Transmission-oriented large model compression and acceleration method and system

Abstract

The application discloses a large model compression and acceleration method and system for transmission, relates to the technical field of large model transmission, and comprises the following steps: before and during model transmission, a plurality of groups of probe data packets are sent between a sending end and a receiving end, quality characteristic parameters in a network transmission channel are continuously collected and analyzed, five quality characteristic parameters, i.e., available bandwidth, transmission delay, delay jitter, packet loss rate and transmission stability level, are obtained, the throughput is improved when the network is improved, the retransmission cost is reduced when the network is deteriorated, concurrency is improved to improve the throughput, when the network is deteriorated, the data block can be reduced, the concurrency can be reduced, the retransmission cost can be reduced by combining the timeout retransmission and the temporary skipping mechanism, and the transmission can be prevented from being stuck, so that more stable and efficient large model transmission can be realized under a complex network environment, and finally the receiving end can quickly obtain a complete model file which can be directly deployed within a controllable precision loss range.

Description

Technical Field

[0001] This invention belongs to the field of large model transmission technology, specifically a method and system for large model compression and acceleration based on transmission. Background Technology

[0002] Large-scale language models are a core infrastructure in the field of artificial intelligence. Their parameter scale has grown from hundreds of millions in the early days to tens or even hundreds of billions, and the corresponding model file size has grown from several gigabytes to hundreds of gigabytes. In application scenarios such as cloud deployment, edge computing, and model distribution, it is necessary to transmit the model from the server to the client or edge device via the network.

[0003] However, existing model compression methods typically employ a uniform compression strategy for the entire model, using the same number of quantization bits or the same pruning ratio for all parameter layers. This completely ignores the fact that different parameter layers within the model have significantly different impacts on the inference results. Key parameters such as the embedding layer and output layer have a much greater impact on model performance than some hidden layers in the middle. Using a uniform compression rate can lead to over-compression of key parameters, resulting in accuracy loss, or under-compression of minor parameters, resulting in wasted transmission resources. At the same time, it completely ignores the actual quality of the network transmission channel. In real network environments, channel quality parameters such as bandwidth, latency, and packet loss rate change dynamically with time and network load. Fixed compression strategies cannot adapt to the dynamic changes in network conditions, leading to either excessively long transmission times that exceed acceptable limits. Based on this, a transmission-oriented large model compression and acceleration method and system are proposed. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for large-scale compression and acceleration for transmission, in order to solve the problems mentioned in the background art.

[0005] Large-scale compression and acceleration methods for transmission include:

[0006] Step 1: Before and during model transmission, send multiple sets of probe data packets between the sending and receiving ends to continuously collect and analyze the quality characteristic parameters in the network transmission channel. Obtain five quality characteristic parameters: available bandwidth, transmission delay, delay jitter, packet loss rate, and transmission stability level. Summarize the five quality characteristic parameters to form a channel quality characteristic set. Step 2: Divide the parameters of the large model to be transmitted into several parameter layers according to the network structure, evaluate the parameter influence of each parameter layer, and divide the parameter layers into three levels: key layer, important layer and secondary layer according to the influence score, and establish a parameter influence hierarchical map. Step 3: Combine the channel quality feature set with the parameter influence layer map for analysis, calculate the overall compression target rate based on available bandwidth and expected transmission time, and calculate the corresponding coding precision quota for each parameter layer group based on parameter level and transmission stability level; Step 4: Perform differentiated linear quantization coding on each parameter layer group according to the parameter layer coding precision quota, determine the data block size according to the channel quality feature set, and divide and package the coded data into blocks according to the hierarchical priority order of key layer, important layer, and secondary layer to form a data block sequence and transmission list file; Step 5: Transmit the transmission manifest file and data block sequence to the receiving end in sequence number. During the transmission process, update the channel quality feature set every 30 seconds and dynamically adjust the data block size and the number of concurrent transmissions according to the bandwidth change ratio. Step 6: The receiving end performs dequantization decoding on the received data blocks, writes the restored parameters into the parameter storage area of ​​the corresponding parameter layer group according to the position index, and performs model parameter reassembly after all data blocks are dequantized to generate a complete model file and save it to the local storage of the receiving end.

[0007] As a further aspect of the present invention, the specific method for obtaining available bandwidth, transmission delay, delay jitter, and packet loss rate is as follows: Ten sets of test data packets are sent, and the transmission time of each set is recorded. The instantaneous bandwidth is obtained by dividing the size of each data packet by the corresponding transmission time, and the average of the ten instantaneous bandwidths is taken as the available bandwidth. Twenty test data packets with timestamps are sent. After receiving each test data packet, the receiving end immediately returns an acknowledgment data packet to the sending end. The sending end divides the round-trip time by 2 to obtain the one-way transmission delay. The 20 one-way transmission delays are sorted from smallest to largest, and the values ​​of the 10th and 11th positions are added together and divided by 2 to obtain the transmission delay. The median delay is calculated by summing the delay values ​​of 20 single-pass transmissions and dividing by 20 to obtain the mean delay. The absolute value of the difference between each measurement and the mean delay is squared, summed, and then divided by 20. The square root of the result is taken to obtain the standard deviation of the delay, which is used as the delay jitter. 100 test data packets are continuously sent from the sending end to the receiving end. The receiving end counts the number of data packets actually received within the waiting time corresponding to 5 times the median transmission delay. The number of lost data packets is divided by 100 and multiplied by 100% to obtain the packet loss rate. As a further aspect of the present invention, the specific method for obtaining the transmission stability level is as follows: If the latency jitter is less than 5 milliseconds and the packet loss rate is less than 1%, it is determined to be a stable channel; if the latency jitter is between 5 and 20 milliseconds or the packet loss rate is between 1% and 5%, it is determined to be a general channel; if the latency jitter is greater than 20 milliseconds or the packet loss rate is greater than 5%, it is determined to be an unstable channel. During the model transmission process, the above measurement and calculation process is repeated every 30 seconds to update the channel quality feature set.

[0008] As a further aspect of the present invention, the specific method for evaluating the parameter influence of each parameter group is as follows: Prepare a standard test dataset containing classification task samples and generation task samples. Add random perturbations to the original model for each parameter layer group, with the perturbation magnitude being 10% of the original parameter values ​​for that layer group. Use the perturbated model to perform inference on the test samples. For classification task samples, count the number of samples whose predicted categories changed and divide by the total number of classification task samples to obtain the classification change rate. For generation task samples, calculate the change in edit distance between the generated text and the original text, divide by the length of the original text, and take the average of all generation task samples to obtain the generation change rate. Use the average of the classification change rate and the generation change rate as the influence score for the corresponding parameter layer group. Parameter layer groups with an influence score greater than or equal to 0.3 are designated as key layers, those with an influence score greater than or equal to 0.1 and less than 0.3 are designated as important layers, and those with an influence score less than 0.1 are designated as minor layers.

