lordx64/phantom-kv
Refusal removal as a loadable KV-cache graft — zero weight modification, no refusal-direction projection, fully reversible.
A star delta has been computed for 61,707 of 62,640 tracked repositories in this period, and only those can be ranked in it.
A repository missing here has not been measured far enough back yet. That does not mean it did not move.
Star counts measured between 00:38 and 02:34 (BKK).
Refusal removal as a loadable KV-cache graft — zero weight modification, no refusal-direction projection, fully reversible.
Serve large Qwen models fast on the GPUs you actually own. Qwen3.8-27B on a single 24 GB card with vLLM: 127 tok/s single-user (381 when the answer quotes the prompt), ~1,035 tok/s at 64 concurrent, 150k-262k context. vLLM patches, requant pipeline, benchmarks.
LMCache: Supercharge Your LLM with the Fastest KV Cache Layer
[ICLR'26] The official code implementation for "Cache-to-Cache: Direct Semantic Communication Between Large Language Models"
LLM KV cache compression made easy
High-performance KV cache storage for LLM inference — GPU offloading, SSD caching, and cross-node sharing via RDMA. Works with vLLM and SGLang.
Pure Rust + CUDA LLM inference engine — no PyTorch, OpenAI-compatible, serves Qwen3 to Kimi-K2
Learn DeepInfer through a small, runnable inference engine. Follow the code on CPU, then run real Qwen3.5 models on Ascend 910C.
Alibaba Cloud's high-performance KVCache system for LLM inference, with components for global cache management, inference simulation(HiSim), and more.
Lossless AI compression: 33% less GPU memory, bit for bit.
Official Pytorch Code of the Paper "FashionChameleon: Towards Real-Time and Interactive Human-Garment Video Customization"
Heterogeneous prefill/decode for DeepSeek-V4-Flash: CUDA prefill (DGX Spark, vLLM) -> Metal decode (Mac Studio, oMLX) over plain 10GbE
High Performance KV Cache for LLM Inference
That's all 13 repositories in this period.