Unverified50% confidenceFactExact time
激活参数描述的是每个 token 的计算量，而非权重的存储需求
1
Sources
50%
Confidence
Long-term
Relevance
7/21/2026
First Seen
Sources
Related Claims
UnverifiedTaking a 7B parameter model as an example, each additional 1K tokens of context consumes approximately 0.5–1 GB of GPU memory.68% similarUnverified上下文越长，模型需要缓存的 KV Cache 越大，KV Cache 使模型生成新 Token 时无需重新计算所有历史 Token 的注意力权重61% similarUnverifiedFP16格式下，大模型显存需求的估算公式为：参数量 × 2 = 所需显存（GB），即每个参数以16位（2字节）浮点数存储61% similarUnverifiedFor a typical 7B-parameter model, the KV Cache per 1K tokens requires approximately 0.5–1 GB of VRAM.60% similarUnverified大模型调用的计费单位是Token而非请求次数，响应时间从毫秒级变为秒级甚至分钟级59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/577409API
curl https://kongchang.com/api/v1/knowledge/claims/577409MCP
get_claim(id=577409)