[0009] As a further aspect of the present invention, the specific method for calculating the corresponding coding precision quota for each parameter layer group based on the parameter level and channel stability level is as follows: The compression target rate is obtained by dividing the total size of the original model parameters by the product of the available bandwidth and the expected transmission time. If the calculated result is less than 1, the compression target rate is set to 1; if the calculated result is greater than or equal to 1, the calculated result is used as the compression target rate. Based on the parameter hierarchy, encoding precision baseline values ​​of 16 bits, 8 bits, and 4 bits are set for the critical layer, important layer, and minor layer, respectively. The encoding precision baseline values ​​are adjusted according to the transmission stability level. When the transmission stability level is stable, the baseline values ​​for each layer remain unchanged. When the transmission stability level is moderate, the critical layer is adjusted to 12 bits (16 bits multiplied by 0.75), the important layer to 4 bits (8 bits multiplied by 0.5), and the minor layer to 2 bits (4 bits multiplied by 0.5). When the transmission stability level is unstable, the critical layer is adjusted to 8 bits (16 bits multiplied by 0.5), the important layer to 2 bits (8 bits multiplied by 0.25), and the minor layer to 1 bit. Multiply the number of parameters in each parameter layer group by the adjusted coding precision corresponding to that layer group, and sum the results for all parameter layer groups to obtain the total number of compressed bits. Divide the total number of compressed bits by 8 to obtain the total number of compressed bytes. Divide the total size of the original model parameters by the size of the compressed model to obtain the actual compression ratio. If the actual compression ratio is less than the compression target ratio, gradually reduce the coding precision of each layer in the order of secondary layers first, important layers second, and critical layers last. Each reduction is equal to the current coding precision divided by 2 and rounded down, with a reduction of at least 1 bit. Recalculate the actual compression ratio and repeat the above process until the actual compression ratio meets the compression target ratio.

[0010] As a further aspect of the present invention, the specific method for performing differentiated linear quantization encoding on each parameter layer group is as follows: Scan all parameter values ​​within the parameter layer group to find the maximum and minimum values. Subtract the minimum value from the maximum value to obtain the parameter value range. Calculate the total number of quantization levels by raising the power of the number of bits corresponding to the encoding precision to base 2. Subtract 1 from the total number of quantization levels to obtain the quantization level. Divide the parameter value range by the quantization level to obtain the quantization interval. Subtract the minimum value of the layer group from each parameter value in the layer group, divide by the quantization interval, and round to obtain the quantization index. Encode and store the quantization index in binary using the corresponding number of bits for encoding precision. Simultaneously, record the maximum value, minimum value, and quantization interval of the parameters as quantization parameters for the receiver to use for dequantization.

[0011] As a further aspect of the present invention: the specific method for determining the data block size based on the channel quality feature set is as follows: When the transmission stability level is stable and the transmission latency is less than 50 milliseconds, the size of each data block is set to 5 megabytes. When the transmission stability level is moderate or the transmission latency is between 50 and 100 milliseconds, the size of each data block is set to 2 megabytes. When the transmission stability level is unstable or the transmission latency is greater than 100 milliseconds, the size of each data block is set to 500 kilobytes. Each data block contains data block header information, data block body, and data block check information. The data block header information includes the data block sequence number, the identifier of the parameter layer group to which it belongs, the layer classification, the start and end position indices of the parameters in the corresponding layer group, and the quantization parameters. The data block body is the binary encoded data of the quantization indices of all parameters in the block. The data block check information is the check code calculated by the cyclic redundancy check algorithm on the data block body.

[0012] As a further aspect of the present invention: the specific method for dynamically adjusting the data block size and the number of concurrent transmissions based on the bandwidth change ratio is as follows: The bandwidth ratio is obtained by dividing the current available bandwidth by the initial available bandwidth recorded at the start of transmission. If the bandwidth ratio is greater than 1.2, the network is considered improved. The current data block size is multiplied by 2.5 and rounded up to a value not exceeding 5 megabytes to obtain the adjusted data block size. Multiple unsent original data blocks are merged and repackaged, and the concurrent transmission quantity is adjusted from 1 to 3 at a time. If the bandwidth ratio is less than 0.8, the network is considered deteriorated. The current data block size is multiplied by 0.25 and rounded down to a value not less than 500 kilobytes to obtain the adjusted data block size. The unsent data blocks are split and repackaged, and the concurrent transmission quantity is adjusted to 1 at a time. For data blocks that have been sent but have not received an acknowledgment message after the median transmission delay multiplied by 5, they are considered to have timed out and are actively retransmitted. Data blocks that fail to be retransmitted after more than 3 retransmissions are marked as temporarily skipped, and subsequent critical layer data blocks are prioritized for transmission. Once the network condition is restored, the skipped data blocks will be retransmitted.

[0013] As a further aspect of the present invention: the specific method by which the receiving end performs inverse quantization decoding on the received data block is as follows: The quantization parameters, including the maximum and minimum values ​​and the quantization interval, are read from the header information of the data block. The binary encoded data stored in the main body of the data block is truncated sequentially according to the number of bits corresponding to the encoding precision recorded in the header of the data block. Each segment of binary encoding is restored to the corresponding quantization index integer, resulting in a sequence of quantization index integers. Each quantization index integer is multiplied by the quantization interval and then added to the minimum value of the parameter to obtain the restored floating-point parameter value. The restored floating-point parameter value is written to the parameter storage area of ​​the corresponding parameter layer group according to the start and end indexes recorded in the header of the data block. After all data blocks have been dequantized, the restored parameter values ​​are read sequentially from the parameter storage area of ​​each parameter layer group according to the network structure order of the original model and filled into the corresponding positions in the model structure. The number of parameters in each parameter layer group is checked to see if it is consistent with the original model definition. If it is inconsistent, an exception report message is sent to the sending end. After all parameters are filled, a complete model file is generated and saved to the local storage of the receiving end.

