A multi-source data fusion domain large model self-adaptive fine-tuning system and method

Abstract

The application relates to the technical field of artificial intelligence, and discloses a multi-source data fusion domain large model self-adaptive fine-tuning system and method, which comprises a multi-source data access topology module, a self-adaptive scheduling arbitration module, a dynamic decoupling partition calculation module, a gradient momentum conservation module and a dynamic logic node remapping module, wherein the self-adaptive scheduling arbitration module issues topology instructions according to feature distribution; the dynamic decoupling partition calculation module locks the shared operator partition as read-only in response to the instructions, and compresses high-dimensional gradients into low-dimensional sequences in situ through a projection operator; the gradient momentum conservation module superimposes a historical momentum compensation term when the step length is low; and the dynamic logic node remapping module reorganizes the route according to the time delay. The application suppresses multi-source feature gradient conflicts, reduces the communication bus load, eliminates cluster deadlocks, and improves the fine-tuning convergence stability.

Description

Technical Field

[0001] This invention relates to an adaptive fine-tuning system and method for a large domain model based on multi-source data fusion, belonging to the field of artificial intelligence technology. Background Technology

[0002] Currently, in distributed computing clusters, the fusion processing of multi-source heterogeneous data streams is used to form model inputs for specific vertical domains. During parameter backpropagation, each distributed computing node frequently exchanges high-dimensional weight matrices, enabling the general computing model to dynamically adjust weight parameters within the global parameter space. This allows it to adapt to specialized intelligent decision-making systems in fields such as industrial control and knowledge retrieval. The input signal characteristics of heterogeneous data sources may exhibit unpredictable dynamic fluctuations. During backpropagation, these fluctuations can easily cause the local gradient matrix direction of each distributed computing node to shift, leading to gradient direction conflicts between different operator nodes within the parameter space. To maintain global parameter consistency, each node needs to synchronize and verify a large amount of data at high frequency. The inter-chip communication bus is prone to reaching its bandwidth limit, resulting in issues such as repeated data jumps on the bus, cache line failures, and inter-chip clock drift. This can also cause multi-source data packet alignment delays in large-scale computing clusters, further triggering physical memory access conflicts and situations where multiple distributed data processing stages sequentially wait for lock release.

[0003] Existing technologies typically increase the bandwidth of the inter-chip interconnect bus to handle high-load parameter transmission, or add a pre-static filtering module at the system input to filter feature noise. However, pre-static filtering modules increase the latency of the serial processing stage and may prematurely filter out feature data reflecting boundary differences. Simply expanding the hardware bus bandwidth cannot directly eliminate gradient direction conflicts caused by high-frequency burst feature noise. As the number of computing nodes increases, the bus interconnect communication overhead will still grow exponentially. Therefore, it is difficult to obtain a feasible solution solely through hardware expansion. However, these conventional methods of expanding hardware or front-end filtering mainly focus on physical-level link topology adjustments. In the actual distributed training and fine-tuning of large language models, software-level gradient compression and transmission control methods also have shortcomings and are unable to directly cope with highly non-stationary multi-source heterogeneous data. According to flow perturbations, for example, Chinese invention patent application CN118690804A discloses a distributed learning information compression method based on shared random projection. It attempts to perform gradient projection compression by sharing Gaussian random vectors among devices and uses a fixed error feedback and periodic reset mechanism to stabilize the compression error. However, this method implicitly relies on an ideal training environment with relatively stable data distribution. Its error feedback loop lacks the ability to dynamically perceive the nonlinear evolution of the signal-to-noise ratio of multi-source heterogeneous data streams and the degree of gradient divergence between nodes. When faced with industrial-grade high-concurrency burst noise, its fixed moving average adjustment and preset error clearing mechanism can easily lead to parameter update lag or loss of boundary-specific features, which in turn leads to stagnation of local operator updates. It is unable to smooth gradient conflicts in situ and maintain global topological balance under extreme uncertain conditions.

[0004] Therefore, how to dynamically reconstruct the logical topology of the parameter calculation graph based on the fluctuations in the physical channel state and the state differences of each part of the parameter calculation graph, while reducing the bandwidth overhead of bus interconnection between distributed computing nodes and improving the stability of parameter updates during backpropagation, remains a problem that existing technologies need to solve. Summary of the Invention

[0005] To address the problems in the background art, the technical solution of the present invention is as follows: A multi-source data fusion domain large model adaptive fine-tuning system, comprising:

[0006] Multi-source data access topology module, used to access multi-source heterogeneous data streams;

[0007] The adaptive scheduling arbitration module is used to calculate the dynamic decoupling sensitivity index based on the data feature distribution of multiple computing partitions, and to issue a topology yield instruction to the dynamic decoupling partition computing module when the dynamic decoupling sensitivity index exceeds the steady-state discrimination threshold of the non-steady-state fluctuation range.

[0008] The dynamic decoupled partitioning computation module is connected to the adaptive scheduling arbitration module and the distributed computing node module. It is used to lock the shared feature operator partition in read-only state in response to the topology retreat instruction, and to compress the high-dimensional backpropagation gradient matrix in situ into a low-dimensional one-dimensional scalar feature sequence through a one-dimensional lightweight random projection operator for distribution to the adaptive scheduling arbitration module.

[0009] The gradient momentum conservation module, connected to the dynamic decoupled partition calculation module, is used to in-situ superimpose historical momentum compensation terms into the current backpropagation gradient flow when the gradient update step size of the shared feature operator partition is lower than the lower limit cutoff value.

[0010] The dynamic logical node remapping module is connected to the adaptive scheduling arbitration module and the distributed computing node module, respectively. It is used to remap the high-load partition to the distributed computing node module with the shortest current queuing latency when the computing throughput queuing latency of the distributed computing node module exceeds the latency threshold window.

[0011] Preferably, the adaptive scheduling arbitration module issues a topology coordination instruction to the dynamic decoupling partition calculation module when the dynamic decoupling sensitivity index is less than or equal to the steady-state discrimination threshold. The dynamic decoupling partition calculation module is used to respond to the topology coordination instruction to unlock the read-only state of the shared feature operator partition so as to restore the data storage module from the read-only state to the writable state, control the synchronous update of multi-source feature gradients in the global space, and the adaptive scheduling arbitration module includes a differential gradient flow monitoring module, which is used to monitor the data feature distribution of multiple calculation partitions.

[0012] Preferably, the dynamic decoupled partitioning calculation module also includes a local static random access storage module; the one-dimensional lightweight random projection operator is used to call the pseudo-random matrix generated by the fixed prior seed in the local static random access storage module, and to compress the high-dimensional backpropagation gradient matrix in situ through matrix multiplication to generate a low-dimensional one-dimensional scalar feature sequence.

[0013] Preferably, under read-only state locking, the data storage module of the shared feature operator partition within the distributed computing node module is in a non-writable state; when the dynamically decoupled partition computing module receives the topology coordination instruction, the data storage module is restored to a writable state.

