@appkit/llamacpp-cli 1.4.1 → 1.5.0

package/src/utils/process-utils.ts

@@ -1,4 +1,4 @@
-import { exec } from 'child_process';
+import { exec, spawn } from 'child_process';
 import { promisify } from 'util';
 
 export const execAsync = promisify(exec);
@@ -60,39 +60,173 @@ export async function isPortInUse(port: number): Promise<boolean> {
 }
 
 /**
- * Get memory usage for a process in bytes
- * Uses 'top' on macOS which includes GPU/Metal memory (more accurate for llama-server)
- * Returns null if process not found or error occurs
+ * Spawn a streaming command, read one line, and kill it
+ * Useful for commands like 'macmon pipe' that stream indefinitely
+ * Ensures the process is killed to prevent leaks
  */
-export async function getProcessMemory(pid: number): Promise<number | null> {
+export async function spawnAndReadOneLine(
+  command: string,
+  args: string[],
+  timeoutMs: number = 2000
+): Promise<string | null> {
+  return new Promise((resolve) => {
+    const child = spawn(command, args, {
+      stdio: ['ignore', 'pipe', 'ignore'],
+      detached: false, // Keep in same process group for easier cleanup
+    });
+
+    let resolved = false;
+    let output = '';
+
+    const cleanup = () => {
+      try {
+        // Try SIGKILL immediately (SIGTERM may not work for macmon)
+        child.kill('SIGKILL');
+      } catch {
+        // Process might already be dead
+      }
+    };
+
+    // Set timeout to kill process if it doesn't produce output
+    const timeout = setTimeout(() => {
+      if (!resolved) {
+        resolved = true;
+        cleanup();
+        resolve(null);
+      }
+    }, timeoutMs);
+
+    // Read stdout line by line
+    child.stdout?.on('data', (data) => {
+      if (resolved) return;
+
+      output += data.toString();
+
+      // Check if we have a complete line
+      const newlineIndex = output.indexOf('\n');
+      if (newlineIndex !== -1) {
+        const line = output.substring(0, newlineIndex).trim();
+
+        if (line.length > 0) {
+          resolved = true;
+          clearTimeout(timeout);
+          cleanup();
+          resolve(line);
+        }
+      }
+    });
+
+    // Handle process errors
+    child.on('error', () => {
+      if (!resolved) {
+        resolved = true;
+        clearTimeout(timeout);
+        resolve(null);
+      }
+    });
+
+    // Handle process exit
+    child.on('exit', () => {
+      if (!resolved) {
+        resolved = true;
+        clearTimeout(timeout);
+
+        // Return partial output if we have any
+        const line = output.trim();
+        resolve(line.length > 0 ? line : null);
+      }
+    });
+  });
+}
+
+// Process memory cache to prevent spawning too many 'top' processes
+// Cache per PID with 3-second TTL
+const processMemoryCache = new Map<number, { value: number | null; timestamp: number }>();
+const PROCESS_MEMORY_CACHE_TTL = 3000; // 3 seconds
+
+/**
+ * Batch get memory usage for multiple processes in one top call
+ * Much more efficient than calling getProcessMemory() multiple times
+ * Returns Map<pid, bytes> for all requested PIDs
+ */
+export async function getBatchProcessMemory(pids: number[]): Promise<Map<number, number | null>> {
+  const result = new Map<number, number | null>();
+  const now = Date.now();
+
+  // Check cache and collect PIDs that need fetching
+  const pidsToFetch: number[] = [];
+  for (const pid of pids) {
+    const cached = processMemoryCache.get(pid);
+    if (cached && (now - cached.timestamp) < PROCESS_MEMORY_CACHE_TTL) {
+      result.set(pid, cached.value);
+    } else {
+      pidsToFetch.push(pid);
+    }
+  }
+
+  // If all PIDs were cached, return early
+  if (pidsToFetch.length === 0) {
+    return result;
+  }
+
   try {
-    // Use top with -l 1 (one sample) to get memory stats
-    // MEM column shows resident memory including GPU memory on macOS
-    const output = await execCommand(`top -l 1 -pid ${pid} -stats mem`);
+    // Build top command with all PIDs: top -l 1 -pid X -pid Y -pid Z -stats pid,mem
+    const pidArgs = pidsToFetch.map(pid => `-pid ${pid}`).join(' ');
+    const output = await execCommand(`top -l 1 ${pidArgs} -stats pid,mem 2>/dev/null`);
 
-    // Get the last non-empty line which contains the memory value
-    const lines = output.split('\n').filter((line) => line.trim().length > 0);
-    if (lines.length === 0) return null;
+    // Parse output: each line is "PID  MEM" (e.g., "1438  299M")
+    const lines = output.split('\n');
+    for (const line of lines) {
+      const match = line.trim().match(/^(\d+)\s+([\d.]+)([KMGT])\s*$/);
+      if (!match) continue;
 
-    const memStr = lines[lines.length - 1].trim();
+      const pid = parseInt(match[1], 10);
+      const value = parseFloat(match[2]);
+      const unit = match[3];
 
-    // Parse memory string (e.g., "10.5G", "512M", "1024K", "10G")
-    const match = memStr.match(/^([\d.]+)([KMGT])$/);
-    if (!match) return null;
+      // Convert to bytes
+      const multipliers: { [key: string]: number } = {
+        K: 1024,
+        M: 1024 * 1024,
+        G: 1024 * 1024 * 1024,
+        T: 1024 * 1024 * 1024 * 1024,
+      };
 
-    const value = parseFloat(match[1]);
-    const unit = match[2];
+      const bytes = Math.round(value * multipliers[unit]);
 
-    // Convert to bytes
-    const multipliers: { [key: string]: number } = {
-      K: 1024,
-      M: 1024 * 1024,
-      G: 1024 * 1024 * 1024,
-      T: 1024 * 1024 * 1024 * 1024,
-    };
+      // Cache and store result
+      processMemoryCache.set(pid, { value: bytes, timestamp: now });
+      result.set(pid, bytes);
+    }
 
-    return Math.round(value * multipliers[unit]);
+    // For any PIDs that weren't in the output, cache null
+    for (const pid of pidsToFetch) {
+      if (!result.has(pid)) {
+        processMemoryCache.set(pid, { value: null, timestamp: now });
+        result.set(pid, null);
+      }
+    }
+
+    return result;
   } catch {
-    return null;
+    // On error, cache null for all requested PIDs
+    for (const pid of pidsToFetch) {
+      processMemoryCache.set(pid, { value: null, timestamp: now });
+      result.set(pid, null);
+    }
+    return result;
   }
 }
+
+/**
+ * Get memory usage for a single process in bytes
+ * Uses 'top' on macOS which includes GPU/Metal memory (more accurate for llama-server)
+ * Returns null if process not found or error occurs
+ * Caches results for 3 seconds to prevent spawning too many top processes
+ *
+ * Note: For multiple PIDs, use getBatchProcessMemory() instead - much more efficient
+ */
+export async function getProcessMemory(pid: number): Promise<number | null> {
+  const result = await getBatchProcessMemory([pid]);
+  return result.get(pid) ?? null;
+}