dsh-codebase-chat 0.25.5 → 0.25.6

package/dist/index.js

@@ -2824,8 +2824,16 @@ function loadLocalModel() {
         throw new Error(`Local model not found: ${spec}`);
       }
       const { getLlama } = await importLlama();
-      const llama = await getLlama();
-      return llama.loadModel({ modelPath });
+      const cpuOnly = /^(0|off|false|cpu)$/i.test(process.env.CODEBASE_LOCAL_GPU || "");
+      const llama = await getLlama(cpuOnly ? { gpu: false } : {});
+      try {
+        return await llama.loadModel({ modelPath });
+      } catch (err) {
+        if (cpuOnly) throw err;
+        console.error(`[local-llm] GPU load failed (${err instanceof Error ? err.message : err}) \u2014 falling back to CPU. Slower, but works. Set CODEBASE_LOCAL_GPU=off to skip GPU entirely.`);
+        const cpuLlama = await getLlama({ gpu: false });
+        return cpuLlama.loadModel({ modelPath });
+      }
     })();
     modelPromise.catch(() => {
       modelPromise = null;