172 results about "Supercomputer" patented technology

The invention relates to the technical field of computers, in particular to an operation method of a DeePMD-kit model in a Shenwei super computer, six core groups CG of the Shenwei super computer are respectively responsible for reasoning of six sub-regions of an MPI process, each core group CG comprises an MPE and 64 CPEs, and the MPE contains a 256-bit SIMD unit; each CPE contains a 512-bit SIMD (Single Instruction Multiple Data) unit; each CPE comprises a local data memory, and each local data memory LDM realizes continuous data transmission through a DMA strategy and a main memory; the CPEs in the same core group CG realize data transmission through remote memory access (RMA); the reasoning method of each core group CG comprises the following steps: accelerating a time-consuming operator through an SACA, and distributing a calculation-intensive part of the time-consuming operator to 64 CPEs by MPE; after calculation is completed, the MPE synchronizes calculation results of the 64 CPEs and stores the results in a main memory; when the SACA runs, the SIMD is parallel, and a 512-bit vectorization instruction is sent out through the CPE. By means of the technical scheme, the method has the advantage that the DeePMD-kit model can be applied to the Shenwei computer.

A method for utilizing hierarchical tokens (h-tokens) in large language models (LLMs) including receiving and processing an input prompt to generate tokens, identifying functional components derived from a subset of tokens, associating or generating h-tokens for at least one of the functional components by compressing the subset of tokens into an h-token for each functional component, associating or generating implementation characteristics including a defined function, an identified event, and / or an implementation method for each h-token, performing a retrieval augmented generation (RAG) processing function on each h-token, and generating a response from an output of the RAG processing function.

The invention provides a time-space attention-driven soil humidity prediction method in a domestic supercomputing environment, and the method comprises the steps: constructing a time-space feature dynamic fusion soil humidity prediction network model based on attention guidance; the attention-guided spatio-temporal feature dynamic fusion soil humidity prediction network model comprises a convolutional long-short term memory network basic framework and a spatio-temporal attention mechanism module. The introduction of the ConvLSTM network effectively integrates the time sequence and space information of the soil humidity, the potential space-time dependence in the data is fully utilized, the space-time attention mechanism enables the model to adaptively pay attention to important time periods and space regions by dynamically adjusting the weight between time steps, and the accuracy of the model is improved. The limitation of fixed feature selection in a traditional method is avoided, so that the prediction precision is improved. The model can better capture space-time dynamic feature information, further identifies the relationship between the influence factor and the soil humidity, and especially shows unique advantages in complex space-time feature processing and modeling of long-time sequence data.

The invention provides a parallel iteration solving method and system for an electromagnetic finite element equation set, and relates to the technical field of electromagnetic finite elements, and the method comprises the steps: constructing a to-be-solved electromagnetic finite element equation set # imgabs0 #; dividing the electromagnetic finite element equation set into a plurality of calculation sub-tasks, and distributing a sub-matrix block # imgabs 1 # and a sub-right-end item vector # imgabs 2 # of the sub-tasks to each process in the super computer; by means of the initialized Hoseholder-GMRES algorithm, the processes are subjected to subtask calculation in parallel, and the local optimal solution # imgabs3 # of each process is obtained; summarizing the local optimal solution # imgabs4 # of each process to obtain an optimal solution # imgabs5 # of the electromagnetic finite element equation set; according to the method, the electromagnetic finite element, the Householder-GMRES algorithm and parallel computing are combined, and the solving efficiency of the linear equation set of the electromagnetic finite element is improved by utilizing a parallel optimization technology.

A system for answering queries using one or more families of large language models (h-LLMs) including a processor, a communication device, and storage with software that, when executed by the processor, is operable to receive a user prompt, generate derived prompts from the user prompt, generate prompt embeddings from the derived prompts, transmit the plurality of prompt embeddings to a vector database, the vector database comprising a database of knowledge documents, each knowledge document comprised by the database of knowledge documents having one or more embeddings associated therewith, receive knowledge documents relevant to the prompt embeddings, generate context-aware prompts, transmit the context-aware prompts to the one or more h-LLMs, receive h-LLM results, process the h-LLM results to produce processed h-LLM results, each having a score, identify one or more preferred results responsive to the scores, and transmit the one or more preferred results to a user via the user interface.

