A computing power routing method for an AGI reasoning network

Abstract

The application provides a computing power routing method for AGI reasoning network, relates to the field of AGI computing power routing, and comprises the following steps: collecting physical operation state parameters of different types of nodes in the AGI reasoning network, constructing a hot area distribution, and establishing an association with a reasoning stage; outputting a physical bearing state set and a hot drift association result; statistically analyzing resource occupation characteristics of the reasoning stage and matching the resource occupation characteristics with the hot area distribution; reconstructing physical pressure and generating a bearing decay mark; outputting a physical constraint reasoning set and a bearing change result; identifying a reasoning stage to be released, matching a migration target node, completing context migration, and generating a reasoning release constraint; outputting a physical constraint routing result and a reasoning stage release mapping set; determining whether the node is out of a hot accumulation state and restores bearing capacity; executing a reasoning stage backflow and updating a physical constraint routing; and outputting an AGI reasoning computing power routing result, so that the AGI reasoning efficiency and stability are improved, and the computing power resource utilization rate is significantly improved.

Description

Technical Field

[0001] This invention belongs to the field of AGI computing power routing, and more specifically, relates to a computing power routing method for AGI inference networks. Background Technology

[0002] With the rapid development of AGI technology, the demand for high-density, long-duration computing power for long-chain inference tasks is growing exponentially, and the hybrid deployment of edge data centers and liquid-cooled computing centers has become the mainstream architecture.

[0003] Traditional computing power routing methods rely solely on logical load, queue depth, and static capacity for scheduling, failing to recognize the shifts in physical capacity of computing nodes during continuous inference. In practice, when liquid-cooled nodes perform high-density context inference, local heat dissipation loops are prone to heat buildup, leading to decreased chip heat exchange efficiency. While air-cooled nodes can recover quickly through momentary load release, traditional methods cannot detect this dynamic change. This results in a significant reduction in the actual inference throughput of some nodes with normal logical loads, yet the system continues to allocate long-chain inference tasks to them, ultimately causing physical inference queuing and congestion. Furthermore, existing methods do not differentiate the heat dissipation characteristics of nodes with different cooling types, nor do they establish a correlation between the inference stage and node hot zones, making dynamic path adjustment impossible. This leads to overall low inference efficiency and unbalanced resource utilization, making it difficult to meet the stability and real-time requirements of large-scale AGI long-chain inference.

[0004] Therefore, it is necessary to provide a computational power routing method for AGI inference networks to solve the above-mentioned technical problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to overcome the aforementioned deficiencies and propose a computing power routing method for AGI inference networks.

[0006] The present invention adopts the following technical solution.

[0007] This invention discloses a computational power routing method for AGI inference networks, comprising: Collect and analyze the physical operating state parameters of different types of nodes in the AGI inference network, construct the continuous hot zone distribution and establish the correlation between the inference stage and the continuous hot zone distribution, and output the physical bearing state set and thermal drift correlation results. Based on the physical load-bearing state set and thermal drift correlation results, the resource occupancy characteristics in the statistical inference stage are correlated and matched with the distribution of continuous hot zones. The physical pressure in the inference stage is reconstructed and a load-bearing attenuation marker is generated. The physical constraint inference set and continuous load-bearing change results are output. Based on the physical constraint inference set and the continuous bearing change results, identify the inference stage to be released and match the migration target node, complete the continuous context migration and generate hierarchical inference release constraints, and output the target physical constraint routing result and the inference stage release mapping set; Based on the target physical constraint routing results and the inference phase release mapping set, monitor the heat dissipation recovery status of the nodes, determine whether the nodes have escaped the heat accumulation state and restore the carrying capacity of the nodes in stages, execute the progressive backflow of the inference phase and dynamically update the physical constraint routing, and output the AGI inference computing power routing results with dynamic physical load balancing.

[0008] Preferably, the step of collecting and analyzing the physical operating state parameters of different types of nodes in the AGI inference network, constructing a persistent hot zone distribution and establishing a correlation between the inference stage and the persistent hot zone distribution, and outputting the physical load-bearing state set and thermal drift correlation results includes: The physical operating status parameters of edge data center nodes, liquid-cooled nodes, and air-cooled nodes in the AGI inference network are collected uniformly, preprocessed, and output physical operating status records. Based on the physical operation status record, the instantaneous heat dissipation capacity reflecting the current heat dissipation level of each node is calculated according to the node cooling method, and the instantaneous heat dissipation capacity result is output. Based on the instantaneous heat dissipation capability results, combined with the actual power consumption and continuous inference duration, the continuous heat accumulation results of each node during the continuous inference process are calculated. Based on the continuous heat accumulation results, each node is divided into multiple functional areas and heat accumulation is allocated to construct a continuous heat zone distribution. Then, the continuous heat zone is mapped to the current inference stage for occupancy relationship, and the physical bearing state set and heat drift correlation results are output.

[0009] Preferably, the unified collection of physical operating status parameters of edge data center nodes, liquid-cooled nodes, and air-cooled nodes in the AGI inference network, followed by preprocessing, outputs physical operating status records, including: The data collection cycle is dynamically adjusted according to the load intensity of the AGI inference task. During the peak period of long-chain inference tasks, a short data collection cycle is adopted to simultaneously collect the physical operating status parameters of edge data center nodes, liquid-cooled nodes and air-cooled nodes. The physical operation status parameters of each node are preprocessed to remove outliers that are clearly outside the reasonable range. The missing values ​​are filled in by interpolation between adjacent time points, and the timestamp format of all physical operation status parameters is standardized. The preprocessed physical operation status parameters are classified and stored according to node type and physical operation status parameter category, generating a physical operation status record containing node type identifier, acquisition time, and physical operation status parameters.

[0010] Preferably, the step of calculating the instantaneous heat dissipation capacity reflecting the current heat dissipation level of each node based on the physical operating status record and the node cooling method, and outputting the instantaneous heat dissipation capacity result, includes: Based on the node type identifier in the physical operation status record, automatically select the heat dissipation capacity calculation logic that matches the node cooling method corresponding to the node type identifier; Physical operating status parameters are extracted from the physical operating status records as key heat dissipation capacity calculation parameters, including at least temperature, flow rate, and rotation speed. The basic heat dissipation capacity of each node is calculated and corrected based on the heat dissipation capacity calculation logic to obtain the instantaneous heat dissipation capacity that reflects the current heat dissipation level of each node. The instantaneous heat dissipation capacity of each node is compared with its rated heat dissipation capacity. The heat dissipation margin is calculated and the heat dissipation capacity level is marked. The instantaneous heat dissipation capacity result, which includes node type identifier, instantaneous heat dissipation capacity and heat dissipation margin, is generated.

[0011] Preferably, the calculation of the continuous heat accumulation result of each node during the continuous inference process based on the instantaneous heat dissipation capability result, combined with the actual power consumption and continuous inference duration, includes: Extract the current actual power consumption of each node from the physical operation status record, and obtain the continuous inference time of each node from the inference scheduling system; The actual power consumption is used as the heat input value per unit time, and the instantaneous heat dissipation capacity is used as the heat output value per unit time. The net heat accumulation of each node per unit time is calculated. By introducing a heat retention coefficient that reflects the degree of heat retention within each node, and combining the duration of continuous heat inference and the net heat accumulation, the total continuous heat accumulation of each node during the continuous inference process is calculated. The total continuous heat accumulation is compared with the preset multi-level heat accumulation threshold to determine the heat accumulation risk level of each node, and a continuous heat accumulation result containing node type identifier, total continuous heat accumulation, and heat accumulation risk level is generated.

