Abnormal value-based weight activation quantification method after large model training

CN119005258APending Publication Date: 2024-11-22THE ACAD OF TIANJIN UNIV HEFEI
View PDF 0 Cites 5 Cited by

Patent Information

Application Number
CN202411142408.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-20
Publication Date
2024-11-22

Smart Images

  • Figure CN119005258A_ABST
    Figure CN119005258A_ABST
Patent Text Reader

Abstract

The invention discloses a weight activation quantification method based on abnormal value large model training. The weight activation quantification method comprises a weight quantification process and an activation value quantification process during reasoning. In weight quantization, model bit width is converted into low bit representation in an off-line mode, and in weight activation value quantization during reasoning, a quantized weight model is initialized, input of a linear layer is pre-scaled during reasoning operation, and finally 8-bit quantization operation is executed. According to the method, mathematical equality transformation of weight and activation is combined, a strategy of lossless quantization and low-delay execution is provided, and the significant weight is considered in an off-line weight quantization process, so that the influence of a smooth activation abnormal value on the significant weight is reduced, and the quantization error of the abnormal value is reduced in a model reasoning process. According to the method, effective quantification during large-scale model reasoning is realized, the memory required for running a large model is remarkably reduced, and the running speed is increased.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of model weight activation quantization methods, and in particular to a large model post-training weight activation quantization method based on outliers. Background Art

[0002] Large-scale models such as GPT-4 and LLAMA have achieved significant technological breakthroughs in the field of natural language processing, with the number of model parameters increasing from 1.5 billion in the past to hundreds of billions. The substantial increase in parameters has greatly enhanced the model's ability to handle complex tasks. However, the huge computing resources and communication cost requirements have become obstacles to its actual deployment in real-world applications.

[0003] In the large model inference generation scenario, the data generated by the trained large model occupies the video memory of the large model parameters (i.e. weights), and the activation value is the intermediate calculation result (i.e. activation value) generated when the large model is inferred and generated. Since the video memory occupied by the activation value increases with the model input length, in many scenarios, the video memory occupied by the activation value is the same as the video memory occupied by the weight. Therefore, in order to reduce the video memory occupied during the large model inference generation, weight activation quantization is required.

[0004] In large model compression technology, quantization methods have been proven to be an effective means of reducing computing and storage burdens. Model quantization is divided into two types: training-aware quantization (QAT) and post-training quantization (PTQ). Although QAT has advantages over PTQ in terms of accuracy, its high cost during training is prohibitive, so PTQ has become a more popular quantization strategy. PTQ is divided into two strategies: weight quantization and weight activation quantization. Although simple weight quantization can be reduced to a bit width of 4 bits or even lower, its effect is poor and it cannot reduce memory consumption during inference. For this reason, how to improve the weight activation quantization scheme has become the key to reducing memory overhead and improving inference efficiency when deploying large models.

[0005] The main challenge of current weight activation quantization lies in the errors caused by the round() function and the clamp() function. To address this problem, research focuses on reducing errors by reducing outliers in the inference process. The mainstream methods are divided into two categories: one is based on mathematical equation transformation, such as SmoothQuant, Ominiquant, Outlier Suppression+ and other solutions. Their core is to simplify the quantization process of the model through mathematical equivalent transformation without losing model accuracy, thereby smoothing outliers in the activation value, reducing the stretching of the quantization range, and ensuring that most values ​​maintain effective quantization bits; the other is based on outlier-repartitioned channel quantization methods, such as RPTQ, QLLM, etc. These methods optimize the quantization process by identifying and processing significant differences in the range of activation values ​​in different channels, so that channels with similar numerical ranges can be grouped and quantized, or channels containing outliers can be effectively repartitioned, thereby significantly reducing the quantization error caused by outliers, and further refining and optimizing the model's quantization strategy.

