Search

CN-122021753-A - InfiDeck artificial intelligent operating system and its extended hardware equipment

CN122021753ACN 122021753 ACN122021753 ACN 122021753ACN-122021753-A

Abstract

The application discloses InfiDeck artificial intelligent operating system and its extended hardware equipment, the system includes power management module, AI calculation module, edge learning manager, privacy and data management module, reasoning engine and adapter management module, knowledge index and model warehouse module and communication interface module; the application realizes dynamic power consumption adjustment through SoC/temperature/idle degree ternary gating, adopts low-bit mixed precision operation processing, sets strategies such as a charging trigger continuous learning mechanism, a self-playback learning mechanism, a model compression self-adjusting system, privacy and energy consumption cooperative control and the like, provides a portable device capable of locally completing personalized AI training and reasoning, and solves the problems of energy consumption, temperature and privacy safety.

Inventors

  • LI ZHEN
  • Xie Congkai
  • YANG HONGXIA

Assignees

  • 无界智索(深圳)科技有限责任公司

Dates

Publication Date
20260512
Application Date
20260130

Claims (10)

  1. The InfiniDeck artificial intelligent operating system is characterized by comprising a power management module, an AI calculation module, an edge learning manager, a privacy and data management module, an reasoning engine and adapter management module, a knowledge index and model warehouse module and a communication interface module; The power management module is used for monitoring the state of charge (SoC), the temperature T and the Idle degree Idle of the battery and outputting a training gating signal and a power consumption limiting parameter to the edge learning manager; The edge learning manager is used for decomposing the training task into a plurality of batches and configuring the precision parameters and the scheduling strategies of the training task to the AI computing module, and the edge learning manager is also used for controlling the AI computing module to start, interrupt or continue executing the training task according to the training gating signal and the power consumption limiting parameter; the privacy and data management module is used for performing de-identification, sensitive content filtering and differential privacy disturbance processing on the training data to obtain safe training data, and storing the safe training data into a training sample pool; the AI calculation module is used for providing matrix operation capability to execute training tasks or assisting the main terminal in reasoning; the reasoning engine and the adapter management module are used for executing reasoning and setting personalized model increment parameter modules for different training tasks; The knowledge index and model warehouse module is used for storing version information of the management model and the model increment parameter module; The communication interface module is used for carrying out data interaction with the main terminal or the outside.
  2. 2. The InfiDeck artificial intelligence operating system of claim 1, wherein said power management module comprises a battery management chip, a charge management circuit, a power detection unit, and a temperature sensing unit, and is configured to: The battery management chip detects the state of charge SoC and the temperature T of the battery through the electric quantity detection unit and the temperature sensing unit, and obtains the Idle degree Idle through software driving; when all the conditions that the state of charge SoC is more than or equal to a first threshold value, the temperature T is less than or equal to a first temperature threshold value and the Idle degree Idle is more than or equal to a first Idle threshold value are detected to be met, a gating signal allowing training is generated and sent to an edge learning manager; otherwise, maintaining or switching to an 'reasoning only/standby' state without starting a training task; When detecting that the state of charge SoC is reduced to a second threshold value, or the temperature T exceeds the second temperature threshold value, or the main terminal initiates a high-priority reasoning task, or a user wake-up instruction is received, triggering the safety interrupt of the training process: Writing the current training state into a local nonvolatile storage to form a check point; preferentially processing reasoning tasks; On the next start of training, training is resumed from the most recent checkpoint.
  3. 3. The InfiDeck artificial intelligence operating system of claim 2, wherein the power management module and edge learning manager are configured to: the power management module detects the charging state of the battery, and when the charging state is in the state and the time that the Idle degree Idle is more than or equal to a first Idle threshold exceeds a set duration threshold, a night reinforcement learning signal is sent to the edge learning manager; The edge learning manager extracts typical scenes from the local knowledge index and the history interaction to generate difficult sample, countermeasure sample or composite multi-round dialogue; continuing to train the model increment parameter module in a self-distillation or contrast learning mode; And the training task is completed to evaluate and compress the model and the model increment parameter module.
  4. 4. The InfiDeck artificial intelligence operating system of claim 1, wherein the edge learning manager is deployed on a master processor of an AI computing module, including a task scheduling sub-module, a training control sub-module, and an evaluation and compression sub-module; The task scheduling submodule is used for applying the trained model and the model increment parameter module and switching according to a request scene of a user; the training control submodule is used for executing training to obtain a trained model and a model increment parameter module; the evaluation and compression submodule is used for evaluating the performance of the corresponding model increment parameter module by using a local verification set and an online A/B test index after a certain training task is completed, and if the threshold is met, the model compression and quantization flow is executed: Carrying out INT4/INT8 quantization or pruning on the model increment parameter module or part of the network layer; Assessing the reasoning delay and the energy consumption under different precision configurations; Configuring energy consumption management strategies according to user settings and power supply electric quantity The compressed model version and model increment parameter module is registered in the knowledge index and model warehouse module.
  5. 5. The InfiDeck artificial intelligence operating system of claim 1, wherein the privacy and data governance module is configured to: de-identifying, namely deleting or masking sensitive fields of a processing user; Filtering the sensitive content, namely filtering samples which are illegal or unsuitable for training according to a local sensitive word stock and a pre-training classifier; The differential privacy disturbance processing comprises introducing noise into a characteristic space or a gradient space; the processed samples are marked as safe training data and stored in a training sample pool.
  6. 6. The InfiDeck artificial intelligence operating system of claim 1, wherein the AI computation module includes a master processor, an NPU, an Edge GPU, and an on-chip cache and is configured to employ processing low-bit-mix precision operations including FP8, FP4, INT8, INT4, MXFP8, MXFP, NVFP4, and INT6.
  7. 7. The InfiDeck artificial intelligence operating system of claim 1, wherein the inference engine and adapter management module includes an inference engine and adapter management sub-module; The reasoning engine is integrated with a transducer reasoning engine, tokenizer and a Cache KV-Cache; The adapter management submodule adopts a MCP protocol tool calling mode and an Agent tool calling mode to autonomously load, unload and switch model increment parameter modules of different training tasks.
  8. 8. The InfiDeck artificial intelligence operating system of claim 1, wherein the knowledge index and model repository module stores versions of base models, model delta parameter modules, and their associated metadata including task type, training date, and performance metrics; And recording the pseudo sample and the difficult statistics information generated in training.
  9. 9. The InfiDeck artificial intelligence operating system of claim 1, further comprising a AIOS client and application interface module disposed on the host terminal device and communicatively coupled to the communication interface module for enabling data interaction, firmware upgrades.
  10. An extended hardware device of an infideck artificial intelligence operating system, wherein the extended hardware device is loaded with the InfiDeck artificial intelligence operating system of any one of claims 1-9, and further comprising a power supply, a microphone, a camera, a physical communication interface, a physical privacy control and indication module, and a storage and hierarchy storage module; The physical communication interface comprises a USB-C interface, a Lei Li interface, a Wi-Fi interface, a Bluetooth interface and/or a magnetic contact point interface; the physical privacy control and indication module comprises a physical deflector rod switch, an indicator lamp and a buzzer, and is arranged in such a way that when the physical deflector rod switch is in a closed state, the power supply of the expansion hardware equipment is disconnected, and only the battery charging and discharging functions are reserved; the storage and layering storage module comprises DRAM and nonvolatile flash memory.