A method for assigning tasks to LLMs using h-including receiving documents relevant to embeddings, generating context-aware prompts responsive to at least one of a received prompt, derived prompts, and the received knowledge documents, each context-aware prompt corresponding to a specialized task of the one or more specialized tasks, transmitting each context-aware prompt to a respective h- that is configured to specialize in processing prompts having a specialty corresponding to the specialized task of the context-aware prompt, and receiving results from the h-LLMs.

The invention provides an iterative solver optimization method and system based on a Shenwei architecture, relates to the technical field of electronic information, and aims to solve the problems that an AztecOO suitable for an SW26010Pro architecture in a new generation of Shenwei super computer at present still adopts a double-precision algorithm, the storage and calculation cost is high, and an AztecOO solver is low in calculation efficiency and low in calculation precision. The method comprises the following steps: acquiring a solving task of a sparse linear equation set, and issuing the solving task to a main core; the main core divides a solving task and allocates a memory for the solving task; the master core calls a slave core starting function and executes calculation; and the slave core feeds back a calculation result to the master core to obtain a solution of the sparse linear equation set. According to the method, the iterative solver is optimized through a designed precision selection strategy, a matrix segmentation algorithm and a mixed precision algorithm, so that the calculation speed of the algorithm is improved and memory resources are saved under the condition that sufficient calculation precision is ensured, the whole solving process of the solver is accelerated, and more efficient development efficiency is provided for researchers in the aspect of numerical calculation.

A method for generating a merged large language model (LLM) from input data including training a first LLM on a first subset of the input data for a first specialized task, training a second LLM on a second subset of the input data for a second specialized task, performing a merging process on the first and second LLMs producing one or more merged LLMs. The method further includes receiving a plurality of derived prompts generated from a user prompt and one or more relevant contexts, receiving one or more relevant contexts related to the generation of the plurality of derived prompts, providing the plurality of derived prompts generated and the one or more relevant contexts to the one or more merged LLMs, generating a plurality of results, and sending the plurality of results to an output broker.

The invention relates to a management method and device for a distributed collection system facing the supercomputing Internet and a storage medium, and the method comprises the steps: obtaining the heartbeat information of all collection nodes in the distributed collection system from a first database, and obtaining the management authority according to the heartbeat information of all collection nodes; and when it is determined that the management authority is obtained, an instance list corresponding to the distributed acquisition system is obtained from the second database, and acquisition tasks on all the acquisition nodes in the distributed acquisition system are managed according to the heartbeat information of all the acquisition nodes and the instance list. According to the method, in the process of managing the acquisition tasks in the distributed acquisition system by using the method, a decentration effect can be realized by adopting a mode of mutually monitoring all the acquisition nodes, and the management efficiency and stability of the whole system are improved. And each acquisition node performs task management in a manner of obtaining the management authority, so that management conflicts can be avoided.

The invention relates to an ultrasonic large model sparse tensor optimization training and pushing acceleration method based on new generation supercomputing, and the method comprises the steps: carrying out the fine tuning of an ultrasonic large model, enabling the weight tensor of each expert model in the large model to generate a structured block sparse mode, and obtaining an ultrasonic sparse large model; constructing a sparse block index table and an expert index table, and associating to form a joint coding table; deploying an ultrasonic sparse large model to each parallel group of the new-generation heterogeneous supercomputing cluster; performing multi-round iterative training on the large sparse model; in each round of iterative training, dividing the sample ultrasonic data of the current batch into each parallel group; in each parallel group, determining a target expert model needing to be activated based on the feature data of each sample ultrasonic data, and positioning a target computing node based on the joint coding table; and activating the target expert model on the target computing node and loading the corresponding non-zero weight block so as to compute the feature data, and updating the model weight parameter based on an output result obtained by computation. The method can accelerate the processing efficiency of the ultrasonic large model.

