Unverified50% confidenceBenchmarkExact time
The original two-step allgather+scatter operation in TP LMHead has been merged into a single all-to-all communication, reducing LMHead latency from 320μs to 169μs and improving TPOT from 36.97ms to 35.67ms in DeepSeek-V4-Pro B200 decoding tests.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/5/2026
First Seen
Valid until: 12/4/2026
Sources
Related Entities
Related Claims
UnverifiedBM25 is a significant improvement over TF-IDF, introducing term frequency saturation and document length normalization64% similarUnverifiedWhen using FP16 precision, the KV Cache for LLaMA-2-7B occupies approximately 0.5 MB per token.63% similarUnverifiedMemStitch claims to bring up to a 25× speedup in TTFT for vLLM through "zero-copy context bridging" technology.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/860550API
curl https://kongchang.com/api/v1/knowledge/claims/860550MCP
get_claim(id=860550)