Unverified50% confidenceSolutionExact time
The original team solved the throughput decay caused by KV block fragmentation by fixing --max-num-seqs 32 and recycling worker nodes based on an RSS slope threshold (rather than a fixed cron).
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/18/2026
First Seen
Valid until: 1/16/2027
Sources
Related Entities
Related Claims
UnverifiedBoth position interpolation and NTK-aware base can restore the phase angle of the slowest frequency pair exactly to the 0.162 rad it had during training.63% similarUnverifiedIn a needle-in-a-haystack experiment with a 128K Token context and about 16,000 key-value pairs, changing the position of just one key-value pair produced a maximum accuracy gap of 40.2% for DeepSeek-V4-Flash-Base, 34.8% for DeepSeek-V4-Pro-Base, and still 19.1% and 14.8% respectively after post-training optimization.59% similarUnverifiedQ4_K_M采用分组量化策略，将权重张量划分为64或256维子块并独立计算缩放因子与零点，能将困惑度损失控制在1–3%以内56% similarUnverifiedOn the Erdős minimum overlap problem, the TTT-Discover recipe of "eight groups × 64 rollouts plus one LoRA update" reached a certified C₅ upper bound of 0.38094 within 25 steps.55% similarUnverifiedKV Cache address calculation example: when position=31, logic_block=31//16=1, offset=31%16=15; if physical_block=12, then slot_id=12*16+15=207.55% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1063648API
curl https://kongchang.com/api/v1/knowledge/claims/1063648MCP
get_claim(id=1063648)