter based on the global gradient.

[0115] In some optional implementations of this embodiment, the intra-group gradient providing control unit is further configured to control the second deployment device to provide the intra-group gradient to the first deployment device in response to the total volume of the data blocks of the intra-group gradient already formed on the second deployment device being greater than or equal to a volume threshold.

[0116] In some optional implementations of this embodiment, the intra-group gradient providing control unit is further configured to control the second deployment device to provide the intra-group gradient to the first deployment device in response to the number of data blocks of the intra-group gradient already formed on the second deployment device being greater than or equal to a number threshold.

[0117] In some optional implementations of this embodiment, the intra-group gradient providing control unit is further configured to control the second deployment device to periodically provide intra-group gradients to the first deployment device.

[0118] In some optional implementations of this embodiment, the apparatus 500 further includes: a first communication index reading unit configured to read a first communication index between the first deployment device and each of the second deployment devices in response to the second deployment devices being predetermined; a second communication index reading unit configured to read a second communication index between the first deployment device and each deployment device in the target deployment device group to which the blocking second deployment device is located in response to the determination, based on the first communication index, that there is a blocking second deployment device among the second deployment devices that is currently in a blocking communication state with the first deployment device; and a deployment device replacement unit configured to determine, based on the second communication index, an updated second deployment device for replacing the blocking second deployment device from the deployment devices in the target deployment device group.

[0119] In some optional implementations of this embodiment, the first broadcast unit 501 includes: a deployment device determination subunit, configured to, in response to the second deployment device not being pre-determined, determine, for the deployment device group, a second deployment device corresponding to the first deployment device based on a third communication index between each deployment device in the deployment device group and the first deployment device; and an update parameter broadcast subunit, configured to broadcast the update result generated by the first deployment device for the target parameter to at least one second deployment device.

[0120] In some optional implementations of this embodiment, the apparatus 500 further includes: a deployment device numbering unit, configured to assign corresponding logical numbers to deployment devices in each deployment device group using the same logical numbering rule in response to the fact that the number of deployment devices included in each deployment device group is the same; and a deployment device determining unit, configured to determine a second deployment device in each deployment device group based on the first logical number of the first deployment device, wherein the second logical number assigned to the second deployment device in its deployment device group is the same as the first logical number.

[0121] In some optional implementations of this embodiment, the communication rate between different deployed devices in the deployment device group is greater than or equal to the communication rate threshold.

[0122] This embodiment exists as a device embodiment corresponding to the above method embodiment. The communication device for model training parameters provided in this embodiment splits global broadcast into cross-group fixed-point transmission and intra-group high-speed broadcast by grouping the deployed devices. This enables the acceleration of parameter synchronization by utilizing the deployed device group while alleviating the bandwidth pressure of the source device, thereby improving the efficiency and scalability of distributed training.

[0123] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0124] Figure 6 A schematic block diagram of an example electronic device 600 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0125] like Figure 6As shown, device 600 includes a computing unit 601, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 602 or a computer program loaded into random access memory (RAM) 603 from storage unit 608. RAM 603 may also store various programs and data required for the operation of device 600. The computing unit 601, ROM 602, and RAM 603 are interconnected via bus 604. Input / output (I / O) interface 605 is also connected to bus 604.

[0126] Multiple components in device 600 are connected to I / O interface 605, including: input unit 606, such as keyboard, mouse, etc.; output unit 607, such as various types of monitors, speakers, etc.; storage unit 608, such as disk, optical disk, etc.; and communication unit 609, such as network card, modem, wireless transceiver, etc. Communication unit 609 allows device 600 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0127] The computing unit 601 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 601 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 601 performs the various methods and processes described above, such as the communication method for model training parameters. For example, in some embodiments, the communication method for model training parameters may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 608. In some embodiments, part or all of the computer program may be loaded and / or installed on device 600 via ROM 602 and / or communication unit 609. When the computer program is loaded into RAM 603 and executed by the computing unit 601, one or more steps of the communication method for model training parameters described above may be performed. Alternatively, in other embodiments, the computing unit 601 may be configured to perform the communication method for model training parameters by any other suitable means (e.g., by means of firmware).

[0128] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0129] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0130] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0131] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0132] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0133] Computer systems can include clients and servers. Clients and servers are generally geographically separated and typically interact via communication networks. The client-server relationship is established by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, also known as cloud computing servers or cloud hosts, which are hosting products within the cloud computing service ecosystem to address the management difficulties and weak business scalability inherent in traditional physical hosts and Virtual Private Servers (VPS) services. Servers can also be categorized as distributed system servers or servers incorporating blockchain technology.