[0012] Preferably, the step of statistically analyzing resource occupancy characteristics during the inference phase based on the physical load-bearing state set and thermal drift correlation results, matching these characteristics with the distribution of persistent hot zones, reconstructing the physical pressure during the inference phase, generating load-bearing attenuation markers, and outputting the physical constraint inference set and persistent load-bearing change results includes: Based on the physical load state set and thermal drift correlation results, the continuous execution characteristics of each inference stage at the corresponding node are statistically analyzed, key indicators reflecting the resource consumption intensity of the inference stage are extracted, and the resource consumption characteristic set of the inference stage is output. Based on the resource occupancy feature set of the inference stage, the resource occupancy features of the inference stage are correlated and matched with the distribution of persistent hot zones. The driving effect of the inference stage on the formation and diffusion of persistent hot zones is analyzed, and the hot zone matching results of the inference stage are output. Based on the hot zone matching results during the inference phase, and combined with the hot zone correlation strength, memory usage ratio, continuous dwell time, and heat dissipation pressure ratio, physical pressure results reflecting the physical load pressure on nodes during the inference phase are reconstructed. Based on the physical pressure results, and according to the continuous changes in physical pressure and the changing trend of node inference throughput, different levels of load attenuation markers and corresponding physical constraint relationships are generated, and the physical constraint inference set and continuous load change results are output.

[0013] Preferably, the physical pressure result, which reflects the physical load pressure on the node during the inference phase, is reconstructed based on the hot zone matching results during the inference phase, combined with the hot zone correlation strength, memory usage ratio, continuous dwell time, and heat dissipation pressure ratio. This includes: The dimensions of physical stress affecting the inference stage are obtained, including the thermal zone correlation strength, memory usage ratio, continuous dwell time, and heat dissipation pressure ratio. Based on the node cooling method and node operating environment, differentiated dimension weight coefficients are assigned to each dimension, and the physical pressure components corresponding to each inference stage in each dimension are calculated respectively, and each physical pressure component is processed as dimensionless. The physical pressure components of each dimension are weighted and summed according to the dimension weight coefficients to obtain the physical pressure of each inference stage, and a physical pressure result containing the inference stage identifier and the corresponding physical pressure is generated.

[0014] Preferably, the step of identifying the inference stage to be released and matching the migration target node based on the physical constraint inference set and the continuous bearing change results, completing the context continuous migration and generating hierarchical inference release constraints, and outputting the target physical constraint routing results and the inference stage release mapping set includes: Based on the physical constraint inference set and the continuous load change results, according to the load attenuation level and load change trend of the node, the inference stages that need to be released first in the hot drift node are identified, and the set of inference stages to be released is output. Based on the set of inference stages to be released, candidate target nodes that meet the migration conditions are screened, the remaining physical carrying capacity ratio and hot zone overlap risk of the candidate target nodes are evaluated, the migration target nodes for each inference stage to be migrated are determined, and the migration target node matching results are output. Based on the target node matching results, the context dwelling state in the inference stage of the migration is classified and processed. The state migration is completed by copying and verifying first and then switching, and the continuous context migration results are output. Based on the contextual continuous migration results, hierarchical inference release constraints are generated for the source heat accumulation nodes corresponding to the inference stage to be migrated. The original routing path is corrected and the release mapping relationship between the inference stage and the nodes is established. The target physical constraint routing results and the inference stage release mapping set are output.

[0015] Preferably, the step of screening candidate target nodes that meet the migration conditions based on the set of inference stages to be released, evaluating the remaining physical carrying capacity ratio and hot zone overlap risk of the candidate target nodes, determining the migration target nodes for each inference stage to be migrated, and outputting the migration target node matching results includes: Nodes without severe continuous inference decay are selected from all available nodes and used as candidate target nodes for the inference phase to be released. Based on the instantaneous heat dissipation capacity, rated heat dissipation capacity, and load attenuation intensity of the candidate target nodes, calculate the proportion of remaining physical load that each candidate target node can use to receive the new inference phase. Analyze the degree of overlap between the key resource-occupied area in the migration inference stage and the persistent hot zone location of the candidate target node, and assess the risk of hot zone overlap that will cause heat accumulation again after migration. Taking into account the remaining physical carrying capacity ratio and hot zone overlap risk of the candidate target nodes, the migration fit value between each candidate target node and the migration inference stage is calculated. The candidate target node with the highest migration fit value is selected as the migration target node, and the migration target node matching result is output.

[0016] Preferably, the step of monitoring the heat dissipation recovery status of nodes based on the target physical constraint routing results and the inference phase release mapping set, determining whether nodes have escaped the heat accumulation state and restoring the carrying capacity of nodes in stages, executing the progressive backflow of the inference phase and dynamically updating the physical constraint routing, and outputting the AGI inference computing power routing results with dynamic physical carrying capacity balancing includes: Based on the release mapping set during the inference phase, determine the nodes that need to be monitored after release, continuously monitor the changes in the instantaneous heat dissipation capacity, continuous heat zone intensity and inference throughput of the nodes, and output the heat dissipation recovery status. Based on the heat dissipation recovery status, combined with the degree of heat dissipation recovery, the degree of hot zone reduction, the increase in inference throughput, and the decrease in attenuation intensity, it is determined whether the node has escaped the state of heat accumulation. For nodes that have broken away from the thermal accumulation state, their bearing relationship with the inference stage is reconstructed in stages according to the degree of recovery, and the physical restriction relationship is gradually removed to output the bearing relationship of the restored node stage. Based on the bearing relationship of the recovery node stage, the progressive backflow allocation of the inference stage is executed in order of increasing physical pressure, and the physical constraint route is dynamically updated according to the real-time running status of the node, outputting the AGI inference computing power routing result with dynamic physical bearing balance.

[0017] The beneficial effects of the present invention are as follows: Compared with the prior art, the present invention has the following advantages: (1) This invention collects and analyzes the physical operating state parameters of different types of nodes in the AGI inference network, constructs a continuous hot zone distribution, establishes the correlation between the inference stage and the continuous hot zone distribution, and outputs the physical load state set and thermal drift correlation results; based on the physical load state set and thermal drift correlation results, it statistically analyzes the resource occupancy characteristics of the inference stage and correlates them with the continuous hot zone distribution, reconstructs the physical pressure of the inference stage and generates a load attenuation marker, and outputs the physical constraint inference set and continuous load change results; based on the physical constraint inference set and continuous load change results, it identifies the inference stage to be released and matches the migration target node to complete the context. The system continuously migrates and generates hierarchical inference release constraints, outputting the target physical constraint routing results and the inference phase release mapping set. Based on the target physical constraint routing results and the inference phase release mapping set, it monitors the heat dissipation recovery status of nodes, determines whether nodes have escaped the heat accumulation state, and restores the carrying capacity of nodes in stages. It executes the progressive backflow of the inference phase and dynamically updates the physical constraint routing, outputting the AGI inference computing power routing results with dynamic physical load balancing. This solves the inference blocking problem caused by the drift of computing power nodes' physical load under hybrid deployment, improves AGI inference efficiency and stability, realizes closed-loop scheduling of computing power, and significantly improves the utilization rate of computing power resources.

[0018] (2) This invention effectively solves the inference blocking problem caused by the dynamic drift of the physical carrying capacity of computing nodes in hybrid deployment scenarios, and significantly improves the operating efficiency and stability of the AGI inference network. By collecting physical states in multiple dimensions and continuously mapping hot zones, the drift of node carrying capacity is accurately identified, solving the technical problems of traditional logical routing and realizing the coordinated scheduling of the physical layer and the logical layer. By reconstructing physical pressure and marking hierarchical decay in the inference stage, high-risk stages are restricted from entering hot drift nodes, reducing the probability of hot accumulation from the source. By using the context continuous inheritance migration mechanism, task migration is completed under the premise of ensuring the integrity of the inference chain, avoiding inference interruption and data loss. By using dynamic recovery monitoring and progressive computing power return, closed-loop adaptive adjustment of computing power resources is realized, which greatly improves the overall inference throughput of the system, reduces the inference latency, improves the utilization rate of computing power resources, and effectively extends the service life of computing nodes, and can stably support the continuous operation of large-scale AGI long-chain inference tasks. Attached Figure Description

[0019] Figure 1 This is a flowchart of a computing power routing method for AGI inference networks provided in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention.

