Intermediate storage device, computer system, and preprocessing method of computer command

Abstract

To provide an intermediate storage device capable of quickly calling a language model, a computer system and a preprocessing method for a computer command.SOLUTION: In the intermediate storage 300, when the inputted command 210 is the model manipulation command 331, the access controlling unit 340 forwards the model manipulation command 331 to the AI selecting unit 330 to make the selected language model 311 generate the output data 334, and makes the access controlling unit 340 generate the generation result 335 based on the output data 334, and when the inputted command 210 is the data-accessing command 332, the access controlling unit 340 accesses the corresponding document 321 from the data-storing unit 320 based on the data-accessing command 332.SELECTED DRAWING: Figure 2A

Description

[Technical Field]

[0001] The present invention relates to an intermediate storage device, a computer system, and a method for preprocessing computer commands, and more particularly to an intermediate storage device, a computer system, and a method for preprocessing computer commands that enable fast invocation of a language model. [Background technology]

[0002] The rise of artificial intelligence (AI) has brought many new possibilities to human society. Among these, the development of large language models (LLMs) has laid the groundwork for AI technology. Summary of the Invention [Problem to be solved by the invention]

[0003] LLM is an AI model that can process and generate natural language, and its training requires a large amount of data and computing resources, so the LLM record issue has also become a significant challenge for the development of AI technology. The current method involves transmitting interactive information between LLMs and users over the Internet. Although the information transmission process can be encrypted, there is still a risk of personal information or trade secrets being leaked. [Means for solving the problem]

[0004] In view of the above problems, the present invention has the following configuration. That is, in an intermediate storage device comprising a model storage unit, a data storage unit, an AI selection unit, and an access control unit, the model storage unit stores a plurality of language models, the data storage unit stores a plurality of documents, the AI ​​selection unit selects the language model based on a model operation command and has the selected language model execute the model operation command to generate output data, the access control unit is connected to the data storage unit and the model storage unit, the access control unit receives an input command, and the access control unit determines whether the input command is the model operation command or a data access command, when the input command is the model operation command, the access control unit transfers the model operation command to the AI ​​selection unit, causes the selected language model to generate the output data, and has the access control unit generate a generation result based on the output data, and when the input command is the data access command, the access control unit accesses the corresponding document from the data storage unit based on the data access command.

[0005] Moreover, the AI ​​selection unit further includes a real-time operating system.

[0006] The access control unit also transmits the generated result or the selected document to a higher-level operating system.

[0007] The access control unit may further include a transmission interface, and the access control unit may be connected to the transmission interface. The type of the transmission interface may be an Advanced Technology Attachment, a Serial Attack Attachment, a Universal Serial Bus, a Peripheral Component Interconnect Expansion Interface, or a Non-Volatile Memory Express.

[0008] Further, in an intermediate storage device having a plurality of model storage units, a data storage unit, an AI selection unit, and an access control unit, each of the plurality of model storage units stores a language model, and the data storage unit stores a plurality of documents, the AI ​​selection unit is connected to the model storage unit, and a model operation command of the AI ​​selection unit selects the language model, and the language model executes the model operation command to generate output data, the access control unit is connected to the data storage unit and the AI ​​selection unit, and the access control unit receives an input command and determines whether the input command is the model operation command or a data access command, and when the input command is the model operation command, the access control unit transfers the model operation command to the AI ​​selection unit, causes the selected language model to generate the output data, and causes the access control unit to generate a generation result based on the output data, and when the input command is the data access command, the access control unit accesses the corresponding document from the data storage unit based on the data access command.

[0009] Moreover, the AI ​​selection unit further includes a real-time operating system.

[0010] The access control unit also transmits the generated result or the selected document to a higher-level operating system.

[0011] The access control unit may further include a transmission interface, and the access control unit may be connected to the transmission interface. The type of the transmission interface may be an Advanced Technology Attachment, a Serial Attack Attachment, a Universal Serial Bus, a Peripheral Component Interconnect Expansion Interface, or a Non-Volatile Memory Express.