The supercomputing charging pile system provided by the embodiment of the present invention includes: a power distribution main station device, a fusion terminal device and a plurality of charging piles; each charging pile reports work information and computing power information to the power distribution main station through the fusion terminal device; the power distribution main station device obtains the flow calculation task issued by the supercomputing terminal, and at the same time, according to the work information and computing power information, decomposes the flow calculation task into a plurality of subtasks, and determines the work information corresponding to each subtask; wherein the subtask corresponds to the charging pile one by one; each subtask and the work information corresponding to each subtask are issued to each charging pile; the charging pile obtains the processing result according to each subtask and the work information corresponding to each subtask, and reports the processing result to the supercomputing terminal through the fusion terminal device and the power distribution main station device in sequence. The present invention can effectively alleviate the computing pressure of the supercomputing center by utilizing the distributed computing power of the charging piles and handing over a part of the flow tasks of the supercomputing to the distributed computing power.

A computing blade for a liquid-cooled supercomputer, including as many network cards as there are processors, each network card being associated with a processor. Each network card includes a plug-in end that receives a connector, and each network card allows data communication between the processor and another processor. Each network card is mounted to the cold plate being connected to the motherboard via a connection system that includes a first connector, connected to the plug-in end of the network card, and a second connector, connected on the one hand to the first connector via a cable and on the other hand to the motherboard.

Systems and methods for intelligent processing of requests in an artificial intelligence system, including receiving an incoming request at an AI broker, evaluating the incoming request to determine routing characteristics, selecting selected specialized language models from a plurality of specialized language models based on the routing characteristics, computing performance metrics for the one or more selected specialized language models, routing the incoming request to the selected specialized language models based on the performance metrics, and generating a final result by receiving and processing results from the selected specialized language models.

The invention relates to the technical field of super-computing clusters, and discloses a vnc resource allocation method and device in a super-computing cluster, and the method comprises the steps: a control node splices a job script to obtain a job script file, compares the computing resource of a computing node and the vnc resource data of a display node with a preset scheduling condition, and sends the job script file to the control node; if the vnc resource data and the computing resources meet the preset scheduling condition, sending the job script file to the computing node; the computing node executes the job script file and starts the vnc service; after the operation script file is executed, the computing node obtains and executes the user application program; and the control node calls the operation state in the computing node, sends a vnc service stop request to the display node based on the operation state, and controls the display node to stop the vnc service. According to the method and the device, the vnc resources in the supercomputing cluster are reasonably allocated, node collapse caused by unreasonable allocation of the vnc resources is solved, and the graphic display effect of the application program is improved.

A method for assigning tasks to LLMs using h-including receiving documents relevant to embeddings, generating context-aware prompts responsive to at least one of a received prompt, derived prompts, and the received knowledge documents, each context-aware prompt corresponding to a specialized task of the one or more specialized tasks, transmitting each context-aware prompt to a respective h- that is configured to specialize in processing prompts having a specialty corresponding to the specialized task of the context-aware prompt, and receiving results from the h-LLMs.

A method for assigning tasks to h-LLMs including receiving a received prompt, generating derived prompts from the user prompt, generating prompt embeddings from the derived prompts, transmitting the prompt embeddings to a vector database, receiving received knowledge documents from the vector database relevant to the prompt embeddings, generating context-aware tasks responsive to one of the user prompt, the derived prompts, and the received knowledge documents, transmitting each context-aware task an h-LLM configured to specialize in processing tasks having a specialty corresponding to the specialty of the context-aware task, and receiving produced results from the plurality of h-LLMs. At least one of the received prompt or the produced results is processed and transmitted to another h-LLM. New requests are created by the h-LLMs as part of processing the at least one or more of the received prompt and the plurality of produced results.

A system for answering queries using one or more families of large language models (h-LLMs) including a processor, a communication device, one or more h-LLMs implemented as microservices in cloud container environments and being accessible via a cloud service API, and software that is executable to operate an input broker having a broker API operable to receive a user prompt from a user interface, generate a plurality of derived prompts, transmit the plurality of derived prompts to one or more h-LLMs via the cloud service API, operate an output broker operable to receive a plurality of h-LLM results, process the plurality of LLM results at the output broker to generate a result, and transmit the result to the user interface via the broker API.

A method for generating a merged large language model (LLM) from input data including receiving derived prompts generated from a user prompt and relevant contexts, receiving the relevant contexts related to the generation of the plurality of derived prompts, providing derived prompts and the relevant contexts to one or more merged LLMs, generating a plurality of results, and sending the plurality of results to an output broker.