[0021] like Figure 1 As shown in the embodiments of the present invention, the computing power routing method for AGI inference networks includes the following steps: node physical load status acquisition and continuous inference hot zone mapping processing, physical load drift analysis and inference stage pressure reconstruction processing, physical constraint route generation and inference migration release processing, and dynamic physical recovery monitoring and computing power backflow adjustment processing. By combining the node physical load evolution status for dynamic inference path adjustment, the method solves the inference blocking problem caused by the dynamic drift of node physical load capacity in the hybrid deployment scenario of edge data centers and liquid-cooled computing power centers, realizes the collaborative scheduling of the physical layer and the logic layer, and improves the operating efficiency and stability of AGI inference networks. The following is a further detailed description of each step of the present invention.

[0022] S1 collects and analyzes the physical operating state parameters of different types of nodes in the AGI inference network, constructs the continuous hot zone distribution and establishes the correlation between the inference stage and the continuous hot zone distribution, and outputs the physical bearing state set and thermal drift correlation results.

[0023] The physical operational status of edge data center nodes, liquid-cooled nodes, and air-cooled nodes in the AGI inference network is collected. Correlation analysis is performed on the heat dissipation cycle status, local heat accumulation locations, and continuous inference duration of each node during continuous inference. Based on the heat accumulation trend of different nodes during long-chain inference, a distribution of persistent hot zones is constructed. The occupancy relationship between persistent hot zones and corresponding inference stages is mapped to identify the degree of impact of different inference stages on the physical carrying capacity of nodes. A set of node physical carrying capacity statuses is generated based on the distribution of persistent hot zones, and the node thermal drift correlation results are output.

[0024] The process involves collecting and analyzing the physical operating state parameters of different types of nodes in the AGI inference network, constructing a persistent hot zone distribution, establishing the correlation between the inference stage and the persistent hot zone distribution, and outputting a set of physical load-bearing states and thermal drift correlation results, including: The physical operating status parameters of edge data center nodes, liquid-cooled nodes, and air-cooled nodes in the AGI inference network are collected uniformly, preprocessed, and output as physical operating status records.

[0025] The unified collection of physical operating status parameters of edge data center nodes, liquid-cooled nodes, and air-cooled nodes in the AGI inference network is performed and preprocessed to output physical operating status records, including: The data acquisition cycle is dynamically adjusted based on the load intensity of the AGI inference task. During peak periods of long-chain inference tasks, a short data acquisition cycle is used to simultaneously acquire the physical operating status parameters of edge data center nodes, liquid-cooled nodes, and air-cooled nodes.

[0026] The physical operating status parameters of each node are preprocessed to remove outliers that are clearly outside the reasonable range. Missing values ​​are filled in using the adjacent time-time interpolation method, and the timestamp format of all physical operating status parameters is standardized.

[0027] The preprocessed physical operation status parameters are classified and stored according to node type and physical operation status parameter category, generating a physical operation status record containing node type identifier, acquisition time, and physical operation status parameters.

[0028] For each node, the system collects data on chip temperature, inlet air temperature, outlet air temperature, cold plate inlet temperature, cold plate outlet temperature, coolant flow rate, fan speed, node power consumption, inference time, and inference throughput per unit time. Different key data collection parameters are set for different node types: for edge data center nodes, the focus is on ambient temperature, cabinet inlet / outlet air temperature difference, and node power consumption; for liquid-cooled nodes, the focus is on coolant inlet / outlet temperature difference, coolant flow rate, and cold plate temperature; for air-cooled nodes, the focus is on fan speed, inlet / outlet air temperature difference, and chip surface temperature.

[0029] The data acquisition period can be set from 1 to 10 seconds, with 1 to 3 seconds being preferred during peak periods of long-chain inference tasks to ensure the real-time nature of the physical state data. The acquired raw data is preprocessed to remove outliers that are significantly outside the reasonable range. Missing data is filled in using adjacent time-time interpolation, and the timestamp format of all data is standardized to the UTC standard format. This generates a standardized record of the node's physical operating status, including node identifier, acquisition time, and various physical parameter values.

[0030] By collecting physical parameters of nodes with different cooling types in a differentiated manner, the problem of traditional methods only collecting logical load data is solved.

[0031] Based on the physical operation status records, the instantaneous heat dissipation capacity reflecting the current heat dissipation level of each node is calculated according to the node cooling method, and the instantaneous heat dissipation capacity result is output.

[0032] The instantaneous heat dissipation capacity, based on the physical operating status record and according to the node cooling method, is calculated to reflect the current heat dissipation level of each node, and the instantaneous heat dissipation capacity result is output, including: Based on the node type identifier in the physical operation status record, automatically select the heat dissipation capacity calculation logic that matches the node cooling method corresponding to the node type identifier; Physical operating status parameters are extracted from the physical operating status records as key heat dissipation capacity calculation parameters, including at least temperature, flow rate, and rotation speed. The basic heat dissipation capacity of each node is calculated and corrected based on the heat dissipation capacity calculation logic to obtain the instantaneous heat dissipation capacity that reflects the current heat dissipation level of each node.

[0033] The instantaneous heat dissipation capacity of each node is compared with its rated heat dissipation capacity. The heat dissipation margin is calculated and the heat dissipation capacity level is marked. The instantaneous heat dissipation capacity result, which includes node type identifier, instantaneous heat dissipation capacity and heat dissipation margin, is generated.

[0034] Based on the node type identifier in the node's physical operation status record, the system automatically selects the heat dissipation capacity calculation logic that matches the node's cooling method. For liquid-cooled nodes, the system calculates the amount of heat that the node can effectively remove based on the collected coolant mass flow rate, coolant specific heat capacity, coolant inlet and outlet temperature difference, and liquid cooling loop heat transfer correction coefficient. For air-cooled nodes, the system calculates the amount of heat that the node can effectively remove based on the collected air mass flow rate, air specific heat capacity, inlet and outlet air temperature difference, and air duct effective heat transfer coefficient. For edge data center nodes, a rack environment correction item is overlaid on the air-cooled heat dissipation capacity calculation results to reflect the impact of hot air recirculation inside the rack on the heat dissipation capacity.

[0035] The calculated instantaneous heat dissipation capacity is compared with the node's rated heat dissipation capacity to calculate the heat dissipation margin and label the heat dissipation capacity level, generating a node instantaneous heat dissipation capacity result that includes the node identifier, instantaneous heat dissipation capacity value, and heat dissipation margin. The heat transfer correction factor for the liquid cooling loop is typically taken as 0.7 to 0.95, and the effective heat transfer coefficient for the air duct is typically taken as 0.5 to 0.85.

[0036] Based on the instantaneous heat dissipation capability results, combined with the actual power consumption and continuous inference duration, the continuous heat accumulation results of each node during the continuous inference process are calculated.

[0037] The calculation of continuous heat accumulation for each node during continuous inference, based on instantaneous heat dissipation capability results, actual power consumption, and continuous inference duration, includes: Extract the current actual power consumption of each node from the physical operation status record, and obtain the continuous inference time occupied by each node from the inference scheduling system.

[0038] The actual power consumption is used as the heat input value per unit time, and the instantaneous heat dissipation capacity is used as the heat output value per unit time. The net heat accumulation of each node per unit time is calculated.

[0039] By introducing a heat retention coefficient that reflects the degree of heat retention within each node, and combining the duration of continuous heat inference and the net heat accumulation, the total continuous heat accumulation of each node during the continuous inference process is calculated.

[0040] The total continuous heat accumulation is compared with the preset multi-level heat accumulation threshold to determine the heat accumulation risk level of each node, and a continuous heat accumulation result containing node type identifier, total continuous heat accumulation, and heat accumulation risk level is generated.

