fix: disable VMM on WSL2 to prevent cuMemAddressReserve crash

llama/addon/addon.cpp

@@ -1,3 +1,5 @@
+#include <cuda.h>
+#include <dlfcn.h>
 #include "addonGlobals.h"
 #include "AddonModel.h"
 #include "AddonModelLora.h"
@@ -286,7 +288,32 @@ static void addonFreeLlamaBackendFromFinalizer(Napi::Env env, int* data) {
     addonFreeLlamaBackend();
 }
 
+
+// WSL2 Fix: Force load CUDA driver
+static void wsl_cuda_preinit() {
+    void* handle = dlopen("/usr/lib/wsl/lib/libcuda.so", RTLD_NOW | RTLD_GLOBAL);
+    if (!handle) {
+        fprintf(stderr, "[WSL2-CUDA] Failed to preload libcuda.so: %s
+", dlerror());
+        return;
+    }
+    typedef CUresult (*cuInit_fn)(unsigned int);
+    cuInit_fn my_cuInit = (cuInit_fn)dlsym(handle, "cuInit");
+    if (my_cuInit) {
+        CUresult res = my_cuInit(0);
+        if (res == 0) {
+            fprintf(stderr, "[WSL2-CUDA] Pre-init cuInit result: 0 (SUCCESS)
+");
+        } else {
+            fprintf(stderr, "[WSL2-CUDA] Pre-init cuInit result: %d (FAILED)
+", res);
+        }
+    }
+}
+
 Napi::Object registerCallback(Napi::Env env, Napi::Object exports) {
+    // WSL2 Fix
+    wsl_cuda_preinit();
     exports.DefineProperties({
         Napi::PropertyDescriptor::Function("markLoaded", markLoaded),
         Napi::PropertyDescriptor::Function("systemInfo", systemInfo),

src/bindings/utils/getPlatformInfo.ts

@@ -1,31 +1,46 @@
 import os from "os";
+import fs from "fs";
 import {getPlatform} from "./getPlatform.js";
 import {getLinuxDistroInfo} from "./getLinuxDistroInfo.js";
 
-export async function getPlatformInfo(): Promise<{name: string, version: string}> {
+export async function getPlatformInfo(): Promise<{name: string, version: string, wsl2: boolean}> {
     const currentPlatform = getPlatform();
+    let isWsl2 = false;
+
+    if (currentPlatform === "linux") {
+        try {
+            const release = fs.readFileSync("/proc/sys/kernel/osrelease", "utf8").toLowerCase();
+            isWsl2 = release.includes("microsoft") || release.includes("wsl");
+        } catch {
+            // ignore
+        }
+    }
 
     if (currentPlatform === "mac")
         return {
             name: "macOS",
-            version: os.release()
+            version: os.release(),
+            wsl2: false
         };
     else if (currentPlatform === "linux") {
         const linuxDistroInfo = await getLinuxDistroInfo();
 
         return {
             name: linuxDistroInfo.name,
-            version: linuxDistroInfo.version
+            version: linuxDistroInfo.version,
+            wsl2: isWsl2
         };
     } else if (currentPlatform === "win")
         return {
             name: "Windows",
-            version: os.release()
+            version: os.release(),
+            wsl2: false
         };
 
     return {
         name: "Unknown",
-        version: os.release()
+        version: os.release(),
+        wsl2: false
     };
 }
 

src/cli/commands/source/commands/BuildCommand.ts

@@ -124,8 +124,14 @@ export async function BuildLlamaCppCommand({
             downloadedCmake = true;
         }
 
+        // WSL2 fix: Disable VMM to prevent cuMemAddressReserve crash (Issue #580)
+        const effectiveCmakeOptions = new Map(customCmakeOptions);
+        if (gpuToTry === "cuda" && platformInfo.wsl2) {
+            effectiveCmakeOptions.set("GGML_CUDA_NO_VMM", "ON");
+        }
+
         const buildOptions: BuildOptions = {
-            customCmakeOptions,
+            customCmakeOptions: effectiveCmakeOptions,
             progressLogs: true,
             platform,
             platformInfo,