Matrix-free rigid equation solution adaptation method and system based on large model reasoning acceleration

Abstract

The invention belongs to the cross technical field of artificial intelligence large model reasoning acceleration and rigid differential equation numerical solution, and discloses a matrix-free rigid equation solution adaptation method and system based on large model reasoning acceleration. According to the method, a core calculation module without matrix rigid solution and a large model reasoning acceleration framework are subjected to lightweight adaptation, a spectral radius estimation link driven by quantum annealing is optimized through tensor rearrangement and operator fusion, and a hybrid precision calculation process is optimized in combination with a video memory dynamic scheduling strategy of a large model; resource sharing and collaborative acceleration of rigid equation solving and large model reasoning are achieved; and meanwhile, the rigidity ratio pre-judgment logic is optimized based on the feature extraction capability of the large model, and the real-time performance of adaptive step length control is improved. According to the method, the problem of resource competition when traditional matrix-free rigid solution and large model reasoning coexist is solved, cross-scene calculation efficiency is greatly improved, and the method can be widely applied to the fields of large model and numerical solution fusion such as intelligent equipment real-time simulation, aerospace intelligent deduction and chemical process intelligent regulation and control.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary fields of accelerating large-scale artificial intelligence inference and numerical solution of rigid differential equations, specifically involving a matrix-free rigid equation solution adaptation method and system based on accelerating large-scale model inference. Summary of the Invention

[0002] To address the shortcomings of existing technologies, this invention aims to provide a matrix-free rigid equation solving adaptation method and system based on large model inference acceleration. It lightweightly adapts the core computational modules of matrix-free rigid equation solving to a large model inference acceleration framework. Through tensor rearrangement and operator fusion optimization of the quantum annealing-driven spectral radius estimation stage, and by combining a large model's dynamic memory scheduling strategy to optimize the mixed-precision computation process, resource sharing and collaborative acceleration of rigid equation solving and large model inference are achieved. Simultaneously, the rigidity ratio prediction logic is optimized based on the feature extraction capabilities of the large model, improving the real-time performance of adaptive step size control. This invention solves the resource competition problem when traditional matrix-free rigid equation solving and large model inference coexist, significantly improving cross-scenario computational efficiency. It can be widely applied in fields such as real-time simulation of intelligent equipment, intelligent aerospace simulation, and intelligent control of chemical processes, where large models and numerical solutions are integrated.

[0003] The adaptive method for solving matrix-free rigid equations based on large model inference acceleration described in this invention includes the following steps:

[0004] S1 Lightweight Module Adaptation: The core module of matrix-free rigid solution is decomposed into spectral radius estimation, mixed precision calculation, and adaptive step size control sub-modules. Tensor rearrangement and operator fusion of the sub-modules are completed based on the tensor calculation specification of the large model inference acceleration framework, so that the sub-modules are compatible with the hardware calculation logic of large model inference.

[0005] S2 Collaborative Resource Scheduling: Extract the dynamic memory scheduling strategy of the large model inference acceleration framework, construct a shared memory pool for rigid solving and large model inference, and dynamically allocate and reclaim memory resources based on the computational load of spectral radius estimation and the token generation speed of large model inference.

[0006] S3 Large Model-Assisted Prediction: Input the initial parameters and solution interval of the rigid equation into the large model feature extraction network to obtain the rigidity ratio prediction value. Use this prediction value as the initial reference value for spectral radius estimation in matrix-free rigidity solution to narrow the search range of spectral radius.

[0007] S4 Collaborative Accelerated Solution: Based on the adapted sub-modules, dynamically scheduled memory resources, and the initial stiffness ratio predicted by the large model, quantum annealing-driven spectral radius estimation, adaptive mixed-precision calculation, and step size control for stiffness ratio adaptation are executed to complete the collaborative accelerated solution of matrix-free stiffness equations.

