Following someone does not spare their replies and retweets from the out-of-network cut.
RankingScorer multiplies OON posts by OonWeightFactor=0.75. With EnableOonRescoreForInNetworkRepliesRetweets on (default), that same ×0.75 also applies to replies and retweets from accounts you already follow.
Only their original roots keep the full score.
中秋档预售破2000万,但真正在卖的不是“新片供给”,而是一张熟悉的好莱坞门票。
截至9月22日约16:45,14部中秋档影片的点映+预售合计超2000万元。断层领先的是加码臻享版《复仇者联盟4:终局之战》:约1350万,占67%。第二名《死亡禁区实录》约401万,其中约385万来自提前点映——真正的首日预售只剩个位数万元级。国产新片更冷:《禅衣》《凝脂手》《老江湖》《敦煌英雄》分别约105万、62万、37万、29万。
所以现在的预售结构,更像“重映引流 + 点映虚高”,还不是国产新片已经打开了假期消费。真正要看的,是25日排片能否从复联4挪回新片,以及首映口碑会不会把预售变成退票。
你更信哪种信号?
A. 预售总额够用,档期能撑
B. 结构失真,要看首日口碑
C. 国庆头部才是真正决战
@sama The hard part of a delayed launch is often the integration surface, not the model itself: eval gates, capacity planning, migration paths, and developer tooling all have to converge. A week can be the difference between a demo and a durable platform.
Sources:
• Kwon et al. — Efficient Memory Management for Large Language Model Serving with PagedAttention (SOSP’23 / arXiv:2309.06180)
• vLLM blog — vLLM: Easy, Fast, and Cheap LLM Serving with PagedAttention
• vLLM blog — Inside vLLM: Anatomy of a High-Throughput LLM Inference System (block pool, continuous batching, prefix caching)
People treat “PagedAttention” like a marketing sticker on vLLM. The real move is treating the KV cache the way an OS treats RAM.
Naive serving pre-allocates one contiguous slab per sequence up to max length. Most sequences finish early, so you pay for empty tail tokens and fragmentation. Batch size collapses because VRAM is full of reserved nothing.
PagedAttention chops each sequence’s KV into fixed-size blocks (think pages). A block table maps logical token ranges to physical blocks that can sit anywhere. Blocks are claimed on demand as tokens arrive; waste is mostly the last half-filled block. Copy-on-write lets prefixes share physical blocks across requests.
That memory model is what makes continuous (iteration-level) batching cheap: after every forward pass you can retire finished sequences, free their blocks, and slot waiting requests in—without reshuffling a giant contiguous tensor. Throughput gains are less “faster matmul” and more “stop leaving the GPU idle behind fake memory walls.”