[0041] Extract the node's current actual power consumption data from the node's physical operating status record, and obtain the node's continuous inference occupancy time data from the inference scheduling system. Using the node's actual power consumption as the heat input value per unit time and the node's instantaneous heat dissipation capacity as the heat output value per unit time, calculate the node's net heat accumulation per unit time.

[0042] A heat retention coefficient, reflecting the degree of heat retention within a node, is introduced. Combined with the continuous inference duration, the total continuous heat accumulation of the node during the entire continuous inference process is calculated. This total continuous heat accumulation is compared with preset multi-level heat accumulation thresholds to determine the node's heat accumulation risk level, generating a node continuous heat accumulation result that includes the node identifier, total continuous heat accumulation, and heat accumulation risk level. The heat retention coefficient is typically set between 0.6 and 1.3, with higher values ​​used for enclosed cabinets with severe heat accumulation.

[0043] By calculating the heat accumulation during continuous inference, the risk of heat buildup at nodes can be predicted in advance, rather than waiting until the node temperature exceeds the limit before taking action, thus achieving early warning of thermal drift.

[0044] Based on the continuous heat accumulation results, each node is divided into multiple functional areas and heat accumulation is allocated to construct a continuous heat zone distribution. Then, the continuous heat zone is mapped to the current inference stage for occupancy relationship, and the physical bearing state set and heat drift correlation results are output.

[0045] For example, the internal structure of a node is divided into functional areas: chip area, memory area, power supply area, cold plate contact area, and airflow outlet area. Based on the thermal conductivity characteristics of different areas within the node, a corresponding heat distribution coefficient is assigned to each area, distributing the total continuous heat accumulation of the node to each functional area. Combining the current temperature of each area, the node's average temperature, and the temperature stability correction value, the sustained heat intensity of each area is calculated.

[0046] Different hot zone determination criteria are set for different types of nodes: For liquid-cooled nodes, the focus is on whether a hot zone forms between the cold plate contact area and the coolant outlet side; for air-cooled nodes, the focus is on whether a hot zone forms between the chip area and the end of the air duct; for edge data center nodes, the focus is on whether a hot zone forms in the upper hot air recirculation area of ​​the rack. Hot zone determination considers not only the temperature of the area but also the duration of continuous inference and the rate of heat accumulation for a comprehensive assessment, generating a continuous hot zone distribution result for the node that includes node identifiers, the intensity of hot zones in each area, and the location of the hot zones. The heat distribution coefficient is typically set between 0.1 and 0.6, determined by the node structure and sensor location; the temperature stability correction value is typically set between 273 Kelvin and 310 Kelvin to avoid misjudgment due to amplified low-temperature areas.

[0047] It should be noted that this step is not simply to determine whether a node is generating heat, but to accurately locate which physical region the continuous heat accumulation is occurring in, thereby providing a basis for mapping the relationship between the heat zone and the subsequent inference stage, and realizing the fine-grained location of thermal drift.

[0048] The results of the persistent hot zone distribution on a node are correlated with the inference phase currently being executed on that node, recording the chip area, video memory area, cache resident area, and duration of occupation for each inference phase. If an inference phase occupies a high hot zone on the same node for an extended period, and the inference throughput per unit time of that node decreases as the intensity of the hot zone increases, then that inference phase is marked as a thermal drift correlated phase.

[0049] The set of node physical load-bearing states includes at least the node type, effective heat dissipation capacity, continuous heat accumulation, sustained hot zone intensity, current inference stage, stage occupancy duration, and load-bearing attenuation state. The node thermal drift correlation results include at least the hot zone location, the correlated inference stage, the thermal drift intensity, and the type of inference stage that needs to be restricted from subsequent entry. The thermal drift intensity is calculated based on the sustained hot zone intensity, the duration of the hot zone occupied during the inference stage, the node's effective heat dissipation capacity, and the evaluation time window. For example, below 0.3 can be defined as slight thermal drift, 0.3 to 0.7 as moderate thermal drift, and above 0.7 as significant thermal drift. The evaluation time window is typically between 60 and 600 seconds.

[0050] S2, based on the physical load-bearing state set and thermal drift correlation results, statistically analyzes the resource occupancy characteristics in the inference stage and correlates them with the distribution of continuous hot zones, reconstructs the physical pressure in the inference stage and generates load-bearing attenuation markers, and outputs the physical constraint inference set and continuous load-bearing change results.

[0051] Based on the node physical load state set and node thermal drift correlation results, a physical load adaptation analysis is performed on the inference stages in the AGI inference network. Specifically, the continuous occupancy duration, context dwell depth, and inference resource release interval of each inference stage during continuous execution are statistically analyzed. Different inference stages are correlated and matched with the continuous hot zones of nodes to identify inference stages that may cause thermal accumulation and diffusion. The physical load consumption state of the inference stage in different nodes is reconstructed to generate stage physical pressure results. Continuous inference decay markers are generated for nodes with thermal drift risk, and physical constraint relationships are established for the corresponding inference stages. A physically constrained inference set is generated based on the stage physical pressure results and continuous inference decay markers, and the node continuous load change results are output.

[0052] The method, based on the physical load-bearing state set and thermal drift correlation results, statistically analyzes the resource occupancy characteristics during the inference phase and correlates them with the distribution of persistent hot zones. It then reconstructs the physical pressure during the inference phase, generates load-bearing attenuation markers, and outputs the physical constraint inference set and persistent load-bearing change results, including: Based on the physical load state set and thermal drift correlation results, the continuous execution characteristics of each inference stage at the corresponding node are statistically analyzed, key indicators reflecting the resource consumption intensity of the inference stage are extracted, and the resource consumption characteristic set of the inference stage is output.

[0053] The continuous execution process of each inference stage on the corresponding node is statistically analyzed, extracting at least the stage's duration of continuous occupation, context residency depth, GPU memory residency ratio, inference resource release interval, number of repeated stage entries, and the corresponding thermal drift intensity. Specifically, duration of continuous occupation indicates the length of time the inference stage continuously occupies the same node; context residency depth indicates the number of context levels retained in GPU memory or cache for the stage; and inference resource release interval indicates the time distance between two resource releases within the same inference stage.

[0054] If an inference phase lasts for a long time, has a high context dwell depth, and a short release interval, it indicates that the phase is likely to keep the node in a state where heat is difficult to release for a long time.

[0055] By statistically analyzing the continuous occupancy characteristics of the inference phase, we can quantify the degree of consumption of node physical capacity at different inference phases.

[0056] Based on the resource occupancy feature set of the inference stage, the resource occupancy features of the inference stage are correlated and matched with the distribution of persistent hot zones. The driving effect of the inference stage on the formation and diffusion of persistent hot zones is analyzed, and the hot zone matching results of the inference stage are output.

[0057] The feature set occupied during the inference phase is matched with the node thermal drift correlation results to determine whether each inference phase continuously affects the node's hot zone. Different matching priorities are set for different types of nodes: for liquid-cooled nodes, the focus is on whether the inference phase occupies the cold plate outlet side, the high-power chip area, or the memory stack area for a long time; for air-cooled nodes, the focus is on whether the inference phase is concentrated in the heat recirculation area at the end of the air duct; for edge data center nodes, the focus is on whether the inference phase is located in the hot air recirculation area at the top of the rack.

[0058] During matching, it's not just about determining whether a stage is running at a high-temperature node, but also whether there's a stable correspondence between its location, duration, heat zone intensity, and decrease in heat dissipation capacity. The matching strength between each inference stage and its corresponding heat zone is calculated, generating stage heat zone matching results that include the inference stage identifier, heat zone location, and matching strength.

[0059] Based on the hot zone matching results during the inference phase, and combined with the hot zone correlation strength, memory usage ratio, continuous dwell time, and heat dissipation pressure ratio, the physical pressure results reflecting the physical load pressure on the nodes during the inference phase are reconstructed.