[0008] The matrix-free rigid equation solving adaptation system based on large model inference acceleration described in this invention employs the above-mentioned method, including:

[0009] The module adaptation unit is used to complete tensor rearrangement and operator fusion of the matrix-free rigid solution submodule, so as to achieve hardware logic compatibility with the large model inference acceleration framework.

[0010] The resource scheduling unit is used to build a shared memory pool and to dynamically allocate and reclaim memory resources based on large model inference load and rigid solution computation load.

[0011] The stiffness ratio prediction unit is used to call the large model feature extraction network, input stiffness equation parameters to obtain the stiffness ratio prediction value and transmit it to the spectral radius estimation unit;

[0012] The collaborative solution unit includes a quantum annealing spectral radius estimation subunit, a mixed-precision calculation subunit, and an adaptive step size control subunit. Based on the adaptation results, resource scheduling, and prediction values, it completes the solution of matrix-free rigid equations.

[0013] The interface unit is used to realize bidirectional data interaction between this system and the large model inference acceleration framework, and supports real-time feedback of solution results to the large model inference process.

[0014] The system is deployed on a large model inference server and shares CPU / GPU / TPU computing resources with the large model inference framework. The system is lightweight and packaged as a whole, with a startup memory usage of ≤512MB and supports hot-swappable access.

[0015] Beneficial effects

[0016] 1. High module compatibility: This invention has achieved deep adaptation between the matrix-free rigid solution module and the large model inference acceleration framework, realizing the sharing of hardware computing resources, improving the overall computing efficiency in the fusion scenario by ≥40%, and improving it by more than 30% compared with the TensorRT acceleration solution;

[0017] 2. Resolving resource contention: By constructing a shared memory pool and a dynamic scheduling strategy, the competition for memory and computing cores between large model inference and rigid solution is eliminated. The speed reduction of a single task when the two coexist is ≤10%, which is far better than the speed reduction of more than 40% of the existing solutions.

[0018] 3. Improve stiffness ratio estimation efficiency: Based on the feature extraction capability of large models, stiffness ratio prediction is achieved, reducing the search range of spectral radius estimation by more than 80%, and reducing the time consumption of spectral radius estimation by ≥70%, which is more than 60% lower than the Transformers plugin solution, meeting the low latency requirements of real-time fusion scenarios;

[0019] 4. Unified precision specification: Adopting the FP16 / FP8 hybrid precision specification for large model inference, the overhead of cross-frame precision conversion is eliminated, and the precision conversion time is reduced by 100% (no additional conversion required);

[0020] 5. Lightweight and easy to deploy: The system of this invention adopts lightweight packaging, has low startup memory consumption, supports hot-swappable calls, and can be directly deployed on existing large model inference servers without additional hardware modifications, resulting in low deployment costs;

[0021] 6. Adaptation to general frameworks: This invention is designed for general large model inference frameworks that support tensor parallelism and pipelined parallelism, giving full play to the advantages of tensor parallelism and pipelined parallelism of general frameworks. Compared with dedicated framework adaptation solutions, the acceleration effect is improved by ≥25%.

Claims

1. A matrix-free rigid equation solving adaptation method based on large model inference acceleration, characterized in that, Includes the following steps: S1 Lightweight Module Adaptation: The core module of matrix-free rigid solution is decomposed into spectral radius estimation, mixed precision calculation, and adaptive step size control sub-modules. Tensor rearrangement and operator fusion of the sub-modules are completed based on the tensor calculation specification of the large model inference acceleration framework, so that the sub-modules are compatible with the hardware calculation logic of large model inference. S2 Collaborative Resource Scheduling: Extract the dynamic memory scheduling strategy of the large model inference acceleration framework, construct a shared memory pool for rigid solving and large model inference, and dynamically allocate and reclaim memory resources based on the computational load of spectral radius estimation and the token generation speed of large model inference. S3 Large Model-Assisted Prediction: Input the initial parameters and solution interval of the rigid equation into the large model feature extraction network to obtain the rigidity ratio prediction value. Use this prediction value as the initial reference value for spectral radius estimation in matrix-free rigidity solution to narrow the search range of spectral radius. S4 Collaborative Accelerated Solution: Based on the adapted sub-modules, dynamically scheduled memory resources, and the initial stiffness ratio predicted by the large model, quantum annealing-driven spectral radius estimation, adaptive mixed-precision calculation, and step size control for stiffness ratio adaptation are executed to complete the collaborative accelerated solution of matrix-free stiffness equations.