[0014] A transmission-oriented large-model compression and acceleration system, implemented using transmission-oriented large-model compression and acceleration methods, including: The channel quality acquisition module is used to send multiple sets of probe data packets between the sending end and the receiving end before and during model transmission, continuously collect the quality characteristic parameters of the network transmission channel, obtain available bandwidth, transmission delay, delay jitter, packet loss rate and transmission stability level, summarize them to form a channel quality feature set, and update the channel quality features every 30 seconds during transmission. The parameter influence assessment module is used to divide the parameters of the large model to be transmitted into several parameter layer groups according to the network structure, assess the parameter influence of each parameter layer group, divide the parameter layer group into key layer, important layer and secondary layer according to the influence score, and establish a parameter influence hierarchical map. The coding precision quota calculation module is used to combine and analyze the channel quality feature set with the parameter influence hierarchical map, calculate the overall compression target rate based on the available bandwidth and expected transmission time, and calculate the corresponding coding precision quota for each parameter layer group based on the parameter level and transmission stability level. The quantization encoding and packaging module is used to perform differentiated linear quantization encoding on each parameter layer group according to the encoding precision quota, determine the data block size according to the channel quality feature set, and package the encoded data into blocks according to the hierarchical priority order of key layer, important layer, and secondary layer to form a data block sequence and a transmission list file. The dynamic transmission scheduling module is used to transmit the transmission list file and data block sequence to the receiving end in sequence number. During the transmission process, the channel quality feature set is updated every 30 seconds and the data block size and concurrent transmission quantity are dynamically adjusted according to the bandwidth change ratio. The dequantization decoding module is used to perform dequantization reconstruction on the data blocks that arrive successively. It extracts the quantization parameters from the header of the data block, performs dequantization calculation on the quantization index to obtain the restored parameter values, and fills them into the corresponding parameter layer group storage area according to the parameter position index. After all data blocks are dequantized, the model parameters are reassembled to generate a complete model file.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This invention addresses the problems of concentrated loss of key parameter accuracy and inability to respond to network fluctuations caused by indiscriminate compression strategies in existing large model transmission technologies. It achieves end-to-end collaboration through five stages: real-time network channel awareness, hierarchical evaluation of parameter influence, adaptive coding decision-making, priority transmission scheduling, and accurate reconstruction at the receiver. By implementing a collaborative approach, it prioritizes transmission and ensures higher coding accuracy based on channel quality and parameter importance. Secondary layer parameters undertake more compression tasks when bandwidth is limited, thereby maximizing transmission efficiency while ensuring core inference capabilities. A dynamic scheduling mechanism enables the transmission process to adaptively respond to network fluctuations, increasing throughput when the network improves and reducing retransmission costs when the network deteriorates. It also increases concurrency to improve throughput. When the network deteriorates, it reduces data blocks, lowers concurrency, and combines timeout retransmission and temporary skipping mechanisms to reduce retransmission costs and avoid transmission deadlocks. This results in more stable and efficient large model transmission in complex network environments, ultimately enabling the receiver to quickly obtain a complete, directly deployable model file within a controllable range of accuracy loss. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the system framework structure of the present invention; Figure 2 This is a schematic diagram of the method framework structure of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1: Please refer to Figure 1 This application provides a method for large-scale compression and acceleration for transmission, including: Step 1: Before and during model transmission, multiple sets of probe data packets are sent between the sending and receiving ends to continuously collect quality characteristic parameters in the network transmission channel. The collected data is analyzed to obtain five quality characteristic parameters: available bandwidth, transmission delay, delay jitter, packet loss rate, and transmission stability level. These five quality characteristic parameters are then summarized to form a channel quality feature set, providing a decision-making basis for subsequent parameter coding strategies. The specific method is as follows: First, raw data for calculating available bandwidth is collected and the available bandwidth is calculated. Specifically, 10 sets of test data packets are sent from the sending end to the receiving end. Each set of data packets is 1 megabyte in size. The time taken from the sending end to the receiving end returning an acknowledgment message for each set of data packets is recorded. The 1 megabyte is divided by the corresponding transmission time of the set to obtain the instantaneous bandwidth of the set, in megabytes per second. Then, the 10 sets of instantaneous bandwidth values ​​are added together and divided by 10 to obtain the average available bandwidth, which is used as the available bandwidth parameter in the channel quality feature set. Then, 20 test data packets carrying sending timestamps are continuously sent from the sending end to the receiving end. After receiving each test data packet, the receiving end immediately returns an acknowledgment data packet to the sending end. The sending end records the time interval between sending the test data packet and receiving the corresponding acknowledgment data packet as the round-trip time of this test. The round-trip time is divided by 2 to obtain the one-way transmission delay of this test. The 20 one-way transmission delay values ​​are sorted from smallest to largest, and the values ​​of the 10th and 11th positions are added together and divided by 2 to obtain the median transmission delay, which is used as the transmission delay parameter in the channel quality feature set. The median is used instead of the mean because the median has a stronger resistance to interference from occasional extreme delay values ​​and can more stably reflect the typical delay level of the channel. Next, the delay jitter is calculated based on 20 single-pass transmission delay measurements. Specifically, the mean of the 20 single-pass transmission delay values ​​is taken as the delay mean. Then, the absolute value of the difference between each measurement value and the delay mean is calculated. The 20 absolute differences are squared and summed. The sum is then divided by 20. Finally, the square root of the result is taken to obtain the delay standard deviation. This standard deviation is taken as the delay jitter value and used as the delay jitter parameter in the channel quality feature set. The larger the delay jitter value, the more severe the channel delay fluctuation and the more unstable the network. Then, raw data for calculating the packet loss rate is collected and the packet loss rate is calculated. Specifically, 100 test data packets are continuously sent from the sending end to the receiving end. The receiving end counts the number of data packets actually received within the specified waiting time. The number of lost data packets is obtained by subtracting the number of data packets actually received from 100. The number of lost data packets is divided by 100 and then multiplied by 100% to obtain the packet loss rate, which is used as the packet loss rate parameter in the channel quality feature set. Here, the waiting time is set to 5 times the median transmission delay.

[0019] Finally, based on the calculated latency jitter and packet loss rate, the transmission stability level is comprehensively determined. Specifically: if the latency jitter is less than 5 milliseconds and the packet loss rate is less than 1%, it is considered a stable channel; if the latency jitter is between 5 and 20 milliseconds or the packet loss rate is between 1% and 5%, it is considered a general channel; if the latency jitter is greater than 20 milliseconds or the packet loss rate is greater than 5%, it is considered an unstable channel. The determination result is used as the transmission stability level parameter in the channel quality feature set. During model transmission, the above measurement and calculation process is repeated every 30 seconds to update the channel quality feature set to adapt to dynamic changes in network conditions.

[0020] By actively sending probe packets between the sender and receiver, raw response data of the network channel is collected from actual transmission behavior. This raw data is then transformed into five structured parameters: available bandwidth, transmission delay, delay jitter, packet loss rate, and transmission stability level, which are summarized into a channel quality feature set. Available bandwidth reflects the channel's data carrying capacity, transmission delay reflects the channel's basic response speed, delay jitter reflects the channel's stability, packet loss rate reflects the channel's reliability, and the transmission stability level is a comprehensive judgment of these parameters, providing a unified decision anchor for subsequent steps. The channel quality feature set is continuously updated every 30 seconds during transmission, enabling the entire transmission system to maintain real-time awareness of dynamic changes in network status. This solves the fundamental problem of the disconnect between compression strategies and the transmission environment in existing technologies, providing objective, continuous, and structured channel status input for subsequent dynamic coding decisions.

