Unverified50% confidenceFactExact time
The architecture uses a single shared KV head (num_key_value_heads=1, i.e. MQA form), with head_dim of 512 to compensate for expressiveness.
1
Sources
50%
Confidence
Long-term
Relevance
10/18/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn llama.cpp, --n-gpu-layers N means putting the N layers of the model closest to the input side on the GPU for computation, with the remaining layers falling back to the CPU.60% similarUnverifiedDSpark 将马尔可夫头（Markov Head）设为默认核心架构，参数量通常不超过原模型的0.1%59% similarUnverifiedHQC-256使用n=57637，其密钥封装核心操作是在GF(2)[x]/(x^n - 1)环上进行多项式乘法59% similarUnverifiedKV Cache stores the already-computed Key and Value matrices in GPU VRAM, reducing generation complexity from O(n²) to O(n).59% similarUnverifiedP106-100基于Pascal架构（GP106芯片），与GTX 1060共享相同的计算核心59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1061804API
curl https://kongchang.com/api/v1/knowledge/claims/1061804MCP
get_claim(id=1061804)