[0014] Preferably, the gradient momentum conservation module includes a step size monitoring submodule, a cache extraction submodule, and a momentum superposition submodule. The step size monitoring submodule is used to monitor the gradient update step size of the shared feature operator partition in real time, and sends a compensation trigger signal to the cache extraction submodule when the gradient update step size is determined to be less than 10μs. The cache extraction submodule is used to respond to the compensation trigger signal and retrieve the average gradient history calibration vector of the shared feature operator partition stored in the secondary cache module in three consecutive calculation batches. The momentum superposition submodule is used to use the average gradient history calibration vector as a historical momentum compensation term and superimpose a historical momentum compensation term with a proportional constant of 0.15 in situ into the current backpropagation gradient flow.

[0015] Preferably, the adaptive scheduling arbitration module includes a parameter fine-tuning operator module and a distributed multi-throughput data routing topology network module; the parameter fine-tuning operator module is used to fine-tune the parameters of multi-source heterogeneous data streams; the distributed multi-throughput data routing topology network module is used to perform routing scheduling and distribute the fine-tuned feature data to the shared feature operator partitions within the distributed computing node module.

[0016] Preferably, the dynamic logical node remapping module includes a latency acquisition submodule, a load estimation submodule, and a remapping route decision submodule. The latency acquisition submodule is used to acquire the computational throughput queuing latency of each distributed computing node module. The load estimation submodule is used to receive the computational load estimation parameters of the shared feature operator partitions output by the adaptive scheduling arbitration module. The remapping route decision submodule is used to split and remap the high-load partitions to the two distributed computing node modules with the shortest current computational throughput queuing latency according to the pre-stored load filtering rules when the computational throughput queuing latency of the distributed computing node module exceeds the latency threshold window.

[0017] Preferably, the system is built on a distributed graph computing architecture, and the multi-source heterogeneous data streams accessed by the multi-source data access topology module are cross-time series multi-index distributed data streams or multi-modal parallel feature data streams.

[0018] Preferably, the system is deployed in a distributed graph computing cluster, where the distributed computing node module, the adaptive scheduling arbitration module, and the dynamic logical node remapping module are interconnected through the inter-chip communication bus channel to form a closed-loop network for global parameter adjustment.

[0019] An adaptive fine-tuning method for a domain-wide large model based on multi-source data fusion, used to run an adaptive fine-tuning system for a domain-wide large model based on multi-source data fusion, includes the following steps:

[0020] Step S101: Multi-source heterogeneous data stream access, controlling the multi-source data access topology module to access multi-source heterogeneous data streams;

[0021] Step S102, Adaptive scheduling arbitration, the adaptive scheduling arbitration module calculates the dynamic decoupling sensitivity index based on the data feature distribution of multiple computing partitions, and issues a topology yield instruction to the dynamic decoupling partition computing module when the dynamic decoupling sensitivity index exceeds the steady-state discrimination threshold of the non-steady-state fluctuation range;

[0022] Step S103: Logical topology decoupling, control the dynamic decoupling partition calculation module to respond to the topology retreat instruction and lock the shared feature operator partition in read-only state, use a one-dimensional lightweight random projection operator to compress the high-dimensional backpropagation gradient matrix in situ into a low-dimensional one-dimensional scalar feature sequence, and distribute it to the adaptive scheduling arbitration module through the inter-chip communication bus.

[0023] Step S104, gradient momentum conservation control: when the gradient update step size of the shared feature operator partition is lower than the lower limit cutoff value of 10μs, the gradient momentum conservation module is controlled to add the historical momentum compensation term in situ into the current backpropagation gradient flow to maintain the active state of the operator.

[0024] Step S105, Node logical remapping: When the computing throughput queuing latency of the distributed computing node module exceeds the latency threshold window, the dynamic logical node remapping module is controlled to remap the high-load partition to the distributed computing node module with the shortest current queuing latency.

[0025] Compared with the prior art, the beneficial effects of the present invention are:

[0026] 1. In the adaptive fine-tuning of a domain-wide large model with multi-source data fusion, the cooperation of the multi-source data access topology module, the dynamic decoupling partition calculation module, and the adaptive scheduling arbitration module can determine the degree of gradient conflict between multi-source data based on the data feature distribution of each calculation partition. When the gradient conflict increases, the shared feature operator partition is locked in a read-only state, and the conflict gradient is restricted from continuing to propagate in the global calculation graph. This reduces the impact of local data fluctuations on shared parameters, allowing the parameter update path to be adjusted with the changes in heterogeneous features, thereby improving the fitting stability during the fine-tuning of the domain-wide large model.

[0027] 2. By using a one-dimensional lightweight random projection operator, the high-dimensional backpropagation gradient matrix is ​​compressed into a low-dimensional one-dimensional scalar feature sequence before being transmitted between nodes. This reduces the amount of data synchronized across nodes and lowers the bandwidth usage of the inter-chip communication bus. This can alleviate the repeated jumps in bus data and communication link saturation caused by frequent synchronization of high-dimensional parameters, and reduce the risk of data waiting and deadlock in the distributed computing process.

[0028] 3. By using the gradient momentum conservation module, a historical momentum compensation term is added when the gradient update step size of the shared feature operator partition is too low. This can preserve the necessary gradient update capability for the suppressed local operators, thus avoiding the interruption of gradient flow and the deactivation of operators due to insufficient updates. This ensures that the parameter updates of each computation partition remain continuous and maintains the stability of the global parameter space. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating the adaptive scheduling and logical topology decoupling of multi-source heterogeneous data streams in this invention.

[0030] Figure 2 This is the state diagram of the dynamic decoupling sensitivity evaluation and feature operator partitioning of the present invention.

[0031] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0032] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0033] An adaptive fine-tuning system for a domain-wide large model based on multi-source data fusion, comprising:

[0034] Multi-source data access topology module, used to access multi-source heterogeneous data streams;

[0035] The adaptive scheduling arbitration module is used to calculate the dynamic decoupling sensitivity index based on the data feature distribution of multiple computing partitions, and to issue a topology yield instruction to the dynamic decoupling partition computing module when the dynamic decoupling sensitivity index exceeds the steady-state discrimination threshold of the non-steady-state fluctuation range.

[0036] The dynamic decoupled partitioning computation module is connected to the adaptive scheduling arbitration module and the distributed computing node module. It is used to lock the shared feature operator partition in read-only state in response to the topology retreat instruction, and to compress the high-dimensional backpropagation gradient matrix in situ into a low-dimensional one-dimensional scalar feature sequence through a one-dimensional lightweight random projection operator for distribution to the adaptive scheduling arbitration module.

[0037] The gradient momentum conservation module, connected to the dynamic decoupled partition calculation module, is used to in-situ superimpose historical momentum compensation terms into the current backpropagation gradient flow when the gradient update step size of the shared feature operator partition is lower than the lower limit cutoff value.

[0038] The dynamic logical node remapping module is connected to the adaptive scheduling arbitration module and the distributed computing node module, respectively. It is used to remap the high-load partition to the distributed computing node module with the shortest current queuing latency when the computing throughput queuing latency of the distributed computing node module exceeds the latency threshold window.