[0021] Step 2: Perform parameter influence analysis on the large model to be transmitted, establish a parameter influence hierarchical map, and assign different transmission priorities to different parameter groups. The specific method is as follows: First, the large model to be transmitted is grouped into parameter layers according to the network structure. The large language model is divided into four categories: embedding layer parameters, multi-head attention layer parameters, feedforward network layer parameters, and output layer parameters. For multi-head attention layer parameters and feedforward network layer parameters, they are further subdivided into three subgroups according to the model depth: front-end, middle-end, and back-end. The specific division method is as follows: for a model with a total of N layers, the layers corresponding to the number of layers from layer 1 to layer N multiplied by one-third and rounded down are the front-end parameter layer group; the layers corresponding to the number of ... Then, the parameter influence of each parameter layer group is evaluated. Specifically, a standard test dataset is prepared, containing 1000 representative inference samples covering classification and generation tasks. For each parameter layer group, the following operations are performed sequentially: Based on the original model, random perturbations are added to all parameter values ​​within the parameter layer group, with a perturbation magnitude of 10% of the original parameter values. The perturbated model is then used to perform inference on the aforementioned 1000 test samples, and the magnitude of change in the inference results relative to the original model is statistically analyzed. For classification task samples, the magnitude of change is calculated by dividing the number of samples whose predicted category changed by the total number of classification task samples, yielding the classification change rate. For generation task samples, the magnitude of change is calculated by dividing the change in edit distance between the generated text and the original text by the length of the original text, and then averaging the results across all generation task samples, yielding the generation change rate. The classification change rate and the generation change rate are added together and divided by 2 to obtain the influence score of the parameter layer group. A higher influence score indicates a more significant and important impact of the parameter on the inference results.

[0022] Next, all parameter layers are classified hierarchically based on their influence scores. The classification rules are as follows: parameter layers with an influence score greater than or equal to 0.3 are classified as key layers; parameter layers with an influence score greater than or equal to 0.1 and less than 0.3 are classified as important layers; and parameter layers with an influence score less than 0.1 are classified as secondary layers.

[0023] Finally, the hierarchical identifiers, influence scores, number of parameters, and position indices in the model for each parameter group are compiled and summarized to form a hierarchical graph of parameter influence.

[0024] By applying controlled perturbations to a large model and observing changes in inference output, the actual contribution of each parameter layer to the model's inference results is quantified. All parameter layers are then categorized into three levels—critical, important, and minor—based on their influence scores, forming a parameter influence hierarchy map. The core mechanism of this evaluation method is that parameter layers with a more significant impact on inference output suffer greater damage to model usability due to transmission accuracy loss, thus requiring priority in ensuring transmission quality. Parameter layers with a weak impact on inference output can undergo greater accuracy compression when bandwidth is limited, thereby freeing up bandwidth resources for high-precision transmission of critical parameters. This solves the problem of concentrated accuracy loss of critical parameters caused by indiscriminate compression in existing technologies, providing a basis for subsequent differentiated coding strategies based on the intrinsic semantics of the model's parameters.

[0025] Step 3: Combine the channel quality feature set with the parameter influence hierarchy map for analysis, and calculate the corresponding coding precision quota for each parameter layer group. This ensures that the allocation of coding precision simultaneously reflects the importance level of the parameter itself and the actual carrying capacity of the current channel. The specific method is as follows: First, based on the available bandwidth and expected transmission time in the channel quality feature set, the overall compression target rate for this transmission is calculated. The expected transmission time is preset according to the application scenario. For the model deployment scenario, the expected transmission time is set to 70% of the actual available time, leaving a 30% time margin to cope with network fluctuations. The overall compression target rate is calculated by dividing the total size of the original parameters of the model by the product of the available bandwidth and the expected transmission time to obtain the required compression ratio. If the calculation result is less than 1, it means that the bandwidth is sufficient and no compression is needed, and the compression target rate is set to 1; if the calculation result is greater than 1, the calculation result is taken as the compression target rate.

[0026] Then, based on the hierarchical classification of each parameter layer group in the parameter influence stratification map output in step two, a baseline value for coding precision is set for the critical layer, important layer, and minor layer. Coding precision is represented by the number of bits, i.e., how many bits are used to encode and store each parameter value. A higher number of bits results in higher precision and greater storage capacity. The baseline value setting rule is as follows: the baseline value for the critical layer is set to 16 bits, the baseline value for the important layer is set to 8 bits, and the baseline value for the minor layer is set to 4 bits. These baseline values ​​represent the recommended coding precision for each layer parameter under good network conditions.

[0027] Next, based on the transmission stability level in the channel quality feature set, the baseline values ​​of coding precision for each layer are adjusted. The adjustment rules are as follows: if the transmission stability level is stable, the coding precision of each layer remains unchanged, i.e., 16 bits for the critical layer, 8 bits for the important layer, and 4 bits for the minor layer; if the transmission stability level is moderate, the coding precision of each layer is moderately compressed, with the coding precision of the critical layer adjusted to 12 bits (16 bits multiplied by 0.75), the coding precision of the important layer adjusted to 4 bits (8 bits multiplied by 0.5), and the coding precision of the minor layer adjusted to 2 bits (4 bits multiplied by 0.5); if the transmission stability level is unstable, the coding precision of each layer is significantly compressed, with the coding precision of the critical layer adjusted to 8 bits (16 bits multiplied by 0.5), the coding precision of the important layer adjusted to 2 bits (8 bits multiplied by 0.25), and the coding precision of the minor layer adjusted to 1 bit.

[0028] Next, verify whether the adjusted encoding precision meets the overall compression target rate. The verification method is as follows: multiply the number of parameters in each parameter layer group by the adjusted encoding precision corresponding to that layer group, sum the calculation results of all parameter layer groups to obtain the total number of compressed bits, divide the total number of bits by 8 to obtain the total number of compressed bytes, and then convert it to the compressed model size in gigabytes; divide the total size of the original model parameters by the compressed model size to obtain the actual compression ratio; compare the actual compression ratio with the overall compression target rate. If the actual compression ratio is greater than or equal to the compression target rate, the transmission requirements are met. If it is less than the compression target rate, further compression is required. The adjustment strategy is to gradually reduce the encoding precision of each layer in the order of secondary layers first, important layers second, and critical layers last. The reduction amount each time is the current encoding precision divided by 2 and rounded down, and the adjusted encoding precision is not less than 1 bit. Recalculate the actual compression ratio and repeat the above process until the actual compression ratio meets the compression target rate.

[0029] The final determined coding precision of each parameter layer group is recorded as the parameter layer coding precision quota.

[0030] By jointly analyzing the channel quality feature set output from step one and the parameter influence hierarchy map output from step two, the required compression level under the current transmission conditions is determined through overall compression target rate calculation. Then, differentiated coding precision is assigned to parameter layers of different importance levels based on transmission stability level, and iterative verification ensures that the actual compression ratio meets the target requirements. The core mechanism of this step is that the allocation of coding precision is constrained by both parameter importance and channel carrying capacity. Parameters with higher importance preferentially obtain higher coding precision under any channel conditions, while the worse the channel conditions, the greater the overall coding precision and overall compression level. These two constraints work together to determine the coding precision of each parameter layer, achieving optimal allocation of transmission resources. This solves the problem in existing technologies where fixed compression parameters cannot respond to dynamic channel changes, forming an adaptive coding mechanism driven by channel awareness and parameter importance.