[0060] The hot zone matching results based on the inference phase, combined with the hot zone correlation strength, memory usage ratio, continuous dwell time, and heat dissipation pressure ratio, reconstruct the physical pressure results reflecting the physical load pressure on the nodes during the inference phase, including: The dimensions of physical stress affecting the inference stage are obtained, including the thermal zone correlation strength, memory usage ratio, continuous dwell time, and heat dissipation pressure ratio.

[0061] Based on the node cooling method and node operating environment, differentiated dimension weight coefficients are assigned to each dimension, and the physical pressure components corresponding to each inference stage in each dimension are calculated respectively, and each physical pressure component is processed as dimensionless.

[0062] The physical pressure components of each dimension are weighted and summed according to the dimension weight coefficients to obtain the physical pressure of each inference stage, and a physical pressure result containing the inference stage identifier and the corresponding physical pressure is generated.

[0063] Based on the phase hot zone matching results, the impact of heat accumulation, long-term memory residency, heat dissipation capacity occupation, and insufficient resource release on nodes during the inference phase are comprehensively reconstructed to form the phase physical stress result. This result is not equivalent to logical load, but rather represents the degree of pressure exerted on the node's continuous carrying capacity at the physical level during a certain inference phase.

[0064] Four dimensions influencing the physical stress during the inference phase are identified: thermal zone correlation strength, memory usage ratio, continuous dwell time, and heat dissipation pressure ratio. Differentiated weighting coefficients are assigned to these four dimensions based on the node's cooling method and operating environment, ensuring the physical stress calculation results better reflect the actual operating characteristics of different node types. The stress components for each inference phase across the four dimensions are calculated separately, with each component being dimensionless to eliminate dimensional differences. The stress components across the four dimensions are then weighted and summed according to the pre-assigned weighting coefficients to obtain the comprehensive physical stress value for each inference phase, generating a phase physical stress result that includes the inference phase identifier and the corresponding physical stress value.

[0065] The weight coefficients for the four dimensions range from 0 to 1, and the sum of the four is 1. Liquid-cooled nodes can increase the weight of the thermal zone correlation strength dimension and the heat dissipation pressure ratio dimension; air-cooled nodes can increase the weight of the thermal zone correlation strength dimension and the continuous dwell time dimension; and edge data center nodes can increase the weight of the continuous dwell time dimension and the heat dissipation pressure ratio dimension.

[0066] By reconstructing the physical pressure during the inference phase, the abstract physical load consumption is transformed into a quantifiable indicator, providing a unified judgment standard for subsequent routing decisions and solving the problem that traditional methods cannot quantify the physical pressure during the inference phase.

[0067] Based on the physical pressure results, and according to the continuous changes in physical pressure and the changing trend of node inference throughput, different levels of load attenuation markers and corresponding physical constraint relationships are generated, and the physical constraint inference set and continuous load change results are output.

[0068] Based on the stage physical pressure results, a continuous inference decay flag is generated for nodes with thermal drift risk. If a node continuously experiences an increase in stage physical pressure across multiple consecutive evaluation windows, and the node's inference throughput per unit time shows a decreasing trend, it indicates that the node has experienced physical capacity drift, and physical constraint relationships should be established for the relevant inference stages.

[0069] Physical restrictions include prohibiting further entry, limiting continuous dwell time, limiting context dwell depth, and requiring migration to low-hotspot nodes. Based on the degree of physical load attenuation of a node, continuous inference attenuation is categorized into three levels: mild attenuation, moderate attenuation, and severe attenuation. Mild attenuation only reduces the routing priority of the node; moderate attenuation restricts entry for long-chain inference phases; and severe attenuation requires the release of currently running high-dwelling phases.

[0070] The continuous inference attenuation intensity is calculated based on the stage physical pressure results, the time the node is continuously under high pressure, the node's rated heat dissipation capacity, the node's effective heat dissipation capacity, and the evaluation time window. Furthermore, an attenuation intensity less than 0.3 is considered mild attenuation, 0.3 to 0.7 is considered moderate attenuation, and greater than 0.7 is considered severe attenuation.

[0071] The relationship between continuous inference decay markers and physical constraints is summarized to form a set of physically constrained inference. This set includes at least the constrained node, constrained inference stage, constraint type, allowed dwell time, allowed context dwell depth, recommended migration direction, and constraint removal conditions.

[0072] Simultaneously, based on the changes in attenuation intensity within different evaluation windows, the node's sustained load-bearing capacity change rate is calculated to indicate whether the node's physical load-bearing capacity continues to decline, tends to stabilize, or is recovering. A positive load-bearing capacity change rate indicates that the attenuation intensity is decreasing and the node's load-bearing capacity is tending to recover; a near-zero rate indicates that the node's load-bearing state is stable; and a negative rate indicates that the attenuation intensity continues to increase and the node's load-bearing capacity continues to decline. The system generates node sustained load-bearing capacity change results that include node identifiers and load-bearing capacity change rates.

[0073] S3, based on the physical constraint inference set and the continuous carrying change results, identifies the inference stage to be released and matches the migration target node, completes the continuous context migration and generates hierarchical inference release constraints, and outputs the target physical constraint routing results and the inference stage release mapping set.

[0074] Based on the physical constraint inference set and the results of continuous node load changes, physical constraint routes are regenerated for the inference phases in the AGI inference network. Specifically, phase migration analysis is performed on nodes with continuous inference decay markers to identify long-chain inference phases that need to be released. The remaining physical load space of the migration target node is matched to prevent the inference phase from continuing to enter the hot drift region. The context dwell state during the migration process is continuously inherited to maintain the state continuity of the inference chain during the migration process. Inference release constraints are executed on nodes that have entered the hot accumulation state to restrict subsequent high-dwelling inference phases from continuing to enter the corresponding nodes. Target physical constraint routes are generated according to the physical constraint relationships, and the inference phase release mapping set is output.

[0075] The system, based on the physical constraint inference set and continuous bearing change results, identifies the inference stage to be released and matches the migration target node, completes continuous context migration, generates hierarchical inference release constraints, and outputs the target physical constraint routing results and the inference stage release mapping set, including: Based on the physical constraint inference set and the continuous load change results, according to the load attenuation level and load change trend of the node, the inference stages that need to be released first in the hot drift node are identified, and the set of inference stages to be released is output.

[0076] Read the restricted nodes, restricted inference stages, continuous inference decay markers, and physical constraint relationships from the physical constraint inference set. Combined with the results of continuous load changes of the node, determine whether the node's physical load capacity is recovering, remaining stable, or continuing to decline.

[0077] Based on the node's attenuation level and load change trend, determine the inference phases that need to be released first: for slightly attenuated nodes, do not release the currently executing inference phase directly, but restrict the entry of new high-resident inference phases; for moderately attenuated nodes, only release the high physical stress phase; for severely attenuated nodes, release the high physical stress phase and the ordinary inference phase that has been continuously occupied for more than the allowed residence time.

[0078] Calculate the release priority value for each inference stage to be released. The release priority value comprehensively considers the stage's physical pressure result, the node's continuous inference decay intensity, the stage's timeout dwell degree, and the context depth exceedance degree. The higher the release priority value, the more likely that the inference stage should be released from the current hot drift node. Sort the release priority values ​​from high to low to generate a set of inference stages to be released.

[0079] Based on the set of inference stages to be released, candidate target nodes that meet the migration conditions are screened, the remaining physical carrying capacity ratio and hot zone overlap risk of the candidate target nodes are evaluated, the migration target nodes for each inference stage to be migrated are determined, and the migration target node matching results are output.

[0080] Based on the set of inference stages to be released, candidate target nodes that meet the migration conditions are screened, the remaining physical carrying capacity ratio and hot zone overlap risk of the candidate target nodes are evaluated, the migration target nodes for each inference stage to be migrated are determined, and the migration target node matching results are output, including: Nodes without severe continuous inference decay are selected from all available nodes and used as candidate target nodes for the inference phase to be released.