[0039] Preferably, the adaptive scheduling arbitration module issues a topology coordination instruction to the dynamic decoupling partition calculation module when the dynamic decoupling sensitivity index is less than or equal to the steady-state discrimination threshold. The dynamic decoupling partition calculation module is used to respond to the topology coordination instruction to unlock the read-only state of the shared feature operator partition so as to restore the data storage module from the read-only state to the writable state, control the synchronous update of multi-source feature gradients in the global space, and the adaptive scheduling arbitration module includes a differential gradient flow monitoring module, which is used to monitor the data feature distribution of multiple calculation partitions.

[0040] Preferably, the dynamic decoupled partitioning calculation module also includes a local static random access storage module; the one-dimensional lightweight random projection operator is used to call the pseudo-random matrix generated by the fixed prior seed in the local static random access storage module, and to compress the high-dimensional backpropagation gradient matrix in situ through matrix multiplication to generate a low-dimensional one-dimensional scalar feature sequence.

[0041] Preferably, under read-only state locking, the data storage module of the shared feature operator partition within the distributed computing node module is in a non-writable state; when the dynamically decoupled partition computing module receives the topology coordination instruction, the data storage module is restored to a writable state.

[0042] Preferably, the gradient momentum conservation module includes a step size monitoring submodule, a cache extraction submodule, and a momentum superposition submodule. The step size monitoring submodule is used to monitor the gradient update step size of the shared feature operator partition in real time, and sends a compensation trigger signal to the cache extraction submodule when the gradient update step size is determined to be less than 10μs. The cache extraction submodule is used to respond to the compensation trigger signal and retrieve the average gradient history calibration vector of the shared feature operator partition stored in the secondary cache module in three consecutive calculation batches. The momentum superposition submodule is used to use the average gradient history calibration vector as a historical momentum compensation term and superimpose a historical momentum compensation term with a proportional constant of 0.15 in situ into the current backpropagation gradient flow.

[0043] Preferably, the adaptive scheduling arbitration module includes a parameter fine-tuning operator module and a distributed multi-throughput data routing topology network module; the parameter fine-tuning operator module is used to fine-tune the parameters of multi-source heterogeneous data streams; the distributed multi-throughput data routing topology network module is used to perform routing scheduling and distribute the fine-tuned feature data to the shared feature operator partitions within the distributed computing node module.

[0044] Preferably, the dynamic logical node remapping module includes a latency acquisition submodule, a load estimation submodule, and a remapping route decision submodule. The latency acquisition submodule is used to acquire the computational throughput queuing latency of each distributed computing node module. The load estimation submodule is used to receive the computational load estimation parameters of the shared feature operator partitions output by the adaptive scheduling arbitration module. The remapping route decision submodule is used to split and remap the high-load partitions to the two distributed computing node modules with the shortest current computational throughput queuing latency according to the pre-stored load filtering rules when the computational throughput queuing latency of the distributed computing node module exceeds the latency threshold window.

[0045] Preferably, the system is built on a distributed graph computing architecture, and the multi-source heterogeneous data streams accessed by the multi-source data access topology module are cross-time series multi-index distributed data streams or multi-modal parallel feature data streams.

[0046] Preferably, the system is deployed in a distributed graph computing cluster, where the distributed computing node module, the adaptive scheduling arbitration module, and the dynamic logical node remapping module are interconnected through the inter-chip communication bus channel to form a closed-loop network for global parameter adjustment.

[0047] Example 1: In this example, the multi-source data access topology module receives a multimodal parallel feature data stream with distributed characteristics. This data stream serves as a specific input form of the multi-source heterogeneous data stream. The feature data of each modality are accessed in parallel within the same computation cycle. The multi-source data access topology module extracts the dimensionless statistical feature scalar of each feature data stream within the current computation cycle from the header identifier of the data transmission buffer, i.e., the input data signal-to-noise ratio variation. and the input data signal-to-noise ratio variation It is written in real time to the 32-bit floating-point register of the data storage module.

[0048] The adaptive scheduling arbitration module includes a parameter fine-tuning operator module, a distributed multi-throughput data routing topology network module, and a differential gradient flow monitoring module. The parameter fine-tuning operator module performs parameter fine-tuning on multi-source heterogeneous data streams, while the distributed multi-throughput data routing topology network module performs routing scheduling. The fine-tuned feature data is distributed to shared feature operator partitions within each distributed computing node module via an inter-chip communication bus channel. During the parameter backpropagation cycle, each distributed computing node module receives fine-tuned feature data with dynamic burst noise. The dynamically decoupled partitioning computation module constructs a parameter computation graph within the computation space, dividing the logical topology of the parameter computation graph into shared feature operator partitions and specific adaptive sensitive partitions corresponding one-to-one with each feature data stream. It also calculates the local gradient matrix difference generated by the current batch of data in each specific adaptive sensitive partition. .

[0049] The differential gradient flow monitoring module is connected to both the multi-source data access topology module and the dynamic decoupling partition calculation module, and reads the input data signal-to-noise ratio variation from the register. and the difference in local gradient matrix Based on this, the data feature distribution of multiple computing partitions is monitored, and the adaptive scheduling arbitration module calculates a dynamic decoupling sensitivity index according to this data feature distribution. The specific calculation formula is as follows: ,in, As a dynamic decoupling sensitivity index, The local gradient matrix dissimilarity is a matrix norm scalar. For the input data signal-to-noise ratio variation; and All are dimensionless pure numbers.

[0050] The adaptive scheduling arbitration module will calculate the dynamic decoupling sensitivity index. The steady-state discrimination threshold between the non-steady-state fluctuation range and the solid-state discrimination threshold in non-volatile memory. Perform discrete conditional control flow determination, where... The steady-state discrimination threshold is set within the non-steady-state fluctuation range, when the dynamic decoupling sensitivity index... Greater than the steady-state discrimination threshold of the non-steady-state fluctuation range When the current multi-source data fusion conflict exceeds the stable fitting boundary, the adaptive scheduling arbitration module issues a topology yield instruction to the dynamic decoupling partitioning calculation module. The dynamic decoupling partitioning calculation module responds to the topology yield instruction and triggers a one-dimensional lightweight random projection operator. This operator calls a fixed prior seed in the local static random access storage module to generate a pseudo-random matrix. Through matrix multiplication, it projects and compresses the high-dimensional backpropagation gradient matrix in situ into a low-dimensional one-dimensional scalar feature sequence, which is then distributed to the adaptive scheduling arbitration module. This in-situ compression reduces the cross-chip bus interconnect bandwidth load between distributed computing node modules by 95%, mitigating bus physical saturation and inter-chip clock drift caused by high-frequency synchronization of high-dimensional parameters. At the same time, the dynamic decoupling partitioning calculation module locks the storage area corresponding to the physical memory address of the weight parameter of the shared feature operator partition into a read-only state. The data storage module is made writable, and the regularization penalty factor of the corresponding specific adaptive sensitive partition is increased by rewriting the register configuration to reduce the flow of conflicting gradients in the global computation graph. During this process, the distributed low-dimensional one-dimensional scalar feature sequence is not used for the evolution and update of the weight parameters of the large model, but is used as the evaluation criterion for topology reconstruction. The adaptive scheduling arbitration module calculates the variance and mean of the sequence and evaluates the feature fluctuation degree between the modules of each distributed computing node in real time. Since the shared feature operator partition is in a read-only state, the model suspends the synchronous interaction of the complete gradient matrix in the global space and only maintains the regularization penalty adjustment of the specific adaptive sensitive partition allowed by the topology retreat. Under the constraint of a 95% reduction in bus bandwidth load, this scalar sequence is used to maintain the adaptive adjustment and state stability of the logical topology.