[0031] Step 4: Based on the parameter-level coding precision quota, perform differentiated linear quantization coding on each parameter level group, and determine the block segmentation strategy according to the channel quality feature set. Then, divide the encoded data into blocks and package them according to hierarchical priority to form a sequence of data blocks to be transmitted. Specifically: First, linear quantization is performed on each parameter layer group according to its corresponding coding precision quota. The specific steps are as follows: scan all parameter values ​​within the parameter layer group, find the maximum and minimum values, and subtract the minimum value from the maximum to obtain the parameter value range of the layer group; calculate the quantization levels by multiplying the base 2 by the number of bits corresponding to the coding precision to the power of 2, obtaining the total quantization levels, and then subtracting 1 from the total quantization levels. For example, when the coding precision is 6 bits, the total quantization levels are 2 to the power of 6, which equals 64, and the quantization levels are 64 minus 1, which equals 63; calculate the quantization interval by dividing the parameter value range by the number of quantization levels. When the parameter value range is 4.4 and the quantization level is 63, the quantization interval is approximately 0.07 (4.4 divided by 63). Quantization is performed on each parameter value in the layer group. The quantization operation is calculated by subtracting the minimum value of the layer group from the parameter value, dividing by the quantization interval, and rounding the result to the nearest integer to obtain the quantization index corresponding to the parameter value. The quantization index ranges from 0 to the quantization level. The quantization index is stored in binary encoding using the corresponding number of bits for the encoding precision. Simultaneously, the quantization parameters of the parameter layer group are recorded, including the maximum value, minimum value, and quantization interval. These three quantization parameters will be used at the receiving end to perform dequantization to restore the parameter values.

[0032] Then, based on the transmission stability level and transmission delay in the channel quality feature set output in step one, the data block size is determined. The specific rules are as follows: if the transmission stability level is stable and the transmission delay is less than 50 milliseconds, a large block transmission strategy is adopted, and the size of each data block is set to 5 megabytes; if the transmission stability level is moderate or the transmission delay is between 50 and 100 milliseconds, a medium block transmission strategy is adopted, and the size of each data block is set to 2 megabytes; if the transmission stability level is unstable or the transmission delay is greater than 100 milliseconds, a small block transmission strategy is adopted, and the size of each data block is set to 500 kilobytes. Next, following the hierarchical priority order of critical layer, important layer, and secondary layer, the encoded data of each parameter layer group is sequentially divided into blocks and packaged. Within the same layer, the data is arranged according to the position index of the layer group in the model. Each data block contains the following: data block header information, including data block sequence number, parameter layer group identifier, layer classification, start and end position indices of the parameters in the corresponding layer group, quantization parameters, i.e., maximum value, minimum value, and quantization interval; data block body, which is the binary encoded bit stream of all parameter quantization indices in the block; and data block verification information, which uses a cyclic redundancy check algorithm to calculate a checksum for the data block body, used by the receiving end to verify the integrity of the data block.

[0033] Finally, a transmission manifest file is generated, which records the metadata of all data blocks, including the total number of data blocks, the sequence number, size, level, and checksum of each data block. The transmission manifest file is sent to the receiving end before the data blocks are transmitted, so that the receiving end can understand the overall situation of this transmission task and prepare storage space in advance.

[0034] Linear quantization coding is performed on each parameter layer group according to the coding precision quota determined in step three. The data block size is determined based on the channel quality feature set, and the blocks are packaged in a priority order based on critical layers to generate a data block sequence with complete metadata and a transmission list file. The core mechanism of linear quantization is as follows: the quantization interval is determined by the parameter value range and the quantization level, mapping continuous floating-point parameter values ​​to discrete integer indices, storing approximate parameter values ​​with fewer bits, and recording quantized parameters for inverse quantization reconstruction at the receiver; the data block size is adaptively adjusted according to channel stability, using larger blocks to reduce transmission overhead in stable channels and smaller blocks to reduce the cost of single retransmission in unstable channels; the priority packaging order based on critical layers gives the transmission sequence a natural priority characteristic, ensuring that important parameters arrive at the receiver first without additional scheduling logic. Transforming the coding decision in step three into a physically transmittable data structure is the key link in the entire scheme from decision-making to execution.

[0035] Step 5: Transmit data blocks to the receiving end in the order of the data block sequence output in Step 4, and dynamically adjust the transmission strategy according to the real-time updated channel quality feature set during the transmission process. The specific method is as follows: First, the sending end sends the transmission manifest file generated in step four to the receiving end, and waits for the receiving end to return a manifest confirmation message, confirming that the receiving end understands the overall situation of this transmission task.

[0036] Then, data blocks are sent sequentially in ascending order of their sequence numbers. Since the data blocks were numbered according to the priority order of critical layer, important layer, and secondary layer in step four, the transmission order naturally ensures that parameters of higher importance arrive at the receiving end first. After sending each data block, the sending end waits for an acknowledgment message from the receiving end. Upon receiving the data block, the receiving end uses the checksum recorded in the data block header to verify the integrity of the data block body. If the verification is successful, it returns an acknowledgment success message containing the data block sequence number to the sending end; if the verification fails, it returns a retransmission request message containing the data block sequence number to the sending end. The sending end maintains a transmission status table to record the current status of each data block, with four status types: not sent, sent and awaiting acknowledgment, acknowledgment successful, and retransmission required.

[0037] Next, a dynamic scheduling mechanism is executed during transmission. Every 30 seconds, the channel quality parameters are remeasured and the channel quality feature set is updated according to the method described in step one. The bandwidth ratio is obtained by dividing the current available bandwidth by the initial available bandwidth recorded at the start of transmission. If the ratio is greater than 1.2, it is determined that the network has improved. If the ratio is less than 0.8, it is determined that the network has deteriorated. If the ratio is between 0.8 and 1.2, it is determined that the network is stable.

[0038] If the network is determined to be improved, the transmission efficiency is increased. The specific strategy is as follows: multiply the current data block size by 2.5 and round up. If the result exceeds 5 megabytes, take 5 megabytes as the adjusted data block size. Merge multiple unsent original data blocks into a larger data block and repackage them. Adjust the number of concurrent transmissions from 1 data block at a time to 3 data blocks at a time before waiting for confirmation, thereby improving the transmission throughput.

[0039] If network degradation is detected, a conservative transmission strategy is adopted. Specifically, the current data block size is multiplied by 0.25 and rounded down. If the result is less than 500 kilobytes, 500 kilobytes is used as the adjusted data block size. Subsequent unsent data blocks are split into smaller blocks and repackaged. The concurrent transmission rate is adjusted to send only one data block at a time and wait for confirmation. For data blocks that have been sent but have not received an confirmation message after the median transmission delay is multiplied by 5, they are judged as timeouts and retransmission is initiated. If a data block fails to be retransmitted after more than 3 retransmissions, it is marked as temporarily skipped, and subsequent critical layer data blocks are transmitted first. Once the network condition is restored, the skipped data blocks are retransmitted.

