Similar to This
Previous

Add FP8 KV-cache support by ORippler · Pull Request #28392 · ggml-org/llama.cpp: LLM inference in C/C++. Contribute to ggml-org/llama.cpp development by creating ...
github.com/ggml-org/llama.cpp/pull/28392

qwen4exp : build the QSA selection without the kq_mask by sanmai · Pull Request #30965 · ggml-org/llama.cpp
github.com/ggml-org/llama.cpp/pull/30965

feat: import qwen4exp (Qwen3.8-Flash-Next) support from upstream PR #27742 by giveen · Pull Request #324 · TheTom/llama-cpp-turboquant
github.com/TheTom/llama-cpp-turboquant/pull/324
Add MiniMax-M3 (MSA: MiniMax Sparse Attention) support by timkhronos · Pull Request #24908 · ggml-org/llama.cpp
github.com/ggml-org/llama.cpp/pull/24908

4-bit key cache that scales each channel on its own (q4-size memory without the quality cliff plain q4_0 falls off on some models) · ggml-org/llama ...
github.com/ggml-org/llama.cpp/discussions/24518

kv-cache : avoid kv cells copies by ggerganov · Pull Request #24277 · ggml-org/llama.cpp
github.com/ggml-org/llama.cpp/pull/24277
![[PoC] server: support requantizing kv cache by wadealexc · Pull Request ...](https://external-preview.redd.it/Uwg2OuKYL_cyUNuDl_QbKE8QLJOJXgBPbR0_Jgg2eAo.png?width=320&crop=smart&auto=webp&s=e349140f70c63a9307fcb53420c133424145284f)
[PoC] server: support requantizing kv cache by wadealexc · Pull Request #24134 · ggml-org/llama.cpp
github.com/ggml-org/llama.cpp/pull/24134
Asymmetric KV q8/q4 cache to preserve precision while reducing memory usage · ggml-org/llama.cpp · Discussion #23470
github.com/ggml-org/llama.cpp/discussions/23470
GitHub - ParmesanParty/llama.cpp: Specialized fork for (relatively) fast single-GPU inference (in CUDA) using large MoE models that don't fit fully into ...
github.com/ParmesanParty/llama.cpp

llama : rotate activations for better quantization by ggerganov · Pull Request #21038 · ggml-org/llama.cpp
github.com/ggml-org/llama.cpp/pull/21038
Next