[0051] During the topology backoff instruction period, the regularization penalty factor of the specific adaptive sensitive partition increases, and the local backpropagation gradient value decays accordingly. The hardware multiply-accumulator in the distributed computing node module may experience data underflow or write stalls when performing weight accumulation operations, posing a risk of parameter flow interruption and operator inactivation. The gradient momentum conservation module, connected to the dynamically decoupled partition computing module, monitors the gradient update step size of the shared feature operator partition. This gradient update step size is represented by the time interval between two consecutive valid update pulses written to the physical memory corresponding to the shared feature operator partition. The step size monitoring submodule captures the physical write enable signal on the inter-chip high-speed bus, and... The corresponding time interval is obtained based on the clock cycle span of the signal. When the time interval is lower than the lower limit cutoff value of 10μs, the step size monitoring submodule sends a compensation trigger signal to the cache extraction submodule. The cache extraction submodule responds to the compensation trigger signal and retrieves the average gradient history calibration vector of the shared feature operator partition stored in the secondary cache module in three consecutive calculation batches. The momentum superposition submodule uses the average gradient history calibration vector as the historical momentum compensation term and superimposes the historical momentum compensation term with a proportional constant of 0.15 in situ into the current backpropagation gradient flow to maintain the activation state of the operators in the shared feature operator partition and offset the effects caused by local degradation.

[0052] When a distributed computing node module experiences uneven computing load due to a sudden influx of heterogeneous data streams from multiple sources, and the computing throughput queuing latency exceeds the latency threshold window, the latency acquisition submodule collects the computing throughput queuing latency of each distributed computing node module. The adaptive scheduling arbitration module generates computing load estimation parameters based on the amount of pending computing tasks in each shared feature operator partition within the current computing cycle. The load estimation submodule synchronously receives these parameters. The dynamic logical node remapping module connects to both the adaptive scheduling arbitration module and the distributed computing node module. The remapping routing decision submodule, based on pre-stored load filtering rules, splits and remaps the high-load shared feature operator partitions to the two distributed computing node modules with the shortest current computing throughput queuing latency, thus converting the topology control of the upper-layer parameter graph in situ into the lower-layer inter-segment synchronization overhead and total... The physical reorganization of line data routing alleviates computational hang-ups and cascade latch waits caused by physical hardware packet loss, enabling multi-source feature gradients to maintain steady-state fitting evolution in the global parameter space. The load screening rule is executed using a logical judgment matrix based on latency difference. After receiving the computational throughput queuing latency of each distributed computing node module, the remapping routing decision submodule sorts the current latency in ascending order and calculates the absolute difference between the current highest latency and the upper bound of the latency threshold window. When the absolute difference is greater than 20ms, a high load judgment is triggered, and half of the computational task corresponding to the high-load shared feature operator partition is split off and routed to the top two distributed computing node modules with the shortest computational throughput queuing latency in the sorting, until the computational throughput queuing latency of the original distributed computing node module carrying the partition falls back to within the latency threshold window.

[0053] When the dynamic decoupling sensitivity index Less than or equal to the steady-state discrimination threshold of the non-steady-state fluctuation range When the specific parameter calculation model is in the steady-state fitting range, the adaptive scheduling arbitration module issues a topology coordination command to the dynamically decoupled partition calculation module. The dynamically decoupled partition calculation module responds to the topology coordination command by unlocking the read-only state of the shared feature operator partition, restoring the data storage module from a read-only state to a writable state, and controlling the synchronous update of multi-source feature gradients across the entire space. The distributed graph computing cluster maintains an adaptive dynamic matching state between the high-dimensional parameter computing network topology and the heterogeneous feature flow path by adjusting the physical distribution of the underlying volatile storage medium and changing the data routing of the inter-chip communication bus channel. In specific implementation, the locking and unlocking of the read-only state... This is achieved by modifying the page attribute register of the memory management unit inside the distributed computing node module. When a topology retreat instruction is received, the dynamic decoupling partition computing module inputs the physical memory base address and space length mapped by the shared feature operator partition into the memory management unit, and clears the read / write control bit of the page attribute register corresponding to the address range, thus rejecting the write enable signal for the storage partition at the physical hardware level and making it unwritable. When a topology coordination instruction is received, the module writes a set logic signal 1 to the same control bit in the page attribute register, reopens the write channel, restores the data storage module to a writable state, and allows the global gradient parameters to be synchronously overwritten and updated.

[0054] Example 2: This example uses a dedicated graph computing cabling test bench with a distributed computing architecture to quantitatively test the feature fitting stability of the distributed computing node module when high-concurrency industrial control data streams are input. The test bench is configured with eight distributed processor cores physically interconnected via a high-speed inter-chip bus, each core serving as a distributed computing node module. The local static random access storage module stores the parameter computation graph topology of the large model's backpropagation cycle within the computing space. The basic multi-source data streams used in the experiment are taken from a publicly available dataset of standard industrial equipment monitoring. The physical input channels of the multi-source data access topology module are superimposed with discrete random pulse overload disturbances generated by a signal generator and Gaussian white noise with a signal-to-noise ratio of 20dB to simulate feature distortion caused by sensor electromagnetic interference. When there are high-dimensional spatial differences in the information entropy change rate and noise power distribution of each data stream, the distributed computing node module experiences gradient matrix direction deflection and inter-chip communication bus channel congestion during backpropagation. The experiment adjusts the disturbance intensity of the multi-source heterogeneous data streams, setting three levels of test conditions (low, medium, and high), and records the input data signal-to-noise ratio change rate. Local gradient matrix difference Dynamic decoupling sensitivity index The calculation includes throughput queuing latency, global model convergence recovery time, and the percentage reduction in bandwidth load of inter-chip bus interconnection.

[0055] This experiment determines the increment constant of the regularization penalty factor and the proportionality constant of the historical momentum compensation term according to predetermined physical constraints. During operation, the proportionality constant simultaneously affects the operator activation state of the shared feature operator partition and the numerical stability of the backpropagation gradient. When the step size monitoring submodule determines that the gradient update step size of the shared feature operator partition is lower than the lower limit cutoff value of 10μs, the gradient momentum conservation module superimposes the historical momentum compensation term in situ into the current backpropagation gradient flow. To determine the value of the proportionality constant, the experiment establishes a monotonically increasing response relationship model between the proportionality constant and the gradient decay damping rate. Under the condition that the computational flux of the global feature graph satisfies the homogeneous dimension conservation, the working window of the proportionality constant is determined to be 0. The response relationship model and its discrete step damping solution are implemented using a discrete conditional branch mapping table pre-stored in memory. In typical high-noise sudden operating conditions, when the input data signal-to-noise ratio variation rate reaches its absolute upper limit and the local gradient matrix difference causes the dynamic decoupling sensitivity index to exceed the steady-state discrimination threshold of the non-steady-state fluctuation range, the mapping table uses the gradient update step size and historical gradient change rate as two-dimensional boundary input parameters. When the gradient update step size is between 10μs and 15μs and the historical gradient change rate fluctuation amplitude of three consecutive calculation batches exceeds 25%, the mapping table enters the branch corresponding to the normal median operating condition and outputs a proportional constant of 0.15, which is used for normal median parameter setting.