[0006] The commonly used weight activation quantization method based on mathematical equation transformation and outlier re-division of channels, although in theory it has proposed potential optimization directions for the computation and storage efficiency of large models, has failed to meet the requirements of high precision and low latency when actually deploying models on the end. Specifically, although the quantization method based on mathematical equation transformation shifts the difficulty of quantization from activation values ​​to weights, this method does not fully consider the significant differences in weights, thereby introducing additional errors in the weight quantization process and affecting the overall accuracy of the model. In addition, although the method based on outlier re-division of channels attempts to handle outliers more effectively by reordering channels, the time complexity of the method used is high, which increases latency. Therefore, the existing weight activation quantization strategy is not effective in the actual end-side deployment of large models, and there is an urgent need to develop a new weight activation quantization strategy. Summary of the invention

[0007] The present invention provides a weight activation quantization method based on outlier large model training to solve the problems of severe precision loss and high latency in existing weight activation quantization schemes, thereby reducing the computational and storage burdens while maintaining high precision and low latency.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is:

[0009] The weight activation quantization method after training of a large model based on outliers first performs weight quantization, which includes the following steps:

[0010] Step 1: Activate outlier smoothing. The process is as follows:

[0011] Input the calibration data set into the trained large model for inference. During the inference process, the activation outliers of each layer in the large model, i.e., the maximum activation value, and the maximum weight of each linear layer in the large model are collected.

[0012] The smoothing scaling factor is calculated based on the maximum activation value of each layer and the maximum weight value of each linear layer, and then the weight of each linear layer in the trained large model is smoothly adjusted according to the smoothing scaling factor to obtain the large model after the linear layer weight is smoothly adjusted;

[0013] Step 2: Reduce the impact of activation smoothing on weight quantization. The process is as follows:

[0014] The calibration data set is input into the large model after the linear layer weight smoothing adjustment obtained in step 1 for inference again. During the inference process, the maximum activation value, the minimum activation value, and the average activation value of each layer in the large model are collected;

[0015] The weight adjustment scaling factor of each layer is calculated based on the maximum activation value, the minimum activation value, and the average activation value, and then the weight of each layer in the large model obtained in step 1 is further adjusted based on the weight adjustment scaling factor to obtain the large model after the weight of each layer is adjusted;

[0016] Step 3, traverse all the named modules of the large model after each layer of weight adjustment obtained in step 2, collect the weight value of the one with the largest weight in each layer of each module as the maximum weight of the module, and the weight value of the one with the smallest weight as the minimum weight of the module, and calculate the weight quantization scaling factor of the corresponding module based on the maximum weight and minimum weight of each module; then, based on the weight quantization scaling factor of each module, quantize the weights of each layer in the module respectively, obtain the quantized weight values ​​of each layer in each module, and save the quantized weight values ​​as integer 8-bit data type data;

[0017] Then, the activation value is quantized. The activation value quantization process is as follows:

[0018] Based on the weight quantization scaling factor, the quantized weight values ​​of each layer of each module in the trained large model are dequantized to obtain the weight values ​​of each layer of each module in an integer 8-bit data type;

[0019] In addition, a calibration data set is input into the trained large model for inference, and the input tensor of each linear layer in the large model, that is, the largest one of the activation values ​​is obtained as the maximum activation value, and the second largest one is obtained as the second largest activation value; based on the maximum activation value and the second largest activation value, the outlier mitigation scaling factor and the quantization scaling factor are calculated respectively; then, based on the outlier mitigation scaling factor and the quantization scaling factor, the activation value of each linear layer is quantized to obtain the activation quantization value, and the activation quantization value is stored as data of integer 8-bit data type; thereby, the quantization of the activation value is completed.

[0020] Furthermore, in step 2 of weight quantization, a greedy algorithm is used to calculate the weight adjustment scaling factor of each layer based on the maximum activation value, the minimum activation value, and the average activation value.

[0021] Furthermore, in the process of activation value quantization, the TOPK algorithm is used to obtain the maximum activation value and the second largest activation value.

[0022] Furthermore, for the softmax layer with self-attention mechanism, the activation value quantization process also includes:

[0023] The activation quantization value of each linear layer of the trained large model with an integer 8-bit data type is matrix multiplied with the corresponding weight value of the integer 8-bit data type to obtain a result of the integer 8-bit data type, and the multiplication result is dequantized into a floating-point 16-bit data type, thereby completing the activation value quantization.