[0040] During transmission, the sending end synchronizes transmission progress information to the receiving end every 100 data blocks transmitted, including the number of data blocks successfully transmitted, the current real-time transmission rate, and the estimated remaining transmission time. The current real-time transmission rate is calculated by dividing the total size of the most recent 100 confirmed data blocks by the time taken from the first data block sent to the last confirmed data block received. The estimated remaining transmission time is calculated by dividing the total size of the remaining untransmitted data blocks by the current real-time transmission rate.

[0041] When the transmission status of all data blocks is updated to acknowledge success, the sender sends a transmission completion message to the receiver, and the transmission process ends.

[0042] By transmitting data blocks to the receiving end according to priority, and re-measuring channel quality parameters every 30 seconds during transmission, the network state change trend is judged based on the ratio of the current available bandwidth to the initial available bandwidth, and the data block size and concurrent transmission quantity are dynamically adjusted accordingly. When the network improves, the data block size is increased and the concurrency is increased to improve throughput; when the network deteriorates, the data block size is reduced and the concurrency is decreased to improve transmission reliability. Timed-out data blocks are actively retransmitted, and after more than 3 retransmissions, they are temporarily skipped to prioritize the continuity of critical layer transmission. The core mechanism of this step is that the transmission scheduling strategy and the channel quality feature set form a closed-loop feedback. The channel state updates continuously generated in step one drive the scheduling decision in step five in real time, enabling the transmission process to adaptively respond to dynamic changes in network state and maximize the utilization of available bandwidth while ensuring the continuity of critical parameter transmission. This solves the problem of fixed transmission strategies and inability to respond to network fluctuations in existing technologies, and significantly improves the transmission efficiency of large models in complex network environments.

[0043] Step Six: At the receiving end, dequantize and decode the received data block, and write the restored parameters into the corresponding model structure to generate a complete model file. The specific method is as follows: First, the receiving end retrieves the data blocks to be processed in the order of priority of key layer, important layer, and secondary layer according to the sequence of successfully confirmed data blocks in step five. Then, dequantization is performed on each data block. Specifically, the quantization parameters, including the maximum value, minimum value, and quantization interval, are read from the header information of the data block. The quantization index bitstream in the main body of the data block is read, and the binary encoded data stored in the main body of the data block is truncated sequentially according to the number of bits corresponding to the encoding precision recorded in the header of the data block. Each segment of binary encoding is restored to the corresponding quantization index integer, resulting in a sequence of quantization index integers. Dequantization is performed on each quantization index by multiplying the quantization index by the quantization interval and adding the minimum value of the parameter to obtain the restored floating-point parameter value. The restored floating-point parameter value is written to the parameter storage area of ​​the corresponding parameter layer group according to the start and end indexes recorded in the header of the data block. After writing, the data block is marked as dequantized. Finally, after all data blocks have been marked as dequantized, the model parameters are reassembled. Specifically, the restored parameter values ​​are read from the parameter storage area of ​​each parameter layer group and filled into the corresponding positions in the model structure according to the network structure order of the original model. The number of parameters in each parameter layer group is checked to see if it is consistent with the original model definition. If an inconsistency is found, an exception report message is sent to the sending end. After all parameters are filled, a complete model file is generated and saved to the local storage of the receiving end for subsequent loading and deployment.

[0044] The receiving end performs real-time dequantization reconstruction on incoming data blocks. Quantization parameters are extracted from the data block header, and dequantization calculations are performed to restore floating-point parameter values. The restored parameters are then filled into the storage area of ​​the corresponding parameter layer group according to their position indices. Once all data blocks have been dequantized, model parameters are reassembled. The parameter values ​​of each parameter layer group are filled into the corresponding positions in the model structure according to the original model network structure, generating a complete model file. The core mechanism of this step is that the dequantization operation uses quantization parameters to restore discrete integer indices to continuous floating-point values, achieving approximate restoration of parameter values; the parameter reassembly process ensures that the restored parameter values ​​are accurately filled into the corresponding positions in the model structure, maintaining the integrity of the model's network topology. This solves the problem of restoring model parameters after compressed transmission, enabling the receiving end to obtain a complete model file that can be directly loaded and deployed.

[0045] Example 2: Please refer to Figure 2 As shown, this embodiment also provides a large model compression and acceleration system for transmission, which is implemented through the aforementioned disclosed large model compression and acceleration method for transmission, specifically including: The channel quality acquisition module, located at the transmitting end, is used to send multiple sets of probe data packets to the receiving end before and during model transmission. It continuously collects and analyzes the quality characteristic parameters in the network transmission channel to obtain five quality characteristic parameters: available bandwidth, transmission delay, delay jitter, packet loss rate, and transmission stability level. The five quality characteristic parameters are summarized to form a channel quality feature set, which is updated every 30 seconds during transmission. The parameter influence analysis module, located at the sending end, is used to divide the large model parameters to be transmitted into several parameter layer groups according to the network structure, evaluate the parameter influence of each parameter layer group, and divide the parameter layer group into three levels: key layer, important layer and secondary layer according to the influence score, and establish a parameter influence hierarchical map. The coding precision quota calculation module, located at the transmitting end, is used to combine and analyze the channel quality feature set with the parameter influence hierarchical map, calculate the overall compression target rate based on the available bandwidth and expected transmission time, and calculate the corresponding coding precision quota for each parameter layer group based on the parameter level and transmission stability level. The quantization encoding and packaging module, located at the transmitting end, is used to perform differentiated linear quantization encoding on each parameter layer group according to the encoding precision quota, determine the data block size according to the channel quality feature set, and package the encoded data into blocks according to the hierarchical priority order of key layer, important layer, and secondary layer to generate a data block sequence and transmission list file. The dynamic transmission scheduling module, located at the sending end, is used to transmit the transmission list file and data block sequence to the receiving end in sequence number. During the transmission process, it dynamically adjusts the data block size and the number of concurrent transmissions based on the real-time updated channel quality feature set. It actively retransmits timed-out data blocks and temporarily skips and retransmits data blocks that fail to be retransmitted after more than 3 retransmissions. The dequantization decoding module, located at the receiving end, is used to sequentially extract the data blocks to be processed according to the confirmed successful data block sequence, read the quantization parameters from the header information of the data block, parse the quantization index bit stream in the main body of the data block into a sequence of quantization index integers according to the corresponding encoding precision, multiply each quantization index by the quantization interval and add the minimum value of the parameter to obtain the restored floating-point parameter value, and write the restored floating-point parameter value into the parameter storage area of ​​the corresponding parameter layer group according to the start position index and end position index. After all data blocks have been dequantized, the model parameters are reassembled. The restored parameter values ​​are filled into the corresponding positions of the model structure according to the network structure order of the original model, generating a complete model file and saving it to the local storage of the receiving end.