[0056] The test was conducted according to the data processing order within the system. The multi-source data access topology module first collected the input status of the multi-source heterogeneous data streams under different noise disturbance intensities, and then the adaptive scheduling arbitration module determined the input data signal-to-noise ratio based on the variation. Difference from local gradient matrix Calculate the dynamic decoupling sensitivity index And compare it with the steady-state discrimination threshold of the non-steady-state fluctuation range. Comparison of the steady-state discrimination thresholds for the non-steady-state fluctuation ranges of 10 samples All values ​​were set to 0.50. During the operation of each sample, when the calculated throughput queuing delay exceeded the delay threshold window, the dynamic logical node remapping module reorganized the routes of the high-load partition according to the pre-stored load filtering rules. The test bench recorded the calculated throughput queuing delay after this processing. The experimental groups included control group 1, control group 2 and control group 3 using the traditional static graph computing architecture, partially missing control group 1 lacking a one-dimensional lightweight random projection operator, partially missing control group 2 lacking a gradient momentum conservation module, out-of-range control group 1 and out-of-range control group 2 with the proportional constant deviating from the working window, and sample group 1, sample group 2 and sample group 3 of the present invention using the complete system.

[0057] Under low-disturbance operating conditions, Sample 1 corresponds to Control Group 1, and the signal-to-noise ratio variation of the input data obtained by the multi-source data access topology module is... The difference in the local gradient matrix obtained by the dynamically decoupled partitioning calculation module is 0.12. The value is 0.22, and the adaptive scheduling arbitration module calculates the dynamic decoupling sensitivity index based on this value. The ratio is 0.20; the scaling factor for this group is set to 0.15, the throughput queuing delay is 12.4ms, the global model convergence recovery time is 145.2s, the inter-chip bus interconnect bandwidth load reduction ratio is 0.0%, and sample 8 corresponds to sample group one of this invention, with its input data signal-to-noise ratio variation. The local gradient matrix dissimilarity is 0.05. The calculated dynamic decoupling sensitivity index is 0.11. The value is 0.10; this indicator does not exceed the steady-state discrimination threshold of the non-steady-state fluctuation range. The adaptive scheduling arbitration module issues topology coordination instructions, the shared feature operator partition remains writable, the proportion constant of this group is set to 0.05, the calculation throughput queuing delay is 8.2ms, the global model convergence recovery time is 32.4s, and the inter-chip bus interconnection bandwidth load reduction ratio is 0.0%.

[0058] Under moderate disturbance conditions, Sample 2 corresponds to Control Group 2, and the input data signal-to-noise ratio variation is... The local gradient matrix dissimilarity is 0.45. The value is 0.68, calculated by the adaptive scheduling arbitration module, representing the dynamic decoupling sensitivity index. The ratio is 0.47; the proportionality constant for this group is set to 0.15, the calculated throughput queuing delay is 28.6ms, the global model convergence recovery time is 189.6s, the inter-chip bus interconnect bandwidth load reduction ratio is 0.0%, sample 9 corresponds to sample group two of this invention, and its input data signal-to-noise ratio variation is... The local gradient matrix dissimilarity is 0.46. The calculated dynamic decoupling sensitivity index is 0.69. Both are 0.47; this indicator does not exceed the steady-state discrimination threshold of the non-steady-state fluctuation range. The adaptive scheduling arbitration module issues topology coordination instructions to keep the shared feature operator partition writable and updatable. The proportion constant of this group is set to 0.15, the calculation throughput queuing delay is 11.5ms, the global model convergence recovery time is 45.6s, and the inter-chip bus interconnection bandwidth load reduction ratio is 0.0%.

[0059] In high-disturbance operating conditions, sample 3 corresponds to control group 3, and the input data signal-to-noise ratio variation is... The local gradient matrix dissimilarity is 0.88. The calculated dynamic decoupling sensitivity index is 1.45. The ratio is 0.77; this group uses a scaling constant of 0.15, the throughput queuing latency is 64.3ms, the global model convergence recovery time is 312.4s, the inter-chip bus interconnect bandwidth load reduction ratio is 0.0%, the traditional static graph computing architecture does not adjust the logical topology of the parameter computing graph according to the dynamic decoupling sensitivity index, the high-dimensional backpropagation gradient matrix continues to be synchronized between distributed computing node modules, and the queuing latency of the inter-chip communication bus channel increases accordingly.

[0060] Sample 10 corresponds to sample group three of this invention under the same perturbation level, with the input data signal-to-noise ratio variation. The local gradient matrix dissimilarity is 0.88. The dynamic decoupling sensitivity index is 1.45. The value is 0.77. After this index exceeds the steady-state discrimination threshold in the non-steady-state fluctuation range, the adaptive scheduling arbitration module issues a topology yield instruction. The dynamic decoupling partition calculation module locks the shared feature operator partition to a read-only state and calls the one-dimensional lightweight random projection operator. Through the pseudo-random matrix generated by the fixed prior seed, it performs matrix multiplication on the high-dimensional backpropagation gradient matrix, compresses it in situ into a low-dimensional one-dimensional scalar feature sequence, and then distributes it to the adaptive scheduling arbitration module. The proportion constant of this group is set to 0.35, the calculation throughput queuing latency is 14.8ms, the global model convergence recovery time is 61.2s, and the inter-chip bus interconnect bandwidth load reduction ratio is 95.2%.

[0061] To differentiate the impact of the one-dimensional lightweight stochastic projection operator and the gradient momentum conservation module on the test results, the experiment set up two groups of partially missing samples. Sample 4 corresponds to the partially missing control group 1, and its input data signal-to-noise ratio variation... The local gradient matrix dissimilarity is 0.87. The calculated dynamic decoupling sensitivity index is 1.42. The value is 0.76, and the proportionality constant is set to 0.15. This index exceeds the steady-state discrimination threshold in the non-steady-state fluctuation range. However, this group lacks a one-dimensional lightweight stochastic projection operator. The high-dimensional backpropagation gradient matrix is ​​still transmitted through the inter-chip communication bus channel. The calculation throughput queuing delay is 58.1ms, the global model convergence recovery time is 245.8s, the inter-chip bus interconnect bandwidth load reduction ratio is 4.2%, and the cascade latch wait time increases accordingly. Sample 5 corresponds to the partially missing control group 2, and its input data signal-to-noise ratio variation rate... The local gradient matrix dissimilarity is 0.89. The calculated dynamic decoupling sensitivity index is 1.46. The value is 0.77, and the proportionality constant is set to 0.15. This index exceeds the steady-state discrimination threshold in the non-steady-state fluctuation range. The one-dimensional lightweight random projection operator still performs in-situ compression on the high-dimensional backpropagation gradient matrix, and the inter-chip bus interconnection bandwidth load is reduced by 95.1%, with a computational throughput queuing delay of 24.5ms. However, this group lacks a gradient momentum conservation module, and the global model convergence recovery time is 288.5s.