[0081] Based on the instantaneous heat dissipation capacity, rated heat dissipation capacity, and load attenuation intensity of the candidate target nodes, calculate the proportion of remaining physical load that each candidate target node can use to receive the new inference phase.

[0082] Analyze the degree of overlap between the critical resource occupation area in the migration inference stage and the persistent hot zone location of the candidate target node, and assess the risk of hot zone overlap that will cause heat accumulation again after migration.

[0083] Taking into account the remaining physical carrying capacity ratio and hot zone overlap risk of the candidate target nodes, the migration fit value between each candidate target node and the migration inference stage is calculated. The candidate target node with the highest migration fit value is selected as the migration target node, and the migration target node matching result is output.

[0084] For each inference stage in the set of inference stages to be released, candidate migration target nodes are matched sequentially. Candidate migration target nodes must simultaneously meet three conditions: First, the node does not have a heavy continuous inference decay marker; second, the node's remaining physical carrying capacity can accommodate the stage physical pressure of the stage to be migrated; third, the node's continuous hot zone location must not overlap with the main resource-occupying area of ​​the stage to be migrated.

[0085] First, nodes without severe continuous inference attenuation markers are selected from all available computing power nodes as initial candidate nodes. Then, based on the candidate nodes' effective heat dissipation capacity, heat dissipation protection baseline power, evaluation time window, continuous inference attenuation intensity, and rated heat dissipation capacity, the remaining physical capacity ratio of each candidate node capable of receiving a new inference phase is calculated. Next, the overlap between the critical resource-occupied area of ​​the inference phase to be migrated and the location of the candidate node's persistent hot zone is analyzed to assess the risk of hot zone overlap leading to heat accumulation after migration. Finally, considering both the remaining physical capacity ratio and the risk of hot zone overlap, a migration fit value is calculated for each candidate node and the inference phase to be migrated. The candidate node with the highest migration fit value is selected as the optimal migration target node, generating a target node matching result containing the identifier of the inference phase to be released and the corresponding optimal migration target node identifier.

[0086] For liquid-cooled nodes, priority should be given to nodes with stable coolant inlet temperature and no hot zone formed in the cold plate contact area; for air-cooled nodes, priority should be given to nodes with stable air intake at the front of the air duct and temperature drop in the chip area; for edge data center nodes, priority should be given to nodes in the lower part of the cabinet and with weaker environmental backflow.

[0087] By comprehensively considering the remaining physical carrying capacity and the risk of overlapping hot zones when matching migration target nodes, it is possible to effectively avoid falling back into the hot drift area after migration during the inference stage, thus ensuring the effectiveness of the migration.

[0088] Based on the target node matching results, the context dwelling state in the inference stage is classified and processed. The state transition is completed by copying and verifying first and then switching, and the continuous context transition results are output.

[0089] After identifying the target node, the context-resident state of the inference phase to be migrated is split and inherited. First, the context-resident state is divided into critical context states that must be continuously retained and intermediate states that can be regenerated. The critical context states include the context cache of the current inference phase, attention intermediate states, memory-resident blocks, phase execution positions, completed inference branches, and incomplete inference branches; intermediate states include temporary computation results, cache copies, etc.

[0090] For critical context states that must be continuously retained, a migration approach of copying, validating, and then switching is adopted: first, the critical context state is copied from the source node to the target node; then, the integrity of the copied state is verified; and only after the verification passes is the inference task switched to the target node for execution. For intermediate states that can be regenerated, only the stage index and recovery identifier are retained, and no migration is performed to avoid consuming too much physical storage space during migration.

[0091] During the migration process, the source node must not immediately release the critical context before the target node completes the state reception and verification; the source node releases the corresponding memory resident blocks and high-hot zone resources only after the target node completes the context reception and verification. The context continuity inheritance completeness is calculated, generating a context continuity migration result that includes the inference stage identifier, context inheritance completeness, and migration completion status.

[0092] Based on the contextual continuous migration results, hierarchical inference release constraints are generated for the source heat accumulation nodes corresponding to the inference stage to be migrated. The original routing path is corrected and the release mapping relationship between the inference stage and the nodes is established. The target physical constraint routing results and the inference stage release mapping set are output.

[0093] After the context migration is complete, inference release constraints are applied to the source hot accumulation node. Releasing constraints does not simply clear the node; instead, it restricts subsequent inference stages from entering based on the degree of hot accumulation. For high physical pressure stages that have already migrated out, their memory resident blocks, context cache, and long-occupied high-hot area resources are immediately released. For lightweight inference stages that still need to remain on the source node, only their maximum residency time limit is reduced. For future inference stages waiting to enter the source node, if they are of the high residency depth, high continuous occupancy time, or high memory ratio type, entry into that hot accumulation node is prohibited.

[0094] The generated inference release constraint result includes at least the source node number, the prohibited inference stage type, the allowed lightweight stage type, the maximum dwell time, the maximum context dwell depth, the release completion status, and the constraint removal condition. The release constraint strength is calculated based on a comprehensive analysis of the node's continuous inference decay strength, the integrity of context inheritance, the stage physical pressure result, the stage dwell exceeding the limit, and the degree of hotspot overlap. A higher release constraint strength indicates that a stronger inference release constraint should be imposed on the source node.

[0095] By generating hierarchical inference release constraints, it is possible to alleviate the heat accumulation of source nodes while making full use of the remaining physical carrying capacity of source nodes to handle lightweight inference tasks, thereby improving the utilization rate of computing resources.

[0096] S4, based on the target physical constraint routing results and the inference phase release mapping set, monitors the heat dissipation recovery status of the nodes, determines whether the nodes have escaped the heat accumulation state and restores the carrying capacity of the nodes in stages, executes the progressive backflow of the inference phase and dynamically updates the physical constraint routing, and outputs the AGI inference computing power routing results with dynamic physical carrying capacity balance.

[0097] Based on the target physical constraint routing results and the inference phase release mapping set, dynamic recovery monitoring is performed on nodes in the AGI inference network. Specifically, the heat dissipation recovery status, hotspot reduction status, and continuous inference recovery capability of nodes after inference release are monitored in real time. It is determined whether the node has escaped the continuous heat accumulation state. For nodes that have recovered and stabilized, the phase bearing relationship is re-established, and the corresponding physical constraint relationship is removed. The inference phase after release is re-allocated, so that the recovered nodes gradually re-participate in long-chain inference tasks. The target physical constraint routing results are dynamically updated according to the real-time recovery status of the nodes, forming a dynamically balanced AGI inference computing power route with physical bearing, realizing continuous and stable inference scheduling in hybrid deployment scenarios.

[0098] The process involves monitoring the heat dissipation recovery status of nodes based on the target physical constraint routing results and the inference phase release mapping set, determining whether nodes have escaped the heat accumulation state and restoring the node's carrying capacity in stages, executing progressive backflow in the inference phase and dynamically updating the physical constraint routing, and outputting AGI inference computing power routing results with dynamic physical load balancing, including: Based on the release mapping set during the inference phase, determine the nodes that need to be monitored after release, continuously monitor the changes in the instantaneous heat dissipation capacity, continuous heat zone intensity, and inference throughput of the nodes, and output the heat dissipation recovery status.

[0099] Based on the release mapping set during the inference phase, the source node that has completed inference release is identified, and the effective heat dissipation capacity, actual power consumption, sustained heat zone intensity, inlet and outlet temperature difference of coolant, inlet and outlet temperature difference of airflow, chip temperature drop rate, and inference throughput per unit time of the node after release are continuously monitored.

[0100] Different monitoring priorities are set for different types of nodes: For liquid-cooled nodes, the focus is on whether the temperature at the cold plate outlet decreases and whether the coolant flow rate returns to stability; for air-cooled nodes, the focus is on whether the heat recirculation at the end of the air duct weakens; for edge data center nodes, the focus is on whether the hot air recirculation area above the cabinet shrinks. The monitoring cycle can be set from 10 seconds to 60 seconds, and the evaluation time window can be set from 3 minutes to 15 minutes.