[0134] According to the technical solution of this disclosure, by grouping the deployed devices, global broadcasting is split into cross-group fixed-point transmission and intra-group high-speed broadcasting. This enables the deployment device group to accelerate parameter synchronization and improve the efficiency and scalability of distributed training while alleviating the bandwidth pressure of the source device.

[0135] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution provided in this disclosure can be achieved, and this is not limited herein.

[0136] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for communicating model training parameters, applied to a first deployment device, comprising: Broadcast the update results generated by the first deployment device for the target parameters to at least one second deployment device; After receiving the update result, the second deployment device is controlled to broadcast the update result to the third deployment device, which is in the same deployment device group as the second deployment device, via group broadcast. The update result is broadcast to a fourth deployment device that is in the same deployment device group as the first deployment device, wherein the first deployment device, the second deployment device, the third deployment device, and the fourth deployment device are used to jointly deploy the target model in a distributed deployment manner.

2. The method according to claim 1, further comprising: The third deployment device is controlled to transmit the generated update sub-gradient for the target parameter to the second deployment device, which is in the same deployment device group as the third deployment device; The second deployment device is controlled to form an intra-group gradient of the target parameter based on the transmitted updated sub-gradient; The second deployment device is controlled to provide the intra-group gradient to the first deployment device; Based on the intra-group gradient, the global gradient of the target parameter is formed; Based on the global gradient, the updated result for the target parameters is generated.

3. The method according to claim 2, wherein, The control of the second deployment device to provide the intra-group gradient to the first deployment device includes: The second deployment device is controlled to provide the intra-group gradient to the first deployment device in response to the total volume of the data blocks of the intra-group gradient already formed on the second deployment device being greater than or equal to a volume threshold.

4. The method according to claim 2, wherein, The control of the second deployment device to provide the intra-group gradient to the first deployment device includes: The second deployment device is controlled to provide the intra-group gradient to the first deployment device in response to the number of data blocks of the intra-group gradient already formed on the second deployment device being greater than or equal to a number threshold.

5. The method according to claim 2, wherein, The control of the second deployment device to provide the intra-group gradient to the first deployment device includes: The second deployment device is controlled to periodically provide the group gradient to the first deployment device.

6. The method of claim 1, further comprising, in response to the second deployment device being predetermined: Read the first communication indicator between the first deployment device and each of the second deployment devices; In response to determining, based on the first communication indicator, that there is currently a blocked second deployment device in the second deployment device that is in a blocked communication state with the first deployment device, the second communication indicator between the first deployment device and each deployment device in the target deployment device group to which the blocked second deployment device is located is read; Based on the second communication metric, an updated second deployment device is determined from the deployment devices of the target deployment device group to replace the blocking second deployment device.

7. The method according to claim 1, wherein, In response to the second deployment device not being predetermined, broadcasting the updated result generated by the first deployment device for the target parameter to at least one second deployment device includes: For a deployment device group, based on the third communication index between each deployment device in the deployment device group and the first deployment device, a second deployment device corresponding to the first deployment device is determined in the deployment device group; Broadcast the updated results generated by the first deployment device for the target parameters to at least one of the second deployment devices.

8. The method of claim 1, wherein, in response to the fact that the number of deployment devices included in each of the deployment device groups is the same, the method further comprises: Using the same logical numbering rule, assign corresponding logical numbers to the deployment devices in each of the deployment device groups; Based on the first logical number of the first deployment device, the second deployment device is determined in each of the deployment device groups, wherein the second logical number assigned to the second deployment device in the deployment device group is the same as the first logical number.

9. The method according to any one of claims 1-8, wherein in the group of deployed devices, the communication rate between different deployed devices is greater than or equal to a communication rate threshold.

10. A communication device for model training parameters, applied to a first deployment device, comprising: The first broadcast unit is configured to broadcast the update result generated by the first deployment device for the target parameter to at least one second deployment device; The broadcast control unit is configured to control the second deployment device to broadcast the update result to a third deployment device that is in the same deployment device group as the second deployment device after receiving the update result. The second broadcast unit is configured to broadcast the update result to a fourth deployment device that is in the same deployment device group as the first deployment device, wherein the first deployment device, the second deployment device, the third deployment device, and the fourth deployment device are used to jointly deploy the target model in a distributed deployment manner.

11. An electronic device, comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the communication method for model training parameters according to any one of claims 1-9.

12. A non-transitory computer-readable storage medium storing computer instructions for causing the computer to perform a communication method for model training parameters according to any one of claims 1-9.

13. A computer program product comprising a computer program that, when executed by a processor, implements a communication method for model training parameters according to any one of claims 1-9.