[0046] By integrating five end-to-end processes—real-time network channel awareness, hierarchical evaluation of parameter influence, adaptive coding decision-making, priority transmission scheduling, and accurate reconstruction at the receiver—this approach addresses the problems of indiscriminate compression strategies leading to concentrated loss of key parameter accuracy and fixed transmission strategies failing to respond to network fluctuations in existing large-scale model transmission technologies. It prioritizes transmission and ensures higher coding accuracy through channel quality and parameter importance, allowing secondary layer parameters to undertake more compression tasks when bandwidth is limited. This maximizes transmission efficiency while ensuring core inference capabilities. A dynamic scheduling mechanism enables the transmission process to adaptively respond to network fluctuations, increasing throughput when the network improves and reducing retransmission costs when the network deteriorates. It also increases concurrency to improve throughput. When the network deteriorates, it reduces data blocks, lowers concurrency, and combines timeout retransmission and temporary skipping mechanisms to reduce retransmission costs and avoid transmission deadlocks. This achieves more stable and efficient large-scale model transmission in complex network environments, ultimately enabling the receiver to quickly obtain a complete, directly deployable model file within a controllable range of accuracy loss.

[0047] Example 3: As Example 3 of the present invention, in specific implementation, compared with Example 1 and Example 2, the technical solution of this example is to combine the solutions of Example 1 and Example 2.

[0048] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0049] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for large-scale model compression and acceleration oriented towards transmission, characterized in that, include: Step 1: Before and during model transmission, send multiple sets of probe data packets between the sending and receiving ends to continuously collect and analyze the quality characteristic parameters in the network transmission channel. Obtain five quality characteristic parameters: available bandwidth, transmission delay, delay jitter, packet loss rate, and transmission stability level. Summarize the five quality characteristic parameters to form a channel quality characteristic set. Step 2: Divide the parameters of the large model to be transmitted into several parameter layers according to the network structure, evaluate the parameter influence of each parameter layer, and divide the parameter layers into three levels: key layer, important layer and secondary layer according to the influence score, and establish a parameter influence hierarchical map. Step 3: Combine the channel quality feature set with the parameter influence layer map for analysis, calculate the overall compression target rate based on available bandwidth and expected transmission time, and calculate the corresponding coding precision quota for each parameter layer group based on parameter level and transmission stability level; Step 4: Perform differentiated linear quantization coding on each parameter layer group according to the parameter layer coding precision quota, determine the data block size according to the channel quality feature set, and divide and package the coded data into blocks according to the hierarchical priority order of key layer, important layer, and secondary layer to form a data block sequence and transmission list file; Step 5: Transmit the transmission manifest file and data block sequence to the receiving end in sequence number. During the transmission process, update the channel quality feature set every 30 seconds and dynamically adjust the data block size and the number of concurrent transmissions according to the bandwidth change ratio. Step 6: The receiving end performs dequantization decoding on the received data blocks, writes the restored parameters into the parameter storage area of ​​the corresponding parameter layer group according to the position index, and performs model parameter reassembly after all data blocks are dequantized to generate a complete model file and save it to the local storage of the receiving end.

2. The large-model compression and acceleration method for transmission according to claim 1, characterized in that, The specific methods for obtaining available bandwidth, transmission latency, latency jitter, and packet loss rate are as follows: Send 10 sets of test data packets and record the transmission time of each set. Divide the size of each set of data packets by the corresponding transmission time to obtain the instantaneous bandwidth. Use the average of the 10 sets of instantaneous bandwidth as the available bandwidth. Twenty test data packets with timestamps are sent. After receiving each test data packet, the receiving end immediately returns an acknowledgment data packet to the sending end. The sending end divides the round-trip time by 2 to obtain the one-way transmission delay. The 20 one-way transmission delays are sorted from smallest to largest, and the values ​​of the 10th and 11th positions are added together and divided by 2 to obtain the median transmission delay. The 20 one-way transmission delay values ​​are added together and divided by 20 to obtain the mean delay. The absolute value of the difference between each measurement value and the mean delay is squared, added together, and then divided by 20. The square root of the result is taken to obtain the standard deviation of the delay as the delay jitter. 100 test data packets are continuously sent from the sending end to the receiving end. The receiving end counts the number of data packets actually received within the waiting time corresponding to 5 times the median transmission delay. The packet loss rate is obtained by dividing the number of lost data packets by 100 and then multiplying by 100%.

3. The large-model compression and acceleration method for transmission according to claim 2, characterized in that, The specific method for obtaining the transmission stability level is as follows: If the latency jitter is less than 5 milliseconds and the packet loss rate is less than 1%, it is determined to be a stable channel; if the latency jitter is between 5 and 20 milliseconds or the packet loss rate is between 1% and 5%, it is determined to be a general channel; if the latency jitter is greater than 20 milliseconds or the packet loss rate is greater than 5%, it is determined to be an unstable channel. During the model transmission process, the above measurement and calculation process is repeated every 30 seconds to update the channel quality feature set.

4. The large-model compression and acceleration method for transmission according to claim 1, characterized in that, The specific method for evaluating the parameter influence of each parameter group is as follows: Prepare a standard test dataset containing classification task samples and generation task samples. For each parameter layer group, add random perturbations to the original model, with the perturbation magnitude being 10% of the original values ​​of the parameters for that layer group. Use the perturbated model to perform inference on the test samples. For classification task samples, count the number of samples whose predicted categories changed and divide by the total number of classification task samples to obtain the classification change rate. For generation task samples, calculate the change in edit distance between the generated text and the original text, divide by the length of the original text, and take the average of all generation task samples to obtain the generation change rate. Use the average of the classification change rate and the generation change rate as the influence score for the corresponding parameter layer group. Parameter layers with an influence score greater than or equal to 0.3 are designated as key layers, those with an influence score greater than or equal to 0.1 and less than 0.3 are designated as important layers, and those with an influence score less than 0.1 are designated as secondary layers.

5. The large-model compression and acceleration method for transmission according to claim 3, characterized in that, The specific method for calculating the corresponding coding precision quota for each parameter group based on the parameter level and channel stability level is as follows: The compression target rate is obtained by dividing the total size of the original model parameters by the product of the available bandwidth and the expected transmission time. If the calculated result is less than 1, the compression target rate is set to 1; if the calculated result is greater than or equal to 1, the calculated result is used as the compression target rate. Based on the parameter hierarchy, encoding precision baseline values ​​of 16 bits, 8 bits, and 4 bits are set for the critical layer, important layer, and minor layer, respectively. The encoding precision baseline values ​​are adjusted according to the transmission stability level. When the transmission stability level is stable, the baseline values ​​for each layer remain unchanged. When the transmission stability level is moderate, the critical layer is adjusted to 12 bits (16 bits multiplied by 0.75), the important layer to 4 bits (8 bits multiplied by 0.5), and the minor layer to 2 bits (4 bits multiplied by 0.5). When the transmission stability level is unstable, the critical layer is adjusted to 8 bits (16 bits multiplied by 0.5), the important layer to 2 bits (8 bits multiplied by 0.25), and the minor layer to 1 bit. Multiply the number of parameters in each parameter layer group by the adjusted coding precision corresponding to that layer group, and sum the results for all parameter layer groups to obtain the total number of compressed bits. Divide the total number of compressed bits by 8 to obtain the total number of compressed bytes. Divide the total size of the original model parameters by the size of the compressed model to obtain the actual compression ratio. If the actual compression ratio is less than the compression target ratio, gradually reduce the coding precision of each layer in the order of secondary layers first, important layers second, and critical layers last. Each reduction is equal to the current coding precision divided by 2 and rounded down, with a reduction of at least 1 bit. Recalculate the actual compression ratio and repeat the above process until the actual compression ratio meets the compression target ratio.