[0101] The heat recovery degree of the node is calculated. The heat recovery degree takes into account the changes in the effective heat dissipation capacity of the node before and after release, the changes in the intensity of the continuous hot zone, and the rated heat dissipation capacity of the node. The result is generated as a node heat recovery monitoring result that includes the node identifier, the heat recovery degree, and the changes in various monitoring parameters.

[0102] Based on the heat dissipation recovery status, combined with the degree of heat dissipation recovery, the extent of hot zone reduction, the increase in inference throughput, and the decrease in attenuation intensity, it is determined whether the node has escaped the state of heat accumulation.

[0103] Based on node heat dissipation recovery, a determination is made as to whether a node has escaped the state of persistent heat accumulation. This determination considers not only whether the node temperature has decreased, but also whether the intensity of the persistent hot zone has decreased, whether effective heat dissipation capacity has recovered, whether inference throughput has rebounded, and whether the intensity of persistent inference decay has decreased. If the node temperature drops briefly but the persistent hot zone does not shrink, the physical constraint relationship must not be lifted; if the node's hot zone shrinks, effective heat dissipation capacity recovers, and no heat accumulation forms again within multiple consecutive evaluation windows, the node is considered to have entered a stable recovery state.

[0104] A node thermal accumulation detachment determination value is calculated, which comprehensively considers the degree of heat dissipation recovery, the degree of hot zone reduction, the recovery rate of inference throughput, and the reduction in attenuation intensity. A thermal accumulation detachment determination value of 0.7 or higher, maintained for two or more consecutive evaluation windows, indicates that the node has detached from the persistent thermal accumulation state. A node thermal accumulation detachment determination result, including the node identifier and the determination result, is generated.

[0105] For nodes that have broken away from the thermal accumulation state, their bearing relationship with the inference stage is reconstructed in stages according to the degree of recovery, and the physical restriction relationship is gradually removed to output the bearing relationship of the restored node stage.

[0106] For nodes that have been determined to have exited the sustained thermal accumulation state, the stage bearing relationship between them and the inference stage is re-established. During reconstruction, the inference stages that the node can receive are divided into three levels according to the residence intensity: low residence stage, medium residence stage, and long chain high residence stage.

[0107] The corresponding physical restrictions are not lifted all at once, but are lifted step by step according to the intensity of the restrictions: for newly recovered nodes, the low-stay stage is allowed to enter first; after the node remains stable within the continuous evaluation window, the medium-stay stage is allowed to enter; only when the node's heat dissipation recovery, heat accumulation removal judgment value, and inference throughput all meet the stability requirements are the long chain high-stay stage allowed to return.

[0108] The calculation phase carries a recovery allowance value, which comprehensively considers the thermal stacking detachment judgment value, heat dissipation recovery degree, current continuous inference attenuation strength, and original release constraint strength. Specifically, when the allowance value is less than 0.3, only the lightweight inference phase is allowed; when the allowance value is between 0.3 and 0.7, the medium-dwelling phase is allowed; and only when the allowance value is greater than 0.7 is the gradual backflow of the long-chain high-dwelling phase allowed. A recovery node phase carrying relationship is generated, including node identifiers, acceptable phase types, and corresponding physical constraint release states.

[0109] Based on the bearing relationship of the recovery node stage, the progressive backflow allocation of the inference stage is executed in order of increasing physical pressure, and the physical constraint route is dynamically updated according to the real-time running status of the node, outputting the AGI inference computing power routing result with dynamic physical bearing balance.

[0110] Based on the phase carrying relationship of the recovery node, the migrated inference phases are reverted. The reverting is carried out in a progressive manner, that is, inference phases with low context dwell depth, low phase physical pressure, and short continuous occupation time are reverted first, then medium pressure phases are reverted, and finally long chain high dwell inference phases are reverted.

[0111] During the reflow process, it is necessary to limit the number of stages a single recovery node can receive within an evaluation time window and set a reflow interval to prevent multiple long-chain stages from reflowing simultaneously, which could cause the node to re-enter a hot-stacking state. A reflow priority value is calculated for each inference stage. This priority value comprehensively considers the stage's recovery permission value, context continuity inheritance integrity, stage physical pressure results, and resource release intervals. A higher reflow priority value indicates that the inference stage is more suitable for priority reflow to the recovery node. The inference stage reflow allocation results are generated by sorting the reflow priority values ​​from high to low.

[0112] By gradually redistributing computing resources, we can make full use of the computing power resources of the recovery nodes while avoiding the physical carrying capacity from drifting again due to too many high-dwelling phases redistributed at once, thus achieving a smooth recovery of computing power resources.

[0113] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A computational power routing method for AGI inference networks, characterized in that, include: Collect and analyze the physical operating state parameters of different types of nodes in the AGI inference network, construct the continuous hot zone distribution and establish the correlation between the inference stage and the continuous hot zone distribution, and output the physical bearing state set and thermal drift correlation results. Specifically, it includes: The physical operating status parameters of edge data center nodes, liquid-cooled nodes, and air-cooled nodes in the AGI inference network are collected uniformly, preprocessed, and output physical operating status records. Based on the physical operation status record, the instantaneous heat dissipation capacity reflecting the current heat dissipation level of each node is calculated according to the node cooling method, and the instantaneous heat dissipation capacity result is output. Based on the instantaneous heat dissipation capability results, combined with the actual power consumption and continuous inference duration, the continuous heat accumulation results of each node during the continuous inference process are calculated. Based on the continuous heat accumulation results, each node is divided into multiple functional areas and heat accumulation is allocated to construct a continuous heat zone distribution. Then, the continuous heat zone is mapped to the current inference stage for occupancy relationship, and the physical bearing state set and heat drift correlation results are output. Based on the physical load-bearing state set and thermal drift correlation results, the resource occupancy characteristics in the statistical inference stage are correlated and matched with the distribution of continuous hot zones. The physical pressure in the inference stage is reconstructed and a load-bearing attenuation marker is generated. The physical constraint inference set and continuous load-bearing change results are output. Specifically, it includes: Based on the physical load state set and thermal drift correlation results, the continuous execution characteristics of each inference stage at the corresponding node are statistically analyzed, key indicators reflecting the resource consumption intensity of the inference stage are extracted, and the resource consumption characteristic set of the inference stage is output. Based on the resource occupancy feature set of the inference stage, the resource occupancy features of the inference stage are correlated and matched with the distribution of persistent hot zones. The driving effect of the inference stage on the formation and diffusion of persistent hot zones is analyzed, and the hot zone matching results of the inference stage are output. During matching, it is not only determined whether the stage is running at a high temperature node, but also whether there is a stable correspondence between its location, duration, heat zone intensity and heat dissipation capacity decrease; Based on the hot zone matching results during the inference phase, and combined with the hot zone correlation strength, memory usage ratio, continuous dwell time, and heat dissipation pressure ratio, the physical pressure result reflecting the physical load pressure on the node during the inference phase is reconstructed. Based on the physical pressure results, and according to the continuous changes in physical pressure and the changing trend of node inference throughput, different levels of load attenuation flags and corresponding physical constraint relationships are generated, and the physical constraint inference set and continuous load change results are output. Physical restrictions include prohibiting further entry, limiting continuous dwell time, limiting context dwell depth, and requiring migration to low-hotspot nodes. Based on the physical capacity attenuation of a node, continuous inference attenuation is categorized into three levels: mild attenuation, moderate attenuation, and severe attenuation. Mild attenuation only reduces the routing priority of the node; moderate attenuation restricts entry into long-chain inference phases; and severe attenuation requires the release of currently running high-dwelling phases. Based on the physical constraint inference set and the continuous bearing change results, identify the inference stage to be released and match the migration target node, complete the continuous context migration and generate hierarchical inference release constraints, and output the target physical constraint routing result and the inference stage release mapping set; Based on the target physical constraint routing results and the inference phase release mapping set, monitor the heat dissipation recovery status of the nodes, determine whether the nodes have escaped the heat accumulation state and restore the carrying capacity of the nodes in stages, execute the progressive backflow of the inference phase and dynamically update the physical constraint routing, and output the AGI inference computing power routing results with dynamic physical load balancing.