[0024] Based on the observation that there is a significant difference between the maximum and the second largest value in the activation channel, the present invention combines the transformation of the mathematical equations of weights and activations to propose a strategy for lossless quantization and low-latency execution. This weight activation strategy takes into account significant weights in the offline quantization weight process, thereby reducing the impact of smooth activation outliers on significant weights. During the model inference process, this strategy optimizes the quantization process of activation values ​​according to the distribution characteristics of the activation channel and reduces the quantization error of outliers. This strategy achieves effective quantization during large-scale model inference, significantly reduces the memory required to run large models, improves the running speed, and increases the feasibility of deploying large models on edge devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 2 is a schematic diagram of an embodiment of the present invention.

[0026] Figure 2 A diagram of the large model structure used in an embodiment of the present invention. DETAILED DESCRIPTION

[0027] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0028] This embodiment discloses a method for quantizing weights after training a large model and a method for quantizing weight activations during inference, which greatly reduces the accuracy loss by scaling the maximum value in each channel.

[0029] like Figure 1As shown, the method process of this embodiment includes two major modules: a weight quantization module and a weight activation quantization module during inference. In the method process of this embodiment, weight quantization is first performed through the weight quantization module, and then the activation value is quantized through the weight activation quantization module. The weight quantization module is responsible for converting the model bit width offline to a low-bit representation, while the weight activation quantization module during inference is responsible for initializing the quantized weight model, pre-scaling the input of the linear layer during the inference run, and finally performing an 8-bit quantization operation. Among them, the weight quantization module consists of the following three parts:

[0030] (1) Activation outlier smoothing to weight mechanism: read the model parameter file, input the calibration data set for the model inference process, collect the maximum activation value and the maximum weight value during the inference process channel by channel, and perform mathematical equation transformation to make the activation distribution smoother.

[0031] (2) Weight quantization to low-bit mechanism: The calibration dataset is input again for model inference, and a greedy algorithm is used to calculate the pre-scaling factor of the weight according to the range of activation values ​​to alleviate the negative impact of the smoothing mechanism on the salient weights, and then all weights are quantized layer by layer.

[0032] (3) Model weight preservation mechanism: Use the integer 8-bit (int8) format to package and store the quantized weights to ensure data compactness and reproducibility.

[0033] The weight activation quantization module at inference time consists of the following three parts:

[0034] (1) Weight dequantization mechanism: load the quantized weights, replace the linear layer of the model, and dequantize the loaded weights to 8-bit precision.

[0035] (2) Activation value quantization mechanism: During the linear layer processing, the maximum value of the input channel is pre-scaled to the second largest value. The pre-scaled output is low-bit quantized using a channel-by-channel quantization strategy. Then it is inversely pre-scaled to the maximum value.

[0036] (3) 8-bit matrix multiplication operator: Use cutlass to perform matrix multiplication on the quantized input and the 8-bit weight.

[0037] In the activation quantization module during inference, considering that the self-attention mechanism in large models has higher requirements for data accuracy, it is necessary to use the activation value dequantization mechanism to dequantize the 8-bit value to float16 for self-attention calculation.

[0038] The following embodiment is combined with Figure 2The Llama2-7B large model shown in the figure further illustrates the method of this embodiment. The Llama2-7B large model consists of 32 neural network layers, each of which has a decoder for self-attention mechanism and a multi-layer perceptron MLP to perform nonlinear transformation and feature enhancement on the output of the decoder. Therefore, the initialization of the Llama2-7B large model requires initialization weights, as well as the number of model layers, the composition of the decoder and the multi-layer perceptron MLP.

[0039] In this embodiment, the process of weight quantization by the weight quantization module is as follows:

[0040] Step 1: Activate outlier smoothing. The process is as follows:

[0041] Firstly, the self-attention mechanism and the linear layer of the multi-layer perceptron (MLP) in the trained Llama2-7B large model are preprocessed. The named modules of the Llama2-7B large model are traversed to extract the normalized weights of the input layer and the corresponding Q, K, and V linear layer weights of each module.

[0042] In addition, a hook function is registered to each layer of the trained Llama2-7B large model. The function of this hook function is to find the maximum value of the activation value of each layer by channel.

[0043] Then, the calibration data set is input into the trained Llama2-7B large model for inference. During the inference process, the hook function is triggered to collect the activation outliers of each layer in the Llama2-7B large model, that is, the activation maximum value act max , and the maximum weight of each linear layer in the large model max .

