Unverified50% confidenceBenchmarkExact time
MultiAlloc splits a fixed quantization budget into two parts, "per-layer precision improvement" and "per-token retention," replacing uniform compression with lightweight offline profiling, and under an equal memory budget it outperforms single-granularity baselines on Mistral, Llama, and QwQ.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedDynamic量化通过敏感度分析为每一层甚至每个权重矩阵分配不同的比特预算，采用混合精度策略在总平均比特数不变的条件下降低整体量化误差68% similarUnverifiedQSA reselects important context at micro-block granularity through a compressed lightweight index, reducing the computational overhead of long-sequence Attention.66% similarVerifiedModern quantization schemes such as GPTQ, AWQ, and the grouped quantization of the GGUF format can reduce model size by more than 75% while retaining the vast majority of reasoning capability through layer-by-layer calibration and outlier protection.66% similarUnverifiedA common practice is to use bounding boxes for major categories and segmentation masks for fine-grained categories with frequent occlusion, in order to balance cost and accuracy.65% similarUnverifiedCompared to full fine-tuning, LoRA can compress the required VRAM and time costs by one to two orders of magnitude.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1026261API
curl https://kongchang.com/api/v1/knowledge/claims/1026261MCP
get_claim(id=1026261)