[0012] Also, in a computer system including a processor, an intermediate storage device, and a host operating system, the processor executes the host operating system, the processor receives input commands via the host operating system, the intermediate storage device is connected to the processor, and the intermediate storage device includes at least one model storage unit, a data storage unit, an AI selection unit, and an access control unit, the access control unit is connected to the data storage unit, the AI ​​selection unit, and each of the model storage units, each of the model storage units stores a language model, the data storage unit stores a plurality of documents, and the host operating system executes the input commands via the host operating system. The system sends the input command to the access control unit, and the access control unit determines whether the input command is a model operation command or a data access command. When the input command is the model operation command, the access control unit transfers the model operation command to the AI ​​selection unit, causes the selected language model to generate the output data, and causes the access control unit to generate a generation result based on the output data. When the input command is the data access command, the access control unit accesses the corresponding document from the data storage unit based on the data access command.

[0013] It also receives an input command from an access control unit of an intermediate storage device, and the access control unit determines whether the input command is a model manipulation command or a data access command. When the input command is the model manipulation command, the access control unit transfers the model manipulation command to an AI selection unit. The AI ​​selection unit selects one of a plurality of language models based on the model manipulation command, the selected language model is a selected model, and the selected model generates output data based on the model manipulation command. The access control unit generates a generation result based on the output data.

[0014] The step of the AI ​​selection unit selecting one of the plurality of language models based on the model operation command, and the selected language model being the selected model, includes the AI ​​selection unit selecting one of a plurality of model storage units, and each of the model storage units storing the corresponding language model.

[0015] Furthermore, the step of the AI ​​selection unit selecting one of the plurality of language models based on the model operation command, and the selected language model being the selected model, includes storing the plurality of language models in a model storage unit, and the AI ​​selection unit selecting one of the language models from the model storage unit.

[0016] Also, the input command is the data access command, and the access control unit accesses the corresponding document from the data storage unit based on the data access command. [Effects of the Invention]

[0017] The intermediate storage device, computer system, and computer command preprocessing method according to the present invention provide a large-scale language model that can be used on a local computer. This not only avoids the risk of information leakage during communication over a network, but also speeds up the invocation of the large-scale language model and shortens information response times. Furthermore, since large-scale language models corresponding to various application scenarios can be deployed on a computer (hereinafter also referred to as "computer"), services related to multiple different application scenarios can be provided on the same computer. [Brief explanation of the drawings]

[0018] [Figure 1] FIG. 1 is a block diagram of hardware for explaining an embodiment of the present invention. [Figure 2A] 1 is a block diagram of a system applied to a computer according to an embodiment of the present invention. [Figure 2B]FIG. 10 is a block diagram of loading a selected model into a data storage unit according to the present embodiment. [Figure 3] 1 is a flowchart of a method for preprocessing computer commands according to an embodiment; [Figure 4] FIG. 2 is a block diagram of an AI selection unit according to the present embodiment. [Figure 5] FIG. 10 is a block diagram of a system applied to a computer according to another embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0019] First, an embodiment of the present invention will be described with reference to Figures 1 and 2A. Figure 1 is a block diagram of hardware for explaining an embodiment of the present invention, and Figure 2A is a block diagram of a system applied to a computer according to an embodiment of the present invention.

[0020] The intermediate storage device 300 is connected to the processor 100 and receives input commands 210 output by the processor 100. The computer 10 can output the input commands 210 via the input device 200, and the processor 100 can also output the input commands 210 to the intermediate storage device 300 when executing an application program. The computer 10 may be an electronic device such as a personal computer, a notebook computer, a tablet computer, a mobile phone, etc. The computer 10 executes a corresponding upper operating system (OS) 322. The upper operating system 322 may be, for example, the Microsoft® Windows® operating system, the Apple® OSX® operating system, or the Linux® operating system.