An extension for a supercomputer rack, the extension including a reinforcement frame, defining a rectangular through-opening, and at least one fluidic connection support mounted to the reinforcement frame. The at least one fluidic connection support includes an upstream circuit, configured to convey a refrigerant liquid to a plurality of connectors configured to be each connected via a cold flexible manifold to a cooling circuit of a compute blade of a server, and a downstream circuit, configured to collect the heated refrigerant liquid having passed through a compute blade of a server via a hot flexible manifold and a plurality of collection connectors. The extension is configured to be positioned at a rear end of the rack.

A system for answering queries using one or more families of large language models (h-LLMs) including a processor, a communication device, and storage with software that, when executed by the processor, is operable to receive a user prompt, generate derived prompts from the user prompt, generate prompt embeddings from the derived prompts, transmit the plurality of prompt embeddings to a vector database, the vector database comprising a database of knowledge documents, each knowledge document comprised by the database of knowledge documents having one or more embeddings associated therewith, receive knowledge documents relevant to the prompt embeddings, generate context-aware prompts, transmit the context-aware prompts to the one or more h-LLMs, receive h-LLM results, process the h-LLM results to produce processed h-LLM results, each having a score, identify one or more preferred results responsive to the scores, and transmit the one or more preferred results to a user via the user interface.

A method for training large language models (LLMs) and using them for inference including a base LLM, creating connected models while training the base LLM by coupling multiple trained base a parallel, series, or hybrid architecture, creating specialized connected language models for different specialized processing tasks by supplementally training two or more connected models different respective training sets, supplementally training a subset of connected language models by selectively routing one or more training data inputs to one or more specialized connected language models of the plurality of specialized connected language models responsive to at least one of accuracy optimization and task specialization, providing test prompt sets to the subset, evaluating an accuracy metric of the outputs from the subset, and routing prompts or derived prompts to the subset.

A system and method for creating a merged large language model (h-LLM) using a bagging approach including receiving input data at a computer system, creating a plurality of data subsets from the input data, training a plurality of h-LLMs, each h-LLM of the plurality of h-LLMs being trained on a respective data subset of the plurality of data subsets, creating a merged h-LLM by merging the plurality of h-LLMs, and outputting the merged h-LLM.

The invention comprises (i) a compilation method for automatically converting a single-threaded software program into an application-specific supercomputer, and (ii) the supercomputer system structure generated as a result of applying this method. The compilation method comprises: (a) Converting an arbitrary code fragment from the application into customized hardware whose execution is functionally equivalent to the software execution of the code fragment; and (b) Generating interfaces on the hardware and software parts of the application, which (i) Perform a software-to-hardware program state transfer at the entries of the code fragment; (ii) Perform a hardware-to-software program state transfer at the exits of the code fragment; and (iii) Maintain memory coherence between the software and hardware memories. If the resulting hardware design is large, it is divided into partitions such that each partition can fit into a single chip. Then, a single union chip is created which can realize any of the partitions.

A zettascale supercomputer may be configured using an array of servers organized in computer pods comprising 4-60 servers per pod. Each server includes computer modules immersed in a tank in which cooling water is flowing. Copper bus bars deliver power to each pod in a range of 4-180 MW. Cooling water is delivered to each pod in a range of 200-24,000 gallons per minute. Supercomputers having an operating power in a range of 4 MW-10 GW are described.

The invention discloses a multi-source cooperative green electric power energy optimization scheduling system, and relates to the technical field of electric power system computing power scheduling, and the system comprises an interaction construction module which constructs a direct connection signal channel of a supercomputing CPU and an intelligent computing GPU; the impedance matching module monitors a signal transmission reflection coefficient in real time and obtains signal transmission delay to realize impedance compensation; the supercomputing power supply layer configuration module regulates and controls output voltage in real time through a voltage monitoring unit; the intelligent calculation power supply layer configuration module adapts to the instantaneous peak current demand of the intelligent calculation module; the supercomputing water-cooling heat dissipation execution module is used for adjusting the flow of a water-cooling pump after receiving the temperature signal; the intelligent calculation air cooling linkage control module is used for collecting the real-time temperature of the intelligent calculation module and controlling the rotating speed of a speed-adjustable fan according to a fan rotating speed adjusting formula so as to realize pulse type heating adaptive heat dissipation; and the comprehensive evaluation module constructs a comprehensive evaluation model and evaluates the system performance. The problems of high transmission delay, mismatched heat dissipation and unstable power supply are solved.

