Unverified50% confidenceFactExact time
GPTQ通过Hessian矩阵的近似信息来逐层最小化量化误差
1
Sources
50%
Confidence
Long-term
Relevance
8/31/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedGPTQ引入二阶Hessian矩阵衡量权重重要性，AWQ通过分析激活值分布保护显著通道权重76% similarUnverified量化主流格式包括GPTQ、AWQ、GGUF等，INT4量化相比FP16通常导致模型在复杂推理、数学和代码任务上出现明显退化65% similarUnverified当前主流的4-bit量化方案（如GPTQ、AWQ）普遍采用group size=128的分组量化策略60% similarUnverified倒排乘积量化（IVFPQ）在PQ基础上引入倒排索引结构，通过聚类将向量库分区，将检索复杂度从线性降低至亚线性60% similarUnverifiedGGUF格式中的Q4_K_M采用分组量化（Group Quantization）策略，每组独立计算缩放因子和零点59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/828396API
curl https://kongchang.com/api/v1/knowledge/claims/828396MCP
get_claim(id=828396)