neuron-explorer
Steering-first neuron explorer for Qwen3.5-4B: push every MLP neuron, keep every rollout, browse the results
翻译准备中
尚未检测到更新。GitHub 发布版本后会显示在这里。
个人 GitHub 账号下公开仓库的所有者可以登录并验证所有权,自行管理介绍,无需管理员批准。
登录并生成验证文件,在 30 分钟内将其提交到此仓库的默认分支。其他分支或分叉仓库中的文件无效。然后在下方检查文件,即可无需管理员批准直接管理项目。请保留文件以维持管理权限。
如需登录,将前往登录页面。登录后请回到此处,再次点击所选的验证按钮。
在项目文件夹中运行 /nomorevibe verify,即可验证所有权并自行管理介绍。
一分钟完成验证用 Rust 编写的推理引擎,支持 GGUF、AWQ 和 FP8(W8A8)检查点,使用手写的 CUDA 和 Metal 内核——不使用 PyTorch、不使用 libtorch、也不使用 ggml
AI 翻译
An inference engine written in Rust for GGUF, AWQ, and FP8 (W8A8) checkpoints, with hand-written CUDA and Metal kernels — no PyTorch, no libtorch, no ggml
取自仓库 README
正在加载完整 README。 以下为已保存的摘录。
An inference engine written in Rust for GGUF, AWQ, native FP8 (W8A8), and
native NVFP4 (W4A4) checkpoints, with hand-written kernels for both NVIDIA
(CUDA) and Apple Silicon (Metal) GPUs — no PyTorch, no libtorch, no ggml.
CUDA is the primary, most complete backend; Metal covers dense decode, GQA
attention and GatedDeltaNet today, with MoE, the vision tower, and the
INT4/FP8/NVFP4 tensor-core GEMM paths still CUDA-only (see
Metal below). The whole path from a model
checkpoint on disk to an OpenAI-compatible HTTP response is in this
repository.查看完整 README ↗目前无法加载完整 README。请通过原文链接阅读。 以下为已保存的摘录。
An inference engine written in Rust for GGUF, AWQ, native FP8 (W8A8), and
native NVFP4 (W4A4) checkpoints, with hand-written kernels for both NVIDIA
(CUDA) and Apple Silicon (Metal) GPUs — no PyTorch, no libtorch, no ggml.
CUDA is the primary, most complete backend; Metal covers dense decode, GQA
attention and GatedDeltaNet today, with MoE, the vision tower, and the
INT4/FP8/NVFP4 tensor-core GEMM paths still CUDA-only (see
Metal below). The whole path from a model
checkpoint on disk to an OpenAI-compatible HTTP response is in this
repository.查看完整 README ↗