[0044] Next, based on the maximum activation value act of each layer max , the maximum weight of each linear layer max Calculate the smoothing scaling factor scale smooth , smoothing scaling factor scale smooth The calculation formula is as follows:

[0045]

[0046] Where α is a predefined smoothing coefficient used to adjust the influence ratio of the maximum activation value to the maximum weight value.

[0047] Finally, according to the smoothing scaling factor scale smoothThe weight of each linear layer in the trained Llama2-7B large model is smoothly adjusted to obtain the Llama2-7B large model after the linear layer weight is smoothly adjusted. Specifically, for the self-attention mechanism in the Llama2-7B large model, the normalized weight of the input layer of the self-attention mechanism is divided by the smoothing scaling factor scale smooth , and multiply the weights of the Q, K, and V linear layers by the same smooth scaling factor scale smooth ; For the MLP part, perform similar operations and divide the normalized weights of the MLP input layer by the smoothing scaling factor scale smooth And multiply the weights of the gate, up, and down linear layers in the MLP by the same smooth scaling factor scale smooth .

[0048] Step 2: Reduce the impact of activation smoothing on weight quantization. The process is as follows:

[0049] The hook function is registered again to each layer of the Llama2-7B large model after the linear layer weight adjustment obtained in step 1, and the calibration data set is input into the Llama2-7B large model after the linear layer weight adjustment obtained in step 1 for inference. During the inference process, the maximum activation value act of each layer in the Llama2-7B large model is collected through the hook function. max , activation minimum value act min and the activation mean act mean .

[0050] Then use the greedy algorithm and set a preset exponent α to balance the influence of the maximum and minimum activation values. max , activation minimum value act min and the activation mean act mean Calculate the weight adjustment scaling factor of each layer awq , weight adjustment scaling factor scale awq The calculation formula is as follows:

[0051]

[0052] Finally, the weight of each layer in the Llama2-7B large model obtained in step 1 is further adjusted based on the weight adjustment scaling factor to obtain the Llama2-7B large model after weight adjustment of each layer. Specifically, the weight in each layer is divided by the weight adjustment scaling factor scale awq In the Llama2-7B large model, this step mainly adjusts the weights of the Q, K, and V layers of the self-attention mechanism and the weights of the multi-layer perceptron MLP.

[0053] Step 3: Traverse all named modules of the Llama2-7B large model after the weight adjustment of each layer obtained in step 2, and collect the weight value of the largest weight in each layer of each module as the maximum weight of the module max The weight value of the one with the smallest weight is taken as the minimum weight of the module. min .

[0054] Then, based on the maximum weight of each module max , minimum weight weight min Calculate the weight quantization scaling factor of the corresponding module weight , when calculating the quantization scaling factor, it is necessary to calculate it according to the bit width of the quantized value. The calculation formula is as follows:

[0055]

[0056] Next, the weight quantization scaling factor is scaled based on each module weight , quantize the weights of each layer in the module, and obtain the quantized weight values ​​of each layer in each module. Specifically, divide the weight of each layer by the weight quantization scaling factor scale weight , that is, the quantized weight value is obtained. In this embodiment, the bit width of the quantized value is 4 bits.

[0057] Finally, the quantized weight values ​​are saved as integer 8-bit data types (int8). Bit mask operations and right shift operations are used on the int8 data to extract the lower 4 bits and upper 4 bits. Every two 4-bit quantized weight values ​​are merged into one int8. Finally, the dimension of each linear layer is modified, and then the model is saved locally, thereby completing the weight quantization of the trained Llama2-7B large model.

[0058] In this embodiment, the process of quantizing the activation value by the weighted activation quantization module is as follows:

[0059] Initialize the weights, model layers, and dimensions of each weight of the trained Llama2-7B large model in int8 format, extract the low and high four bits of the int8 data, and decompose an int8 value into two 4-bit quantized weights.