6. The large-model compression and acceleration method for transmission according to claim 5, characterized in that, The specific method for performing differentiated linear quantization encoding on each parameter layer group is as follows: Scan all parameter values ​​within the parameter layer group to find the maximum and minimum values, and subtract the minimum value from the maximum value to obtain the parameter value range; calculate the total number of quantization levels by raising the number of encoding precision bits to the power of 2, and subtract 1 from the total number of quantization levels to obtain the quantization level; Divide the parameter value range by the quantization level to obtain the quantization interval; subtract the minimum value of the layer group from each parameter value in the layer group, divide by the quantization interval and round to obtain the quantization index; encode and store the quantization index in binary using the number of bits corresponding to the coding precision. Simultaneously, the maximum value, minimum value, and quantization interval of the parameters are recorded as quantization parameters for the receiving end to use for dequantization.

7. The large-model compression and acceleration method for transmission according to claim 3, characterized in that, The specific method for determining the data block size based on the channel quality feature set is as follows: When the transmission stability level is stable and the transmission latency is less than 50 milliseconds, the size of each data block is set to 5 megabytes. When the transmission stability level is moderate or the transmission latency is between 50 and 100 milliseconds, the size of each data block is set to 2 megabytes. When the transmission stability level is unstable or the transmission latency is greater than 100 milliseconds, the size of each data block is set to 500 kilobytes. Each data block contains data block header information, data block body, and data block check information. The data block header information includes the data block sequence number, the identifier of the parameter layer group to which it belongs, the layer classification, the start and end position indices of the parameters in the corresponding layer group, and the quantization parameters. The data block body is the binary encoded data of the quantization indices of all parameters in the block. The data block check information is the check code calculated by the cyclic redundancy check algorithm on the data block body.

8. The large-model compression and acceleration method for transmission according to claim 2, characterized in that, The specific method for dynamically adjusting the data block size and concurrent transmission quantity based on the bandwidth change ratio is as follows: The bandwidth ratio is obtained by dividing the current available bandwidth by the initial available bandwidth recorded at the start of transmission. If the bandwidth ratio is greater than 1.2, the network is considered improved. The current data block size is multiplied by 2.5 and rounded up to a value not exceeding 5 megabytes to obtain the adjusted data block size. Multiple unsent original data blocks are merged and repackaged, and the concurrent transmission quantity is adjusted from 1 to 3 at a time. If the bandwidth ratio is less than 0.8, the network is considered deteriorated. The current data block size is multiplied by 0.25 and rounded down to a value not less than 500 kilobytes to obtain the adjusted data block size. The unsent data blocks are split and repackaged, and the concurrent transmission quantity is adjusted to 1 at a time. For data blocks that have been sent but have not received an acknowledgment message after the median transmission delay multiplied by 5, they are considered to have timed out and are actively retransmitted. Data blocks that fail to be retransmitted after more than 3 retransmissions are marked as temporarily skipped, and subsequent critical layer data blocks are prioritized for transmission. Once the network condition is restored, the skipped data blocks will be retransmitted.

9. The large-model compression and acceleration method for transmission according to claim 6, characterized in that, The specific method by which the receiving end performs dequantization decoding on the received data block is as follows: Read the quantization parameters from the header information of the data block, including the maximum value, minimum value and quantization interval; extract the binary encoded data stored in the body of the data block according to the number of bits corresponding to the encoding precision recorded in the header of the data block, and restore each segment of binary encoding to the corresponding quantization index integer to obtain the quantization index integer sequence. Multiply each quantization index integer by the quantization interval and add the minimum parameter value to obtain the restored floating-point parameter value; The restored floating-point parameter values ​​are written into the parameter storage area of ​​the corresponding parameter layer group according to the start and end position indices recorded in the header of the data block. After all data blocks have been dequantized, the restored parameter values ​​are read from the parameter storage area of ​​each parameter layer group in the order of the original model network structure and filled into the corresponding positions of the model structure. The number of parameters in each parameter layer group is checked to see if it is consistent with the original model definition. If it is inconsistent, an exception report message is sent to the sending end. After all parameters are filled, a complete model file is generated and saved to the local storage of the receiving end.

10. A large-model compression and acceleration system for transmission, characterized in that, The system is implemented using the large-model compression and acceleration method for transmission as described in any one of claims 1-9, including: The channel quality acquisition module is used to send multiple sets of probe data packets between the sending end and the receiving end before and during model transmission, continuously collect the quality characteristic parameters of the network transmission channel, obtain available bandwidth, transmission delay, delay jitter, packet loss rate and transmission stability level, summarize them to form a channel quality feature set, and update the channel quality features every 30 seconds during transmission. The parameter influence assessment module is used to divide the parameters of the large model to be transmitted into several parameter layer groups according to the network structure, assess the parameter influence of each parameter layer group, divide the parameter layer group into key layer, important layer and secondary layer according to the influence score, and establish a parameter influence hierarchical map. The coding precision quota calculation module is used to combine and analyze the channel quality feature set with the parameter influence hierarchical map, calculate the overall compression target rate based on the available bandwidth and expected transmission time, and calculate the corresponding coding precision quota for each parameter layer group based on the parameter level and transmission stability level. The quantization encoding and packaging module is used to perform differentiated linear quantization encoding on each parameter layer group according to the encoding precision quota, determine the data block size according to the channel quality feature set, and package the encoded data into blocks according to the hierarchical priority order of key layer, important layer, and secondary layer to form a data block sequence and a transmission list file. The dynamic transmission scheduling module is used to transmit the transmission list file and data block sequence to the receiving end in sequence number. During the transmission process, the channel quality feature set is updated every 30 seconds and the data block size and concurrent transmission quantity are dynamically adjusted according to the bandwidth change ratio. The dequantization decoding module is used to perform dequantization reconstruction on the data blocks that arrive successively. It extracts the quantization parameters from the header of the data block, performs dequantization calculation on the quantization index to obtain the restored parameter values, and fills them into the corresponding parameter layer group storage area according to the parameter position index. After all data blocks are dequantized, the model parameters are reassembled to generate a complete model file.

Images