b12207

<details open> ggml-cuda：为SSM_SCAN内核添加d_state=64快速路径（#31818） Zamba2及类似架构在循环扫描步骤中使用了64大小的状态，这超出了GPU内核现有快速路径（96、128、256）的范围。每个扫描步骤都不得不回退到CPU执行，导致生成速度受限，仅为每秒6-7个token，而其他架构在同一硬件上可达到每秒50-70个token的水平。 按照现有模式，新增针对d_state=64的第4种模板实例化；底层数学运算未作更改。已通过现有CPU参考正确性测试框架验证，并基于Zamba2的真实超参数构建了3个新测试用例（全部通过，28/28）。 在Zamba2-7B-Instruct上测得吞吐量提升约35%（从11提高到15.1 token/s）。尚未与其他架构实现完全对等——仍有约28%的扫描调用因原因未明而回退到CPU执行。 协助者：Claude Sonnet 5 共同作者：Sargentmajord <331229907+sargentmajord@users.noreply.github.com> </details> **网站：** - <https://llama.app> **证明文件：** - <https://github.com/ggml-org/llama.cpp/attestations/65059042> **macOS/iOS：** - [macOS Apple Silicon (arm64)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-macos-arm64.tar.gz) - macOS Apple Silicon (arm64, 启用KleidiAI) [已禁用](https://github.com/ggml-org/llama.cpp/pull/23780) - [macOS Intel (x64)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-macos-x64.tar.gz) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-xcframework.zip) **Linux：** - [Ubuntu x64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-ubuntu-x64.tar.gz) - [Ubuntu arm64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-ubuntu-arm64.tar.gz) - [Ubuntu 24.04 s390x (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-ubuntu-s390x.tar.gz) - [Ubuntu 26.04 s390x (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-ubuntu-26.04-s390x.tar.gz) - [Ubuntu 26.04 s390x (zDNN)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-ubuntu-zdnn-s390x.tar.gz) - [Ubuntu x64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-ubuntu-vulkan-x64.tar.gz) - [Ubuntu arm64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-ubuntu-vulkan-arm64.tar.gz) - [Ubuntu x64 (CUDA 12)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-ubuntu-cuda-12.8-x64.tar.gz) - [CUDA 12.8库](https://github.com/ggml-org/llama.cpp/releases/download/b12207/cudart-llama-b12207-bin-ubuntu-cuda-12.8-x64.tar.gz) - [Ubuntu x64 (CUDA 13)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-ubuntu-cuda-13.5-x64.tar.gz) - [CUDA 13.5库](https://github.com/ggml-org/llama.cpp/releases/download/b12207/cudart-llama-b12207-bin-ubuntu-cuda-13.5-x64.tar.gz) - [Ubuntu arm64 (CUDA 13)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-ubuntu-cuda-13.5-arm64.tar.gz) - [CUDA 13.5库](https://github.com/ggml-org/llama.cpp/releases/download/b12207/cudart-llama-b12207-bin-ubuntu-cuda-13.5-arm64.tar.gz) - [Ubuntu x64 (ROCm 10.0)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-ubuntu-rocm-10.0-x64.tar.gz) - [Ubuntu x64 (OpenVINO)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-ubuntu-openvino-2026.4.1-x64.tar.gz) - [Ubuntu x64 (SYCL FP32)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-ubuntu-sycl-fp32-x64.tar.gz) - [Ubuntu x64 (SYCL FP16)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-ubuntu-sycl-fp16-x64.tar.gz) - [Linux arm64 (Snapdragon：CPU、Adreno GPU、Hexagon NPU)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-linux-arm64-snapdragon.tar.gz) - [设置指南](https://github.com/ggml-org/llama.cpp/blob/master/docs/backend/snapdragon/linux.md) **Android：** - [Android arm64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-android-arm64.tar.gz) - [Android arm64 (Snapdragon：CPU、Adreno GPU、Hexagon NPU)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-android-arm64-snapdragon.tar.gz) - [设置指南](https://github.com/ggml-org/llama.cpp/blob/master/docs/backend/snapdragon/README.md) **Windows：** - [Windows x64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-win-cpu-x64.zip) - [Windows arm64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-win-cpu-arm64.zip) - [Windows arm64 (OpenCL Adreno)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-win-opencl-adreno-arm64.zip) - [Windows x64 (CUDA 12)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-win-cuda-12.4-x64.zip) - [CUDA 12.4 DLLs](https://github.com/ggml-org/llama.cpp/releases/download/b12207/cudart-llama-bin-win-cuda-12.4-x64.zip) - [Windows x64 (CUDA 13)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-win-cuda-13.5-x64.zip) - [CUDA 13.5 DLLs](https://github.com/ggml-org/llama.cpp/releases/download/b12207/cudart-llama-bin-win-cuda-13.5-x64.zip) - [Windows arm64 (CUDA 13)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-win-cuda-13.5-arm64.zip) - [CUDA 13.5 DLLs](https://github.com/ggml-org/llama.cpp/releases/download/b12207/cudart-llama-bin-win-cuda-13.5-arm64.zip) - [Windows x64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-win-vulkan-x64.zip) - [Windows arm64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-win-vulkan-arm64.zip) - [Windows x64 (OpenVINO)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-win-openvino-2026.4.1-x64.zip) - [Windows x64 (SYCL)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-win-sycl-x64.zip) - [Windows x64 (ROCm 10.0)](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-bin-win-rocm-10.0-x64.zip) **openEuler：** - [已禁用](https://github.com/ggml-org/llama.cpp/pull/23705) - openEuler x86 (310p) - openEuler x86 (910b，ACL Graph) - openEuler aarch64 (310p) - openEuler aarch64 (910b，ACL Graph) **UI：** - [UI](https://github.com/ggml-org/llama.cpp/releases/download/b12207/llama-b12207-ui.tar.gz)