[0060] Then, based on the weight quantization scaling factor scale weight , dequantize the quantized weight values ​​of each module and layer in the trained Llama2-7B large model to obtain the weight values ​​of each module and layer of integer 8-bit data type weight 8bits , the inverse quantization formula is as follows:

[0061]

[0062] In addition, the calibration data set is input into the trained Llama2-7B large model for inference, so that each layer in the Llama2-7B large model generates input tensors, i.e., activation values, and quantizes the input tensors, i.e., activation values, of each linear layer in the Llama2-7B large model, including the self-attention mechanism and the linear layer of the multi-layer perceptron MLP. Specifically:

[0063] First, the TOPK algorithm is used to obtain the input tensor of each linear layer in the Llama2-7B large model, that is, the largest one among the activation values ​​as the maximum activation value act max , the second largest one is used as the second largest value act to activate second .

[0064] Then, based on the maximum activation act max and activate the second largest value act second , respectively calculate the outlier reduction scaling factor scale second And the scaling factor scale when quantizing act , the calculation formula is as follows:

[0065]

[0066]

[0067] Next, we reduce the scaling factor based on the outliers. second And the scaling factor scale when quantizing act , quantize the activation value of each linear layer to obtain the activation quantization value. Specifically, firstly, the maximum activation value act max Divide by the outlier to mitigate the scaling factor scale second , to reduce the error caused by outliers in the tensor to the subsequent quantization, then divide all activation values ​​including the maximum activation value by the quantization scaling factor scale act , thus completing the quantization of the activation values ​​of each linear layer.

[0068] Finally, the activation quantization value is stored as an integer 8-bit data type, and the resulting act max Multiply it by the outlier to reduce the scaling factor scale second , to restore the outliers in the corresponding input tensor.

[0069] For the softmax layer with a self-attention mechanism, in this embodiment, the process of quantizing the activation value through the weight activation quantization module also includes:

[0070] After weight dequantization and activation value quantization, the activation quantization value and weight are now both int8 data types. Therefore, in this embodiment, the activation quantization value of each linear layer in the trained Llama2-7B large model obtained in step S1 is respectively subjected to matrix multiplication with the corresponding weight value of the integer 8-bit data type to obtain the result of the integer 8-bit data type, and the multiplication result is dequantized to the floating point 16-bit data type, thereby completing the activation value quantization.

[0071] For the softmax layer of the self-attention mechanism in the Llama2-7B large model, since the softmax of the self-attention softmax layer is an exponential function, the calculation of the self-attention softmax layer s requires higher numerical accuracy, and the input of the self-attention calculation is the output product of the Q and K linear layers, so the input of the softmax needs to be a high-precision value. Therefore, this embodiment needs to multiply the activation quantization value of the linear layer by the corresponding integer 8-bit data type weight value and then dequantize it to a dequantized activation value act of a floating point 16-bit data type (float16). fp16 , the inverse quantization calculation formula is as follows:

[0072] act fp16 = float16(act int8 *scale act *scale weight *(2 4 -1)) / 127

[0073] Among them, act int8 The result of multiplying the activation quantization value of the linear layer by the corresponding integer 8-bit data type weight value, scale weight Scale the previous weight quantization scaling factor weight , scale act Scaling factor when quantization is activated.

[0074] After the dequantization is completed, the dequantized activation value product Z tensor corresponding to the output of the Q and K linear layers is calculated, the maximum value D in each channel of the tensor Z is calculated, and the softmax layer of the self-attention mechanism is calculated based on the maximum value D and Z, thereby obtaining the probability of generating the next token. The calculation formula of the softmax layer is as follows:

[0075]

[0076] Where: z i is a value in the tensor Z. cis a value in the tensor Z. C is the size of the last dimension in the tensor Z. dim(-1) is the token-by-token calculation of softmax, that is, the calculation is based on the granularity of the last dimension.

[0077] After each layer of the Llama2-7B model is self-attentioned, it is then passed through a multi-layer perceptron for nonlinear transformation and feature enhancement. Finally, an output tensor is obtained, the size of the tensor is [1, vocab size], the index in the tensor represents the token id, and the value represents the probability of generating the token.

[0078] The preferred embodiments of the present invention are described in detail above in conjunction with the accompanying drawings. The embodiments described in the present invention are merely descriptions of the preferred embodiments of the present invention, and do not limit the concept and scope of the present invention. The various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction, and such combinations should also be regarded as the contents disclosed in the present disclosure as long as they do not violate the concept of the present invention. In order to avoid unnecessary repetition, the present invention will not further describe various possible combinations.