[0021] The intermediate storage device 300 is compatible with a variety of interfaces, including Serial Advanced Technology Attachment (hereinafter also referred to as "SATA"), Serial Attached SCSI (hereinafter also referred to as "SAS"), Peripheral Component Interconnect Express (hereinafter also referred to as "PCIe"), Non-Volatile Memory Express (hereinafter also referred to as "NVMe"), Universal Serial Bus (hereinafter also referred to as "USB"), External Serial Advanced Technology Attachment (hereinafter also referred to as "eSATA"), Small Computer System Interface (hereinafter also referred to as "SCSI"), Integrated Drive Electronics (hereinafter also referred to as "IDE"), Next Generation Form Factor (hereinafter also referred to as "M.2"), Universal Serial Bus (hereinafter also referred to as "IDE"). Bus, hereinafter simply referred to as "USB.") or Thunderbolt. The intermediate storage device 300 includes a model storage unit 310, a data storage unit 320, an artificial intelligence selection unit (hereinafter referred to as AI selection unit 330), and an access control unit 340. The access control unit 340 is connected to the model storage unit 310, the data storage unit 320, and the AI ​​selection unit 330.

[0022] The model storage unit 310 may be stored in a non-volatile memory (NVM) or other storage medium, and stores a plurality of large language models 311 (hereinafter also referred to simply as "LLMs"). The types of language models 311 include a language model for conversational applications (Language Model for Dialogue Applications, hereinafter also referred to as "LaMDA") 311, a large language model (Large Language Model Meta AI, hereinafter also referred to as "LLaMA") 311, GPT-3 (registered trademark) (Generative Pre-trained Transformer 3, hereinafter also referred to as "GPT-3"), Google's (Google (registered trademark)) BERT (Bidirectional Encoder Representations from Transformers, hereinafter also referred to as "BERT"), GPT-4 (registered trademark) (Generative Pre-trained Transformer 4, hereinafter also referred to as "GPT-4"), Codex, DALL·E (OpenAI DALL-E), BART (Bidirectional and Auto-Regressive Transformers, hereinafter also referred to as "BART"), and RoBERTa (Robustly optimized BERT). The approach may be, but is not limited to, T5 (Text-to-Text Transfer Transformer, hereinafter also referred to as "T5"), or PaLM (Pathways Language Model, hereinafter also referred to as "PaLM"). Based on the above-mentioned language model 311, the model storage unit 310 can develop language models 311 with different requirements based on the requirements of different scenarios, for example, LLaMA can be applied to an English teaching scenario, and DALL-E can be applied to a drawing teaching scenario. The various language models 311 mentioned above can also be applied to fields such as personal chat scenarios, personal knowledge base management, document organization, document translation, healthcare, financial investment, manufacturing, retail, transportation or agriculture.

[0023] The data storage unit 320 stores a plurality of documents 321, which generally refer to text files, image files, video files, audio files, database files, compressed files, presentation files, program code files, etc. In some embodiments, the upper operating system 322 may be stored in the intermediate storage device 300 or other storage device. The upper operating system 322 causes the processor 100 to output the input commands 210 to the intermediate storage device 300 based on the input commands 210 generated by a user or by an application program. Among these, the types of input commands 210 can be further divided into data access commands 332 and model operation commands 331. When the intermediate storage device 300 receives the input command 210, it determines whether the input command 210 is a model operation command 331 or a data access command 332.

[0024] Data access commands 332 may include related operations such as reading, writing, appending, updating, deleting, moving, copying, searching, locking, etc. of documents 321 .

[0025] The model operation commands 331 include a chat interaction command (AI Language Model Chat, AI LLM CHAT), a model configuration command (AI Language Model Configuration Model, AI LLM CONFIG MODEL), a model fine-tuning command (AI Language Model Fine Tune, AI LLM FINE TUNE), a language model 311 authorization command (AI Language Model Authorization, AI LLM AUTH), a model deployment command (AI Language Model Deploy Model, AI LLM DEPLOY MODEL), a dataset load command (AI Language Model Load Dataset, AI LLM LOAD DATASET), a text generation command (AI Language Model Generate Text, AI LLM GENERATE TEXT), a performance evaluation command (AI Language Model Evaluate Performance, AI LLM EVALUATE PERFORMANCE), a parameter update command (AI Language Model Update Parameters, AI LLM UPDATE PARAMETERS), a model save command (AI Language Model Save Model, AI LLM SAVE MODEL), a model test command (AI Language Model Test Model, AI LLM TEST MODEL), and a model reset command (AI Language Model Reset These include instructions for setting permissions (AI Language Model Set Permissions, AI LLM SET PERMISSIONS), model training instructions (AI Language Model Train Model, AI LLM TRAIN MODEL), and so on.