2. The adaptation method according to claim 1, characterized in that, In step S1, operator fusion targets the energy function calculation stage of quantum annealing. It integrates multiple basic operators for the equivalent product of matrices and vectors and the energy difference calculation into a composite operator supported by the large model inference framework. After operator fusion, the calculation time is reduced by ≥30%.

3. The adaptation method according to claim 1, characterized in that, In step S2, the dynamic allocation rule of the shared memory pool is as follows: when the large model inference is at the peak of token generation (token generation rate ≥ 1000 tokens / s), 512MB of basic video memory resources are allocated for rigid solution, and only the coarse estimation of the spectral radius is maintained; when the large model inference is in the idle window period (token generation rate < 100 tokens / s), 4GB of peak video memory resources are allocated for rigid solution, and the fine estimation of the spectral radius and the fine calculation of mixed precision are performed.

4. The adaptation method according to claim 1, characterized in that, In step S3, the large model feature extraction network uses a feedforward neural network layer of the large model. The input is a 128-dimensional feature vector consisting of the initial value statistical features, solution interval features, and directional derivative features of the rigid equation. The output is 0~10. 6 The predicted stiffness ratio is within the range, and the error between the predicted value and the actual stiffness ratio is ≤15%.

5. The adaptation method according to claim 1, characterized in that, In step S4, the mixed-precision calculation for collaborative acceleration adopts the FP16 / FP8 mixed-precision specification for large model inference, the coarse estimation of spectral radius (relative error tolerance > 10⁻³) adopts FP8 precision, and the fine estimation of spectral radius and the linear solution stage (relative error tolerance ≤ 10⁻³ and > 10⁻³) 6 FP16 precision is adopted, and the key residual update process (relative error tolerance ≤10⁻) 6 It uses FP64 precision.

6. The adaptation method according to claim 1, characterized in that, The large model inference acceleration framework is a general-purpose large model inference framework that supports tensor parallelism and pipelined parallelism (including but not limited to ByteDance's self-developed large model inference framework), is compatible with GPU / TPU heterogeneous computing architecture, and supports tensor parallelism and pipelined parallelism.

7. A matrix-free rigid equation solving adaptation system based on large model inference acceleration, characterized in that, The method according to any one of claims 1 to 6 comprises: The module adaptation unit is used to complete tensor rearrangement and operator fusion of the matrix-free rigid solution submodule, so as to achieve hardware logic compatibility with the large model inference acceleration framework. The resource scheduling unit is used to build a shared memory pool and to dynamically allocate and reclaim memory resources based on large model inference load and rigid solution computation load. The stiffness ratio prediction unit is used to call the large model feature extraction network, input stiffness equation parameters to obtain the stiffness ratio prediction value and transmit it to the spectral radius estimation unit; The collaborative solution unit includes a quantum annealing spectral radius estimation subunit, a mixed-precision calculation subunit, and an adaptive step size control subunit. Based on the adaptation results, resource scheduling, and prediction values, it completes the solution of matrix-free rigid equations. The interface unit is used to realize bidirectional data interaction between this system and the large model inference acceleration framework, and supports real-time feedback of solution results to the large model inference process.

8. The adaptation system according to claim 7, characterized in that, The system is deployed on a large model inference server and shares CPU / GPU / TPU computing resources with the large model inference framework. The system is lightweight and packaged as a whole, with a startup memory usage of ≤512MB and supports hot-swappable calls.