【训练营】MXFP8/NVFP4低精度模拟与反量化

02_quant_dequant/黄新颖/.gitignore

@@ -0,0 +1,32 @@
+# CMake 与本地构建目录
+/build/
+/cmake-build-*/
+/CMakeCache.txt
+/CMakeFiles/
+/CTestTestfile.cmake
+/Testing/
+/compile_commands.json
+/CMakeUserPresets.json
+
+# 程序运行产生的文件
+/outputs/
+/logs/
+*.log
+*.jsonl
+*.csv
+
+# Python 实验脚本的解释器缓存
+__pycache__/
+*.pyc
+
+# 大型或临时二进制数据；小型测试样例可显式保留在 tests/data/。
+*.bin
+!tests/data/
+!tests/data/**/*.bin
+
+# CUDA 分析和编译中间产物
+*.ptx
+*.cubin
+*.fatbin
+*.nsys-rep
+*.ncu-rep

02_quant_dequant/黄新颖/CMakeLists.txt

@@ -0,0 +1,50 @@
+cmake_minimum_required(VERSION 3.22)
+
+project(quant_dequant LANGUAGES CXX CUDA)
+
+set(CMAKE_CXX_STANDARD 20)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+set(CMAKE_CXX_EXTENSIONS OFF)
+
+# CMake 3.22 尚未定义 CUDA20 对应的 NVCC 编译选项。host 侧 .cpp 继续
+# 使用 C++20；.cu 文件使用 CUDA 12.1 与 CMake 3.22 均明确支持的 C++17。
+# 当前 codec 和 kernel 约束在 C++17 可用特性内，以便两种编译路径一致。
+set(CMAKE_CUDA_STANDARD 17)
+set(CMAKE_CUDA_STANDARD_REQUIRED ON)
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+
+option(QUANT_DEQUANT_BUILD_TESTS "构建正确性测试" ON)
+option(QUANT_DEQUANT_BUILD_APPS "构建命令行程序" ON)
+option(QUANT_DEQUANT_BUILD_BENCHMARKS "构建性能评测程序" ON)
+set(QUANT_DEQUANT_CUDA_ARCHITECTURES "" CACHE STRING
+    "目标 CUDA 架构，例如 RTX 4060 使用 89；留空时由 CMake 决定")
+
+if(QUANT_DEQUANT_CUDA_ARCHITECTURES)
+    set(CMAKE_CUDA_ARCHITECTURES "${QUANT_DEQUANT_CUDA_ARCHITECTURES}" CACHE STRING
+        "CUDA 架构列表" FORCE)
+endif()
+
+include(CTest)
+find_package(CUDAToolkit REQUIRED)
+
+function(quant_dequant_enable_warnings target_name)
+    target_compile_options(${target_name}
+        PRIVATE
+            $<$<COMPILE_LANGUAGE:CXX>:-Wall -Wextra -Wpedantic>
+    )
+endfunction()
+
+# src 只生成可复用核心库；应用、测试和性能评测各自链接该库。
+add_subdirectory(src)
+
+if(QUANT_DEQUANT_BUILD_APPS)
+    add_subdirectory(apps)
+endif()
+
+if(BUILD_TESTING AND QUANT_DEQUANT_BUILD_TESTS)
+    add_subdirectory(tests)
+endif()
+
+if(QUANT_DEQUANT_BUILD_BENCHMARKS)
+    add_subdirectory(benchmarks)
+endif()

02_quant_dequant/黄新颖/CMakePresets.json

@@ -0,0 +1,60 @@
+{
+  "version": 3,
+  "cmakeMinimumRequired": {
+    "major": 3,
+    "minor": 22,
+    "patch": 0
+  },
+  "configurePresets": [
+    {
+      "name": "base",
+      "hidden": true,
+      "generator": "Unix Makefiles",
+      "binaryDir": "${sourceDir}/build/${presetName}",
+      "cacheVariables": {
+        "QUANT_DEQUANT_BUILD_APPS": "ON",
+        "QUANT_DEQUANT_BUILD_BENCHMARKS": "ON",
+        "QUANT_DEQUANT_BUILD_TESTS": "ON",
+        "QUANT_DEQUANT_CUDA_ARCHITECTURES": "89"
+      }
+    },
+    {
+      "name": "rtx4060-release",
+      "displayName": "RTX 4060 Release",
+      "description": "为 NVIDIA RTX 4060（compute capability 8.9）构建 Release 版本。",
+      "inherits": "base",
+      "cacheVariables": {
+        "CMAKE_BUILD_TYPE": "Release"
+      }
+    },
+    {
+      "name": "rtx4060-debug",
+      "displayName": "RTX 4060 Debug",
+      "description": "为 NVIDIA RTX 4060（compute capability 8.9）构建 Debug 版本。",
+      "inherits": "base",
+      "cacheVariables": {
+        "CMAKE_BUILD_TYPE": "Debug"
+      }
+    }
+  ],
+  "buildPresets": [
+    {
+      "name": "rtx4060-release",
+      "configurePreset": "rtx4060-release"
+    },
+    {
+      "name": "rtx4060-debug",
+      "configurePreset": "rtx4060-debug"
+    }
+  ],
+  "testPresets": [
+    {
+      "name": "rtx4060-release",
+      "configurePreset": "rtx4060-release"
+    },
+    {
+      "name": "rtx4060-debug",
+      "configurePreset": "rtx4060-debug"
+    }
+  ]
+}