[0026] After the access control unit 340 receives the input command 210, the access control unit 340 performs the following steps according to the type of the input command 210, as shown in Figure 3. Here, Figure 3 is a flowchart of the computer command pre-processing method according to this embodiment.

[0027] Step S310: Receive the input command 210 from the access control unit 340 of the intermediate storage device 300. Step S320: The access control unit 340 determines whether the input command 210 is a model operation command 331 or a data access command 332. Step 330 : When the input command 210 is a model operation command 331 , the access control unit 340 sends the model operation command 331 to the AI ​​selection unit 330 . Step S340: The AI ​​selection unit 330 selects one of the groups of language models 311 according to the model operation command 331, and the selected language model 311 becomes the selected model 312. Step S350: The selection model 312 generates the output data 334 based on the model operation command 331. Step S360: The access control unit 340 generates a generation result based on the output data 334. Step S370: When the input command 210 is a data access command 332, the access control unit 340 sends the data access command 332 to the data storage unit 310.

[0028] First, the computer 10 is enabled and booted, and the computer 10 loads the upper operating system 322. The upper operating system 322 receives the input command 210. Generally, a user can output an input command 210 through an input device 200, for example, a user can enter an input command 210 via a keyboard, or a user can form an input command 210 by clicking on an associated option with a mouse. After the upper operating system 322 receives the input command 210, the upper operating system 322 sends the input command 210 to the corresponding intermediate storage device 300 (corresponding to step S310). Here, there is at least one intermediate storage device 300, but in FIG. 1, one intermediate storage device 300 is taken as an example for explanation.

[0029] The access control unit 340 of the intermediate storage device 300 determines that the input command 210 is a data access command 332 or a model operation command 331 (corresponding to step S320). When the access control unit 340 determines that the input command 210 is the model operation command 331, the access control unit 340 transmits the model operation command 331 to the AI ​​selection unit 330 (corresponding to step S330). Next, the AI ​​selection unit 330 selects from the plurality of language models 311 in the model storage unit 310, and the selected language model 311 is referred to as a selected model 312 (corresponding to step S340).

[0030] In this embodiment, the model storage unit 310 has multiple language models 311. Before selecting a language model 311, the AI ​​selection unit 330 determines the corresponding application scenario based on the content of the model operation command 331. The AI ​​selection unit 330 selects the corresponding language model 311 based on the selected application scenario. In Figure 2A, the selected model 312 is shown by a thick black frame as an example. In this embodiment, as shown in Figure 2B, the access control unit 340 can load the selection model 312 into the data storage unit 320. The selection model 312 can execute the model operation command 331 and generate output data 334.

[0031] The data storage unit 320 sends the output data 334 to the access control unit 340 , which generates the generated result 335 . In general, the access control unit 340 encapsulates the output data 334 in an API (Application Programming Interface) interface format compatible with the upper application or operating system. The API interface format is as follows: {”prompt”:”content”,”max_tokens”:60,”temperature”:0.5} The AI ​​selection unit 330 sends a model operation command 331 to the selection model 312. The selection model 312 generates corresponding output data 334 based on the model operation command 331 (corresponding to step S350). For example, when the model operation command 331 is "Translate the following English paragraph into traditional Chinese", the AI ​​selection unit 330 may select the aforementioned LLaMA model as the language model 311 for translation and training.

[0032] The LLaMA model executes the model operation command 331 and generates the translated output data 334: "Translate the following English paragraph into Traditional Chinese."

[0033] The access control unit 340 sends the generated result 335 to the upper operating system 322. Furthermore, the AI ​​selection unit 330 can also invoke a related language model 311 based on the investment-related input command 210, such as GPT-3, GPT-4, or LaMDA. When the input command 210 is a data access command 332, the access control unit 340 sends the data access command 332 to the data storage unit 320 (corresponding to step S370). The access control unit 340 accesses the corresponding document 321 from the data storage unit 320. The intermediate storage device 300 sends the selected file information to the upper operating system 322.