[0079] The present invention is not limited to the specific details of the above-mentioned embodiments. Within the technical concept of the present invention and without departing from the design concept of the present invention, various modifications and improvements made to the technical solution of the present invention by technical personnel in this field should fall within the protection scope of the present invention. The technical contents for which protection is sought in the present invention have been fully recorded in the claims.

Claims

1. A method for quantizing weight activations after training a large model based on outliers, characterized in that: First, weight quantization is performed, which includes the following steps: Step 1: Activate outlier smoothing. The process is as follows: Input the calibration data set into the trained large model for inference. During the inference process, the activation outliers of each layer in the large model, i.e., the maximum activation value, and the maximum weight of each linear layer in the large model are collected. The smoothing scaling factor is calculated based on the maximum activation value of each layer and the maximum weight value of each linear layer, and then the weight of each linear layer in the trained large model is smoothly adjusted according to the smoothing scaling factor to obtain the large model after the linear layer weight is smoothly adjusted; Step 2: Reduce the impact of activation smoothing on weight quantization. The process is as follows: The calibration data set is input into the large model after the linear layer weight smoothing adjustment obtained in step 1 for inference again. During the inference process, the maximum activation value, the minimum activation value, and the average activation value of each layer in the large model are collected; The weight adjustment scaling factor of each layer is calculated based on the maximum activation value, the minimum activation value, and the average activation value, and then the weight of each layer in the large model obtained in step 1 is further adjusted based on the weight adjustment scaling factor to obtain the large model after the weight of each layer is adjusted; Step 3, traverse all the named modules of the large model after each layer of weight adjustment obtained in step 2, collect the weight value of the one with the largest weight in each layer of each module as the maximum weight of the module, and the weight value of the one with the smallest weight as the minimum weight of the module, and calculate the weight quantization scaling factor of the corresponding module based on the maximum weight and minimum weight of each module; then, based on the weight quantization scaling factor of each module, quantize the weights of each layer in the module respectively, obtain the quantized weight values ​​of each layer in each module, and save the quantized weight values ​​as integer 8-bit data type data; Then, the activation value is quantized. The activation value quantization process is as follows: Based on the weight quantization scaling factor, the quantized weight values ​​of each layer of each module in the trained large model are dequantized to obtain the weight values ​​of each layer of each module in an integer 8-bit data type; In addition, a calibration data set is input into the trained large model for inference, and the input tensor of each linear layer in the large model, that is, the largest one of the activation values ​​is obtained as the maximum activation value, and the second largest one is obtained as the second largest activation value; based on the maximum activation value and the second largest activation value, the outlier mitigation scaling factor and the quantization scaling factor are calculated respectively; then, based on the outlier mitigation scaling factor and the quantization scaling factor, the activation value of each linear layer is quantized to obtain the activation quantization value, and the activation quantization value is stored as data of integer 8-bit data type; thereby, the quantization of the activation value is completed.

2. The outlier-based large model post-training weight activation quantization method according to claim 1, characterized in that: In step 2 of weight quantization, a greedy algorithm is used to calculate the weight adjustment scaling factor of each layer based on the maximum activation value, the minimum activation value, and the average activation value.

3. The outlier-based large model post-training weight activation quantization method according to claim 1, characterized in that: During the activation value quantization process, the TOPK algorithm is used to obtain the maximum activation value and the second largest activation value.

4. The outlier-based large model post-training weight activation quantization method according to claim 1, characterized in that: For the softmax layer with self-attention mechanism, the activation value quantization process also includes: The activation quantization value of each linear layer of the trained large model with an integer 8-bit data type is matrix multiplied with the corresponding weight value of the integer 8-bit data type to obtain a result of the integer 8-bit data type, and the multiplication result is dequantized into a floating-point 16-bit data type, thereby completing the activation value quantization.

Citation Information

Cited By

  • Quantitative reasoning acceleration method and system oriented to Grounding DINO

    CN119443167A

  • Neural network optimization method based on outlier separation activation value hybrid quantization

    CN119761429A

  • Model quantification method and device, related equipment and computer program product

    CN120012843A

  • Post-training quantization method suitable for autoregressive mask generation model

    CN121168554A

  • Ai model quantization method and apparatus, and related device

    WO2026148839A1