Description

InfiDeck artificial intelligent operating system and its extended hardware equipment Technical Field The invention belongs to the technical field of AI training/reasoning integrated machines, and particularly relates to a InfiDeck artificial intelligent operating system and an expansion hardware device thereof. Background Large Model (LLM) reasoning is mainly performed at the cloud in exchange for high computational power and large memory, but brings privacy, time delay and cost pressure. Intelligent terminal vendors are turning to the form of "end-side prioritization+cloud assistance if necessary", e.g. Apple will execute smaller models entirely locally, while complex requests are processed on Apple self-grinding security clusters through its "Private Cloud Compute (PCC)". However, the above prior art has the following drawbacks and disadvantages: (1) The network dependence and delay are high, the reasoning and the generation need to depend on cloud computing, and the response is slow and the interaction is not smooth in a weak network environment; (2) Privacy and security risks that voice, text and behavior data of a user must be uploaded to a cloud end, and leakage risks exist; (3) The local AI has high energy consumption, can cause the problems of overhigh power consumption and temperature rise when a medium-large model is directly operated on the mobile terminal, and lacks autonomous management learning (stream Mode) in a charging window/idle window. The industry has 'idle+charging+wi-Fi' participation threshold in federal learning, but most focus is aggregated across devices, how to automatically identify suitable thermally-enabled windows, plan light training/distillation/playback sequences, and safely interrupt and resume when windows are closed inside a single device, and still lacks a general and engineering scheme. (4) The personalized adaptation is insufficient, the existing terminal AI model is fixed parameters and cannot be self-learned according to personal habits, and although AdapterHub/LoRA proves the feasibility of efficient and modularized parameters, how to low-stop hot-change personalized modules of different scenes/characters/applications when the terminal side runs, and the terminal side is distributed with a system-level model (such as Play for On-device AI) to control and roll back in a collaborative mode, so that standard solutions are still lacking. (5) The system integration level is low, AI calculation, data management and a power supply system are separated, and the system is complex to carry and low in efficiency. (6) There is no end-to-end low resource training reasoning scheme, such as AWQ/GPTQ/FP8, but how to adaptively select bit width/calibration strategy/mixed precision path according to load and thermal state at the end side, and dynamically balance quality-energy consumption-delay triangle, and the integrated compression self-adjusting framework is lacking. A system like 'LLM IN A FLASH' shows the possibility of flash memory-DRAM layering, but on a mobile phone/tablet/embedded type, combining with minimum read-write strategy of a training side, check point slicing and breakpoint training, and operating simultaneously with hot plug of an adapter still belongs to front-edge exploration. These problems limit the sustainable use and experience enhancement of AI on personal devices. Disclosure of Invention The invention mainly aims to overcome the defect that the existing mobile terminal lacks a portable device capable of locally completing personalized AI training and reasoning, and meanwhile cannot solve the problems of energy consumption, temperature and privacy safety, and provides a InfiDeck artificial intelligent operating system and an expansion hardware device thereof. The invention aims to solve the following problems in the prior art, namely 1) lack of a mechanism for automatically and continuously learning in a charging idle state, 2) lack of a modularized and multitask adaptive pluggable learning system, 3) lack of a self-adjusting algorithm system for automatically compressing and distilling after training is finished, and 4) lack of a scheduling strategy capable of dynamically controlling batch and precision on the basis of energy-calculation force balance. In order to achieve the above purpose, the present invention adopts the following technical scheme: The invention provides a InfiDeck artificial intelligent operating system, which comprises a power management module, an AI calculation module, an edge learning manager, a privacy and data management module, an reasoning engine and adapter management module, a knowledge index and model warehouse module and a communication interface module; The power management module is used for monitoring the state of charge (SoC), the temperature T and the Idle degree Idle of the battery and outputting a training gating signal and a power consumption limiting parameter to the edge learning manager; The edge learning manage