[0034] In some embodiments, the AI ​​selection unit 330 further includes a real-time operating system (RTOS) 333, as specifically shown in FIG. 4. Here, FIG. 4 is a block diagram of the AI ​​selection unit according to this embodiment. The real-time operating system 333 provides task management, resource management, communication or information security processing for the intermediate storage device 300. The real-time operating system 333 sends output data 334 to the access control unit 340 and causes the access control unit 340 to generate a generated result 335.

[0035] The intermediate storage device 300 further includes a transmission interface, the details of which are shown in Figure 4. The access control unit 340 is connected to the transmission interface, which is also connected to the computer 100. Types of transmission interfaces include Advanced Technology Attachment (hereinafter referred to as "ATA"), Serial AT Attachment (hereinafter referred to as "SATA"), Universal Serial Bus (hereinafter referred to as "USB"), Peripheral Component Interconnect Express (hereinafter referred to as "PCIe"), and Non-Volatile Memory Express (hereinafter referred to as "NVMe"). In some embodiments, the computer 10 includes a processor 100 and an intermediate storage device 300. Please refer to Fig. 5, which is a block diagram of a system applied to a computer according to another embodiment of the present invention. The intermediate storage device 300 of this embodiment includes a plurality of model storage units 310, a data storage unit 320, an AI selection unit 330, and an access control unit 340.

[0036] The access control unit 340 is connected to the data storage unit 320 and the AI ​​selection unit 330. Each model storage unit 310 stores its own language model 311. The types of language models 311 are as described above. Alternatively, the same language model 311 used in different application scenarios may be stored in each model storage unit 310. For example, two sets of LLaMA models are applied to English language teaching and Python programming language teaching, so the English language teaching LLaMA model can be stored in a separate model storage unit 310, and the Python programming language teaching LLaMA model can be stored in another model storage unit 310.

[0037] The access control unit 340 determines whether the received input command 210 is a model manipulation command 331 or a data access command 332 . When the input command 210 is a data access command 332 , the access control unit 340 accesses the corresponding document 321 from the data storage unit 320 based on the data access command 332 , and then returns the selected document 321 to the upper operating system 322 . When the data access command 332 is a model operation command 331 , the access control unit 340 sends the model operation command 331 to the AI ​​selection unit 330 . The AI ​​selection unit 330 selects the corresponding model storage unit 310 and language model 311 based on the application scenario of the model operation command 331. The selected model 312 executes the model operation command 331 and generates output data 334. The access control unit 340 sends the generated result 335 to the upper operating system 322.

[0038] The intermediate storage device 300, the computer 10 system, and the computer 10 command pre-processing method provide a large language model 311 that can be used by the local computer 10. In addition to avoiding the risk of information leakage during network communication, the calling speed of the large-scale language model 311 can be increased, thereby accelerating the information response time. Furthermore, large-scale language models 311 for different application scenarios can be deployed on the computer 10, so that the same computer 10 can provide related services for multiple different application scenarios. [Explanation of symbols]

[0039] 10 Calculator 100 processor 200 input devices 210 Input Commands 300 Intermediate storage device 310 Model Storage Unit 311 Language Models 312 Selection Model 320 Data Storage Unit 321 documents 322 Top Operating Systems 330 AI Selection Unit 331 Model Operation Commands 332 Data Access Commands 333 Real-Time Operating Systems 334 Output Data 335 Generated results 340 Access Control Unit S310, S320, S330, S340, S350, S360, S370 Step

Claims

1. An intermediate storage device comprising a model storage unit, a data storage unit, an AI selection unit, and an access control unit, the model storage unit stores a plurality of language models; the data storage unit stores a plurality of documents; The AI ​​selection unit selects the language model based on a model operation command, and causes the selected language model to execute the model operation command to generate output data; the access control unit is connected to the data storage unit and the model storage unit, the access control unit receives an input command, and the access control unit determines whether the input command is the model operation command or a data access command; When the input command is the model operation command, the access control unit transfers the model operation command to the AI ​​selection unit, causes the selected language model to generate the output data, and causes the access control unit to generate a generation result based on the output data; When the input command is the data access command, the access control unit accesses the corresponding document from the data storage unit based on the data access command. Intermediate storage device.