[0062] To examine the operational status when the proportional constant deviates from the working window, sample 6 corresponds to control group 1 (out-of-range), and the input data signal-to-noise ratio variation is... The local gradient matrix dissimilarity is 0.88. The calculated dynamic decoupling sensitivity index is 1.44. The value is 0.77; this index exceeds the steady-state discrimination threshold in the non-steady-state fluctuation range, and the proportionality constant is set to 0.02, which is lower than the lower limit of the working window; the one-dimensional lightweight random projection operator still performs in-situ compression, with a computational throughput queuing delay of 22.3ms, a global model convergence recovery time of 254.1s, and a 94.8% reduction in inter-chip bus interconnect bandwidth load. Under this proportionality constant, the historical momentum compensation term superimposed on the current backpropagation gradient flow is relatively small, and the activation maintenance effect of operators in the shared feature operator partition is correspondingly weakened. Sample 7 corresponds to the out-of-range control group 2, and its input data signal-to-noise ratio variation rate is... The local gradient matrix dissimilarity is 0.86. The calculated dynamic decoupling sensitivity index is 1.43. The value is 0.77; this index exceeds the steady-state discrimination threshold in the non-steady-state fluctuation range, and the proportionality constant is set to 0.55, which is higher than the upper limit of the working window; the one-dimensional lightweight random projection operator still performs in-situ compression, the computation throughput queuing delay is 25.1ms, the global model convergence recovery time is 295.3s, and the inter-chip bus interconnection bandwidth load is reduced by 94.9%. Under this proportionality constant, the superposition of the historical momentum compensation term in the backpropagation gradient flow increases, the homogeneous dimension conservation constraint of the backpropagation gradient flow is violated, and numerical disturbance accumulation and convergence oscillation occur in the model convergence process.

[0063] As can be seen from the processing of each set of data, when the dynamic decoupling sensitivity index does not exceed the steady-state discrimination threshold of the non-steady-state fluctuation range, the system maintains the writable state of the shared feature operator partition, and the multi-source feature gradient is updated synchronously in the global space. When the index exceeds the threshold, the system locks the shared feature operator partition through the topology backoff instruction, and the one-dimensional lightweight random projection operator reduces the cross-chip transmission volume of the high-dimensional gradient matrix. Some missing samples reflect that the one-dimensional lightweight random projection operator mainly affects the inter-chip bus interconnect bandwidth load and the computational throughput queuing delay, while the gradient momentum conservation module mainly affects the update continuity of the shared feature operator partition during the topology backoff period. The out-of-range samples show that after the proportional constant deviates from the working window, even if the one-dimensional lightweight random projection operator still reduces the inter-chip bus interconnect bandwidth load, the global model convergence recovery time will still increase. The above processing constrains the high-dimensional gradient transmission volume, computational throughput queuing delay, and operator update continuity under high perturbation input, thereby mitigating the risk of cluster deadlock caused by inter-chip bus saturation, distributed computing node module load imbalance, and cascade latch waiting.

[0064] Example 3: The current high-concurrency industrial control computing center accesses a multi-indicator distributed data stream across time series. Its internal distributed computing node module consists of 8 distributed processor cores physically interconnected via a high-speed inter-chip bus. It executes the parameter adaptive fine-tuning task of a large domain model on a distributed graph computing architecture. The dynamic decoupling partition computing module reads the logical topology of the parameter computing graph fixed in the local static random access storage module at the start of each backpropagation cycle. It decouples the weight tensors at each layer of the attention mechanism operator in the parameter computing graph according to the feature dimension, dividing them into shared feature operator partitions and specific adaptive sensitive partitions corresponding one-to-one with the multi-source heterogeneous data streams. The gradient increments generated by each specific adaptive sensitive partition are written to the dynamic buffer. The forward feature stream and the backward feature stream at the attention mechanism operator are logically isolated in the volatile random access memory, which is used to spatially decouple the local gradient matrix direction deflection caused by heterogeneous inputs.

[0065] During the operation of the distributed graph computing network, the multi-source data access topology module collects the raw input of multi-source heterogeneous data streams in real time, and calculates the signal-to-noise ratio variation of the input data by using a hardware counter to count the characteristic packet loss rate on the statistical bus. Then, it is written to a 32-bit floating-point register in a designated memory. The dynamically decoupled partitioning computation module obtains the local gradient matrix generated by the current computation batch in each specific adaptive sensitive partition based on the local gradient change rate exchanged between each distributed computing node module, and calculates the local gradient matrix difference. During calculation, the local gradient matrix of the current batch is first subtracted element-wise from the baseline gradient matrix of the previous historical standard batch stored in volatile memory to obtain the difference matrix. The multiply-accumulator performs a squaring operation on each element of the difference matrix, globally accumulates the squared values, and then performs a square root operation to output a unique positive real scalar as the local gradient matrix difference. .

[0066] The adaptive scheduling arbitration module reads the signal-to-noise ratio variation of the input data. Difference from local gradient matrix ,by and The ratio is used as a dynamic decoupling sensitivity index. To reduce the difference in parameter settings at the decision boundary, the adaptive scheduling arbitration module determines the steady-state discrimination threshold for the non-steady-state fluctuation range through a power-on self-test calibration procedure. During the initialization self-test phase, the system injects standard monochromatic pulses into the physical channel. The step size monitoring submodule continuously collects the inter-chip synchronization delay variance at 256 sampling points, and uses the reciprocal of this synchronization delay variance as the intrinsic fitting benchmark value of the current computing environment. After inputting this benchmark value into the dynamic adjustment function, the steady-state discrimination threshold of the non-steady-state fluctuation range of this fine-tuning task is obtained. Its calibration value is 0.50. The dynamic adjustment function adopts a linear weighted mapping. The arithmetic logic unit first multiplies the intrinsic fitting benchmark value by a constant scaling factor of 1.25 fixed in the firmware using a multiplier, and then adds a fixed engineering correction constant of 0.12 by an adder. The resulting value is used as the steady-state discrimination threshold for the non-steady-state fluctuation range of this fine-tuning task. .