The present invention provides a method for predicting supercomputer job duration based on semantics and timing, which relates to the field of supercomputers and solves the problem of ignoring semantic information in the job path and timing information between jobs in supercomputer job runtime prediction. The method first obtains job log data, distinguishes different user types through data grouping, uses different data storage methods to store the user's job log data, and forms a user model training set through coarse-grained clustering. Subsequently, a prediction model for job runtime is constructed and model training is performed by improving the BERT architecture. During the training process, a timing prediction method is combined so that the prediction model predicts the timing of new jobs based on timing information. After the training is completed, the user submits the job path information of the new job, and the prediction model determines the job category of the new job and outputs the job runtime. The present invention improves the prediction accuracy of job runtime, facilitating subsequent backfill scheduling.

The invention provides an internal interconnection and intercommunication small supercomputing system, device and cluster. The system comprises a programmable processing unit, a plurality of high-speed serial interconnection ports, a high-speed data transmission module and a multi-level heterogeneous storage module, the high-speed serial interconnection port is used for connecting the computing unit; the high-speed data transmission module is used for sending the received to-be-processed task to the cache space of the programmable processing unit; the programmable processing unit is used for distributing to-be-processed tasks to the plurality of computing units for task processing based on a preset distribution strategy and receiving task processing results; and the multi-level heterogeneous storage module is used for storing the received task processing result. The programmable processing unit serves as an architecture center, various high-speed interfaces are managed and coordinated in a unified mode, an architecture and a management mechanism for interconnection and intercommunication of internal components of the small supercomputing system are achieved, and the data transmission efficiency is improved. The programmable processing unit distributes tasks to a plurality of computing units for parallel processing, so that the computing efficiency and the computing power utilization rate are improved.

The application relates to the technical field of supercomputer clusters, and discloses a method and device for allocating vnc resources in a supercomputer cluster, which comprises the following steps: a control node splices a job script to obtain a job script file, and compares the vnc resource data of a display node and the computing resource of a computing node with preset scheduling conditions; if the vnc resource data and the computing resource meet the preset scheduling conditions, the job script file is sent to the computing node; the computing node executes the job script file to start a vnc service; after the execution of the job script file is completed, the computing node acquires and executes a user application program; the control node calls a job state in the computing node, sends a vnc service stop request to the display node based on the job state, and controls the display node to stop the vnc service. The application realizes the reasonable allocation of the vnc resources in the supercomputer cluster, solves the node collapse caused by the unreasonable allocation of the vnc resources, and improves the graphic display effect of the application program.

The application discloses a supercomputer load balancing strategy determination method and device based on a structure grid, equipment and a storage medium, relates to the technical field of computational fluid dynamics, and comprises the following steps: setting a processor of a supercomputer as a computing node, constructing a structure grid with parameters such as the number of grid units and the number of adjacent surfaces based on a master-slave core heterogeneous architecture, establishing a load balancing model in combination with a hardware architecture, utilizing the model to segment the structure grid under constraint conditions such as node calculation time, inter-node communication, inter-core group communication and inter-slave core communication, obtaining segmented grid blocks, determining vertices and their weights according to the segmentation result, determining edges and their weights based on an adjacent relationship, setting the number of adjacent grid blocks as the vertex degree, thereby constructing a directed graph mapping result, processing the mapping result by using a graph partitioning algorithm, and inversely mapping the partitioning result into a final target grid, and accordingly determining an optimal load balancing strategy, so that the efficiency of determining the load balancing strategy can be improved.

The invention provides a method and device for computing a target task through a super computer, and the method comprises the steps: obtaining the topological information of a computing network corresponding to the super computer, the computing network comprises N network nodes, and the N network nodes comprise M computing nodes and L switching nodes, the topology information comprises association between the network nodes; determining a problem domain which corresponds to the given target calculation task and is represented by a first multivariate array; according to the computing performance of each computing node, determining the computing power weight corresponding to each computing node, and according to the topological information and the computing power weight of each computing node, dividing the problem domain into M sub-domains, the M sub-domains respectively corresponding to the M computing nodes; and obtaining a calculation result of the sub-domain corresponding to each calculation node through each calculation node, and determining a calculation result of the target calculation task according to the calculation result of each sub-domain.