2. The computing power routing method for AGI inference networks according to claim 1, characterized in that, The unified collection of physical operating status parameters of edge data center nodes, liquid-cooled nodes, and air-cooled nodes in the AGI inference network is performed and preprocessed to output physical operating status records, including: The data collection cycle is dynamically adjusted according to the load intensity of the AGI inference task. During the peak period of long-chain inference tasks, a short data collection cycle is adopted to simultaneously collect the physical operating status parameters of edge data center nodes, liquid-cooled nodes and air-cooled nodes. The physical operation status parameters of each node are preprocessed to remove outliers that are clearly outside the reasonable range. The missing values ​​are filled in by interpolation between adjacent time points, and the timestamp format of all physical operation status parameters is standardized. The preprocessed physical operation status parameters are classified and stored according to node type and physical operation status parameter category, generating a physical operation status record containing node type identifier, acquisition time, and physical operation status parameters.

3. The computing power routing method for AGI inference networks according to claim 1, characterized in that, The instantaneous heat dissipation capacity, based on the physical operating status record and according to the node cooling method, is calculated to reflect the current heat dissipation level of each node, and the instantaneous heat dissipation capacity result is output, including: Based on the node type identifier in the physical operation status record, automatically select the heat dissipation capacity calculation logic that matches the node cooling method corresponding to the node type identifier; Physical operating status parameters are extracted from the physical operating status records as key heat dissipation capacity calculation parameters, including at least temperature, flow rate, and rotation speed. The basic heat dissipation capacity of each node is calculated and corrected based on the heat dissipation capacity calculation logic to obtain the instantaneous heat dissipation capacity that reflects the current heat dissipation level of each node. The instantaneous heat dissipation capacity of each node is compared with its rated heat dissipation capacity. The heat dissipation margin is calculated and the heat dissipation capacity level is marked. The instantaneous heat dissipation capacity result, which includes node type identifier, instantaneous heat dissipation capacity and heat dissipation margin, is generated.

4. The computing power routing method for AGI inference networks according to claim 1, characterized in that, The calculation of continuous heat accumulation for each node during continuous inference, based on instantaneous heat dissipation capability results, actual power consumption, and continuous inference duration, includes: Extract the current actual power consumption of each node from the physical operation status record, and obtain the continuous inference time of each node from the inference scheduling system; The actual power consumption is used as the heat input value per unit time, and the instantaneous heat dissipation capacity is used as the heat output value per unit time. The net heat accumulation of each node per unit time is calculated. By introducing a heat retention coefficient that reflects the degree of heat retention within each node, and combining the duration of continuous heat inference and the net heat accumulation, the total continuous heat accumulation of each node during the continuous inference process is calculated. The total continuous heat accumulation is compared with the preset multi-level heat accumulation threshold to determine the heat accumulation risk level of each node, and a continuous heat accumulation result containing node type identifier, total continuous heat accumulation, and heat accumulation risk level is generated.

5. The computing power routing method for AGI inference networks according to claim 1, characterized in that, The hot zone matching results based on the inference phase, combined with the hot zone correlation strength, memory usage ratio, continuous dwell time, and heat dissipation pressure ratio, reconstruct the physical pressure results reflecting the physical load pressure on the nodes during the inference phase, including: The dimensions of physical pressure affecting the inference stage are obtained, including the thermal zone correlation strength, memory usage ratio, continuous dwell time, and heat dissipation pressure ratio. Based on the node cooling method and node operating environment, differentiated dimension weight coefficients are assigned to each dimension, and the physical pressure components corresponding to each inference stage in each dimension are calculated respectively, and each physical pressure component is processed as dimensionless. The physical pressure components of each dimension are weighted and summed according to the dimension weight coefficients to obtain the physical pressure of each inference stage, and a physical pressure result containing the inference stage identifier and the corresponding physical pressure is generated.

6. The computing power routing method for AGI inference networks according to claim 1, characterized in that, The system, based on the physical constraint inference set and continuous bearing change results, identifies the inference stage to be released and matches the migration target node, completes continuous context migration, generates hierarchical inference release constraints, and outputs the target physical constraint routing results and the inference stage release mapping set, including: Based on the physical constraint inference set and the continuous load change results, according to the load attenuation level and load change trend of the node, the inference stages that need to be released first in the thermal drift node are identified, and the set of inference stages to be released is output. Based on the set of inference stages to be released, candidate target nodes that meet the migration conditions are screened, the remaining physical carrying capacity ratio and hot zone overlap risk of the candidate target nodes are evaluated, the migration target nodes for each inference stage to be migrated are determined, and the migration target node matching results are output. Based on the target node matching results, the context dwelling state in the inference stage of the migration is classified and processed. The state migration is completed by copying and verifying first and then switching, and the continuous context migration results are output. Based on the contextual continuous migration results, hierarchical inference release constraints are generated for the source heat accumulation nodes corresponding to the inference stage to be migrated. The original routing path is corrected and the release mapping relationship between the inference stage and the nodes is established. The target physical constraint routing results and the inference stage release mapping set are output.

7. The computing power routing method for AGI inference networks according to claim 6, characterized in that, Based on the set of inference stages to be released, candidate target nodes that meet the migration conditions are screened, the remaining physical carrying capacity ratio and hot zone overlap risk of the candidate target nodes are evaluated, the migration target nodes for each inference stage to be migrated are determined, and the migration target node matching results are output, including: Nodes without severe continuous inference decay are selected from all available nodes and used as candidate target nodes for the inference phase to be released. Based on the instantaneous heat dissipation capacity, rated heat dissipation capacity, and load attenuation intensity of the candidate target nodes, calculate the proportion of remaining physical load that each candidate target node can use to receive the new inference phase. Analyze the degree of overlap between the key resource-occupied area in the migration inference stage and the persistent hot zone location of the candidate target node, and assess the risk of hot zone overlap that will cause heat accumulation again after migration. Taking into account the remaining physical carrying capacity ratio and hot zone overlap risk of the candidate target nodes, the migration fit value between each candidate target node and the migration inference stage is calculated. The candidate target node with the highest migration fit value is selected as the migration target node, and the migration target node matching result is output.

8. The computing power routing method for AGI inference networks according to claim 1, characterized in that, The process involves monitoring the heat dissipation recovery status of nodes based on the target physical constraint routing results and the inference phase release mapping set, determining whether nodes have escaped the heat accumulation state and restoring the node's carrying capacity in stages, executing progressive backflow in the inference phase and dynamically updating the physical constraint routing, and outputting AGI inference computing power routing results with dynamic physical load balancing, including: Based on the release mapping set during the inference phase, determine the nodes that need to be monitored after release, continuously monitor the changes in the instantaneous heat dissipation capacity, continuous heat zone intensity and inference throughput of the nodes, and output the heat dissipation recovery status. Based on the heat dissipation recovery status, combined with the degree of heat dissipation recovery, the degree of hot zone reduction, the increase in inference throughput, and the decrease in attenuation intensity, it is determined whether the node has escaped the state of heat accumulation. For nodes that have broken away from the thermal accumulation state, their bearing relationship with the inference stage is reconstructed in stages according to the degree of recovery, and the physical restriction relationship is gradually removed to output the bearing relationship of the restored node stage. Based on the bearing relationship of the recovery node stage, the progressive backflow allocation of the inference stage is executed in order of increasing physical pressure, and the physical constraint route is dynamically updated according to the real-time running status of the node, outputting the AGI inference computing power routing result with dynamic physical bearing balance.

Images