[0067] When the dynamic decoupling sensitivity index Steady-state discrimination threshold for non-steady-state fluctuation range greater than 0.50 At this time, the adaptive scheduling arbitration module enters the topology yield branch and issues a topology yield instruction to the dynamic decoupling partition calculation module. The dynamic decoupling partition calculation module responds to the instruction by locking the physical memory address of the weight parameters of the shared feature operator partition to read-only mode, and simultaneously opening the data physical interface to the one-dimensional lightweight random projection operator. The one-dimensional lightweight random projection operator reads a fixed prior seed from the local static random access storage module and generates a row dimension of [missing information] in the temporary high-speed temporary storage register. Column dimensions are The pseudo-random matrix, where, This represents the compressed low-dimensional scale. Representing the original high-dimensional scale, this operator uses a one-dimensional multiplier to make the feature dimension equal to 1. Perform matrix multiplication between the high-dimensional backpropagation gradient matrix and the pseudo-random matrix to compress the high-dimensional backpropagation gradient matrix in situ into a matrix with a feature dimension of 1. The low-dimensional one-dimensional scalar feature sequence is distributed to the adaptive scheduling arbitration module. After compression, the complete high-dimensional backpropagation gradient matrix is ​​no longer transmitted between distributed computing node modules, and the synchronization overhead of the inter-chip bus is reduced accordingly.

[0068] When the one-dimensional lightweight random projection operator performs in-situ compression, the dynamic decoupling partition calculation module synchronously triggers the gradient momentum conservation module for compensation. The step size monitoring submodule monitors the gradient update step size of the shared feature operator partition in real time. When the gradient update step size is lower than the lower limit cutoff value of 10μs, the cache extraction submodule locates the physical address of the second-level cache corresponding to the shared feature operator partition, retrieves the average gradient history calibration vector stored in the partition in three consecutive calculation batches, and uses it as the historical momentum compensation term. The momentum superposition submodule in-situ superimposes the historical momentum compensation term with a proportional constant of 0.15 into the current backpropagation gradient flow to correct the feature flow direction deflection caused by high-frequency burst noise interference, so that the local gradient update step size converges to 14.2μs.

[0069] After completing in-situ compression and historical momentum compensation, the distributed graph computing cluster adjusts the physical distribution of the underlying storage medium according to the operating status of each partition and changes the data routing of the inter-chip communication bus channel to keep the high-dimensional parameter computing network topology matched with the heterogeneous feature flow path.

[0070] Example 4: This example combines Figures 1 to 2 This document describes an adaptive fine-tuning system and method for a large-scale domain model based on multi-source data fusion, as follows: Figure 1As shown, the method includes step S101: multi-source heterogeneous data stream access, controlling the multi-source data access topology module to access multi-source heterogeneous data streams, and then proceeding to step S102: adaptive scheduling arbitration, controlling the adaptive scheduling arbitration module to calculate the dynamic decoupling sensitivity index based on the data feature distribution of multiple computing partitions, and issuing a topology yield instruction to the dynamic decoupling partition computing module when the dynamic decoupling sensitivity index exceeds the steady-state discrimination threshold of the non-steady-state fluctuation range, followed by step S103: logical topology decoupling, controlling the dynamic decoupling partition computing module to respond to the topology yield instruction and lock the shared feature operator partition in a read-only state, and using a one-dimensional lightweight random projection operator to reverse the high-dimensional... The propagation gradient matrix is ​​compressed in situ into a low-dimensional one-dimensional scalar feature sequence, which is then distributed to the adaptive scheduling arbitration module via the inter-chip communication bus. This process then proceeds to step S104, gradient momentum conservation control. When the gradient update step size of the shared feature operator partition is lower than the lower limit cutoff value of 10μs, the gradient momentum conservation module is controlled to add historical momentum compensation terms in situ to the current backpropagation gradient flow to maintain the active state of the operator. Finally, the process reaches step S105, node logic remapping. When the computation throughput queuing delay of the distributed computing node module exceeds the delay threshold window, the dynamic logic node remapping module is controlled to remap the high-load partition to the distributed computing node module with the shortest current queuing delay.

[0071] like Figure 2 As shown, after the multi-source heterogeneous data stream is accessed, it is input to the dynamic decoupling sensitivity index evaluation and judgment stage. Based on the evaluation and judgment result, it is transferred to the global spatial synchronization update process or the shared feature operator partition lock read-only state process. After entering the shared feature operator partition lock read-only state, the operation of superimposing historical momentum compensation items and remapping high-load partition operation are triggered downstream in parallel. There is a one-way control state recovery process between the shared feature operator partition lock read-only state and the global spatial synchronization update.

[0072] Example 5: Before deploying a domain-wide large-scale adaptive fine-tuning system for multi-source data fusion in a current high-concurrency industrial control data flow center, a standardized pre-deployment calibration procedure is first executed to eliminate the differences in multi-source heterogeneous data flow inputs caused by hardware manufacturing tolerances between distributed computing node modules. After the system enters offline calibration mode, eight distributed processor cores are connected to the same controlled multi-source test signal source. Each core constitutes a corresponding distributed computing node module. The signal source outputs a standard pseudo-random noise sequence with an amplitude fluctuation range within ±10%. Each distributed computing node module samples and extracts features from this sequence within a preset idle computing power window, calculates the average computing power consumption and feature recognition deviation value of each node per unit time, and then obtains the initial computing compensation offset of each node based on the two calculation results and writes it into the read-only register of the corresponding node.

[0073] During real-time system operation, baseline calibration is performed every 1000 backpropagation training cycles. During calibration, the system monitors the weight update saturation index of the specific adaptive sensitive partition in the dynamic decoupled partition calculation module. When the long-term moving average of this index continuously deviates from the initial baseline and exceeds the preset tolerance value of 5%, an online incremental recalibration operation is triggered. The system reweights and normalizes the statistical characteristics of each calculation partition according to the current operating status, and then writes the obtained normalization coefficients into the runtime dynamic configuration table. Through offline calibration and online incremental recalibration, the logical topology of the parameter calculation graph in each distributed computing node module remains within the pre-calibrated optimal fitting range, eliminating the systematic fine-tuning deviation caused by differences in hardware operating environment.

[0074] To address the data processing risks arising from the injection of unstructured features during the dynamic updating of the domain knowledge base, the system implements privacy protection and data compliance filtering procedures at the data fusion layer. The original input data streams participating in data fusion first enter the data desensitization and built-in processing stage. A hardware-level hash mapping unit is set up at the data input end. For potentially sensitive fields such as user identifiers and timestamps in multi-source heterogeneous data streams, this unit uses a pre-built secure hash algorithm to perform real-time desensitization calculations, outputting feature identifiers that are not related to the original identity information. The desensitization process is completed before the data enters the local high-speed cache of the distributed computing node module, ensuring that the data streams entering the logical topology space of the parameter computing graph do not contain unprocessed original user identity information.

[0075] When the fitted risk parameters output by the system exceed the preset compliance warning threshold, the anti-discrimination logic correction mechanism is triggered. This mechanism compares the fitted risk parameters with the pre-stored objective behavioral parameter baseline matrix item by item, and performs backtracking adjustment on the weight parameters of the corresponding partition only according to frequency, density or response time. The adjustment process does not use decision weights associated with natural person identity tags. Through data input anonymization processing and weight adjustment based on objective behavioral parameters, the fine-tuning process focuses on the numerical fitting of multi-source heterogeneous data streams and controls the risks of privacy leakage and decision bias.