2. The AI ​​selection unit further includes a real-time operating system.

2. The intermediate storage device according to claim 1,

3. The access control unit transmits the generated result or the selected document to a higher-level operating system.

2. The intermediate storage device according to claim 1,

4. Further comprising a transmission interface, the access control unit is connected to the transmission interface, and the type of the transmission interface is an advanced technology attachment, a serial attenuation attachment, a universal serial bus, a peripheral component interconnect expansion interface, or a non-volatile memory express.

2. The intermediate storage device according to claim 1,

5. An intermediate storage device comprising a plurality of model storage units, a data storage unit, an AI selection unit, and an access control unit, the plurality of model storage units, each of which stores a language model; the data storage unit stores a plurality of documents; The AI ​​selection unit is connected to the model storage unit, and a model operation command of the AI ​​selection unit selects the language model, and the language model executes the model operation command to generate output data; the access control unit is connected to the data storage unit and the AI ​​selection unit, and the access control unit receives an input command; the access control unit determines whether the input command is the model operation command or a data access command; When the input command is the model operation command, the access control unit transfers the model operation command to the AI ​​selection unit; causing the selected language model to generate the output data, and causing the access control unit to generate a generation result based on the output data; When the input command is the data access command, the access control unit accesses the corresponding document from the data storage unit based on the data access command. Intermediate storage device.

6. The AI ​​selection unit further includes a real-time operating system.

6. The intermediate storage device according to claim 5,

7. The access control unit transmits the generated result or the selected document to a higher-level operating system.

6. The intermediate storage device according to claim 5,

8. Further comprising a transmission interface, the access control unit is connected to the transmission interface, and the type of the transmission interface is an advanced technology attachment, a serial attenuation attachment, a universal serial bus, a peripheral component interconnect expansion interface, or a non-volatile memory express.

6. The intermediate storage device according to claim 5,

9. In a computer system including a processor, an intermediate storage device, and a host operating system, the processor executes a higher-level operating system, and the processor receives input commands via the higher-level operating system; the intermediate storage device is connected to the processor, and the intermediate storage device comprises at least one model storage unit, a data storage unit, an AI selection unit, and an access control unit; the access control unit is connected to the data storage unit, the AI ​​selection unit, and each of the model storage units; each said model storage unit stores a language model and said data storage unit stores a plurality of documents; the upper operating system sends the input command to the access control unit; The access control unit determines whether the input command is a model operation command or a data access command; When the input command is the model operation command, the access control unit transfers the model operation command to the AI ​​selection unit, causes the selected language model to generate output data, and causes the access control unit to generate a generation result based on the output data; When the input command is the data access command, the access control unit accesses the corresponding document from the data storage unit based on the data access command. Computer system.

10. receiving an input command from an access control unit of the intermediate storage device; The access control unit determines whether the input command is a model operation command or a data access command; When the input command is the model operation command, the access control unit transfers the model operation command to an AI selection unit; the AI ​​selection unit selects one of a plurality of language models based on the model operation command, and the selected language model is a selection model; the selected model generates output data based on the model manipulation commands; the access control unit generates a generation result based on the output data. How to preprocess computer commands.

11. The step of the AI ​​selection unit selecting one of the plurality of language models based on the model operation command, and the selected language model being the selected model, includes: The AI ​​selection unit selects one of a plurality of model storage units, and each of the model storage units stores a corresponding one of the language models.

11. The method of claim 10, wherein the computer command is preprocessed.

12. The step of the AI ​​selection unit selecting one of the plurality of language models based on the model operation command, and the selected language model being the selected model, includes: storing the plurality of language models in a model storage unit; The AI ​​selection unit selects one of the language models from the model storage unit.

11. The method of claim 10, wherein the computer command is preprocessed.

13. The input command is the data access command, and the access control unit accesses a corresponding document from a data storage unit based on the data access command.

11. The method of claim 10, wherein the computer command is preprocessed.

Images