[0076] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0077] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A domain-wide large-scale model adaptive fine-tuning system based on multi-source data fusion, characterized in that, include: Multi-source data access topology module, used to access multi-source heterogeneous data streams; The adaptive scheduling arbitration module is used to calculate the dynamic decoupling sensitivity index based on the data feature distribution of multiple computing partitions, and to issue a topology yield instruction to the dynamic decoupling partition computing module when the dynamic decoupling sensitivity index exceeds the steady-state discrimination threshold of the non-steady-state fluctuation range. The dynamic decoupled partitioning computation module is connected to the adaptive scheduling arbitration module and the distributed computing node module. It is used to lock the shared feature operator partition in read-only state in response to the topology retreat instruction, and to compress the high-dimensional backpropagation gradient matrix in situ into a low-dimensional one-dimensional scalar feature sequence through a one-dimensional lightweight random projection operator for distribution to the adaptive scheduling arbitration module. The gradient momentum conservation module, connected to the dynamic decoupled partition calculation module, is used to in-situ superimpose historical momentum compensation terms into the current backpropagation gradient flow when the gradient update step size of the shared feature operator partition is lower than the lower limit cutoff value. The dynamic logical node remapping module is connected to the adaptive scheduling arbitration module and the distributed computing node module, respectively. It is used to remap the high-load partition to the distributed computing node module with the shortest current queuing latency when the computing throughput queuing latency of the distributed computing node module exceeds the latency threshold window.

2. The adaptive fine-tuning system for a large domain model based on multi-source data fusion according to claim 1, characterized in that, When the dynamic decoupling sensitivity index is less than or equal to the steady-state discrimination threshold, the adaptive scheduling arbitration module issues a topology coordination instruction to the dynamic decoupling partition calculation module. The dynamic decoupling partition calculation module is used to respond to the topology coordination instruction to unlock the read-only state of the shared feature operator partition so as to restore the data storage module from the read-only state to the writable state, control the synchronous update of multi-source feature gradients in the global space, and the adaptive scheduling arbitration module includes a differential gradient flow monitoring module, which is used to monitor the data feature distribution of multiple calculation partitions.

3. The adaptive fine-tuning system for a large domain model based on multi-source data fusion according to claim 1, characterized in that, The dynamic decoupled partitioning computation module also includes a local static random access storage module; the one-dimensional lightweight random projection operator is used to call the pseudo-random matrix generated by the fixed prior seed in the local static random access storage module, and to compress the high-dimensional backpropagation gradient matrix in situ through matrix multiplication to generate a low-dimensional one-dimensional scalar feature sequence.

4. The adaptive fine-tuning system for a large domain model based on multi-source data fusion according to claim 2, characterized in that, Under read-only locking, the data storage module of the shared feature operator partition within the distributed computing node module is in a non-writable state; when the dynamically decoupled partition computing module receives the topology coordination instruction, the data storage module is restored to a writable state.

5. The adaptive fine-tuning system for a large domain model based on multi-source data fusion according to claim 1, characterized in that, The gradient momentum conservation module includes a step size monitoring submodule, a cache extraction submodule, and a momentum superposition submodule. The step size monitoring submodule monitors the gradient update step size of the shared feature operator partition in real time and sends a compensation trigger signal to the cache extraction submodule when the gradient update step size is determined to be less than 10μs. The cache extraction submodule responds to the compensation trigger signal by retrieving the average gradient history calibration vector of the shared feature operator partition stored in the secondary cache module in three consecutive calculation batches. The momentum superposition submodule uses the average gradient history calibration vector as a historical momentum compensation term and superimposes a historical momentum compensation term with a proportionality constant of 0.15 into the current backpropagation gradient flow.

6. The adaptive fine-tuning system for a large domain model based on multi-source data fusion according to claim 1, characterized in that, The adaptive scheduling arbitration module includes a parameter fine-tuning operator module and a distributed multi-throughput data routing topology network module. The parameter fine-tuning operator module is used to fine-tune the parameters of multi-source heterogeneous data streams. The distributed multi-throughput data routing topology network module is used to perform routing scheduling and distribute the fine-tuned feature data to the shared feature operator partitions within the distributed computing node module.

7. The adaptive fine-tuning system for a large domain model based on multi-source data fusion according to claim 1, characterized in that, The dynamic logical node remapping module includes a latency acquisition submodule, a load estimation submodule, and a remapping routing decision submodule; The latency acquisition submodule is used to collect the computing throughput queuing latency of each distributed computing node module; The load estimation submodule is used to receive the calculated load estimation parameters of the shared feature operator partitions output by the adaptive scheduling arbitration module; The remapping routing decision submodule is used to split and remap high-load partitions to the two distributed computing node modules with the shortest current computing throughput queuing latency when the computing throughput queuing latency of the distributed computing node module exceeds the latency threshold window, based on the pre-stored load filtering rules.

8. The adaptive fine-tuning system for a large domain model based on multi-source data fusion according to claim 1, characterized in that, The system is built on a distributed graph computing architecture. The multi-source heterogeneous data streams accessed by the multi-source data access topology module are distributed data streams with multiple time series indicators or multimodal parallel feature data streams.

9. The adaptive fine-tuning system for a large domain model based on multi-source data fusion according to claim 1, characterized in that, The system is deployed in a distributed graph computing cluster. The distributed computing node module, the adaptive scheduling arbitration module, and the dynamic logical node remapping module are interconnected through the inter-chip communication bus channel, forming a closed-loop network for global parameter adjustment.

10. A method for adaptive fine-tuning of a domain-wide large model based on multi-source data fusion, used to run the adaptive fine-tuning system for a domain-wide large model based on multi-source data fusion as described in claim 1, characterized in that, Includes the following steps: Step S101: Multi-source heterogeneous data stream access, controlling the multi-source data access topology module to access multi-source heterogeneous data streams; Step S102, Adaptive scheduling arbitration, the adaptive scheduling arbitration module calculates the dynamic decoupling sensitivity index based on the data feature distribution of multiple computing partitions, and issues a topology yield instruction to the dynamic decoupling partition computing module when the dynamic decoupling sensitivity index exceeds the steady-state discrimination threshold of the non-steady-state fluctuation range; Step S103: Logical topology decoupling, control the dynamic decoupling partition calculation module to respond to the topology retreat instruction and lock the shared feature operator partition in read-only state, use a one-dimensional lightweight random projection operator to compress the high-dimensional backpropagation gradient matrix in situ into a low-dimensional one-dimensional scalar feature sequence, and distribute it to the adaptive scheduling arbitration module through the inter-chip communication bus. Step S104, gradient momentum conservation control: when the gradient update step size of the shared feature operator partition is lower than the lower limit cutoff value of 10μs, the gradient momentum conservation module is controlled to add the historical momentum compensation term in situ into the current backpropagation gradient flow to maintain the active state of the operator. Step S105, Node logical remapping: When the computing throughput queuing latency of the distributed computing node module exceeds the latency threshold window, the dynamic logical node remapping module is controlled to remap the high-load partition to the distributed computing node module with the shortest current queuing latency.

Images