Prevent llama output pipe deadlock

app.py

@@ -265,12 +265,14 @@ def _run_bench_locked(model_filter=None):
             to = 90 if mi["size_gb"] <= 6 else 180
             command = [bench, "-m", mp, "-t", str(nt), "-ngl", "0", "-c", "256", "-p", "32", "-n", "64", "-r", "1", "--no-mmap", "-o", "json"]
             _log(f"Running llama-bench for {mi['name']} on CPU with {nt} threads")
-            p = subprocess.run(command, capture_output=True, text=True, timeout=to, env=env)
+            with tempfile.TemporaryFile(mode="w+", encoding="utf-8") as output_file:
+                p = subprocess.run(command, stdout=output_file, stderr=subprocess.PIPE, text=True, timeout=to, env=env)
+                output_file.seek(0)
+                raw = output_file.read().strip()
             wt = time.time()-t0
             diagnostics = (p.stderr or "").strip()
             if p.returncode != 0:
                 raise RuntimeError(f"llama-bench exited with code {p.returncode}: {diagnostics[-1500:]}")
-            raw = (p.stdout or "").strip()
             parsed = json.loads(raw)
             row = parsed[0] if isinstance(parsed, list) and parsed else parsed
             r.update({"status":"success", "total_time_sec":round(wt,2), "prompt_tokens":32, "completion_tokens":64, "prompt_tokens_per_sec":row.get("avg_ts", 0), "generation_tokens_per_sec":row.get("avg_ts", 0), "raw":row})
@@ -338,10 +340,10 @@ def _run_cpu_smoke(model_filter):
     except Exception as exc:
         result["phases"]["help"] = {"ok": False, "elapsed_sec": round(time.time() - help_t0, 2), "error": str(exc)}
 
-    command = [cli, "-m", path, "-t", "1", "-ngl", "0", "-c", "256", "-n", "1", "-p", "Say hi.", "--no-display-prompt", "--no-warmup", "-fa", "0"]
+    command = [cli, "-m", path, "-t", "1", "-ngl", "0", "-c", "256", "-n", "1", "-p", "Say hi.", "--no-display-prompt", "--no-warmup", "--log-disable", "-fa", "0"]
     t0 = time.time()
     try:
-        run = subprocess.run(command, capture_output=True, text=True, timeout=180, env=env)
+        run = subprocess.run(command, stdout=subprocess.DEVNULL, stderr=subprocess.PIPE, text=True, timeout=180, env=env)
         result["phases"]["load_and_one_token"] = {
             "ok": run.returncode == 0,
             "elapsed_sec": round(time.time() - t0, 2),
@@ -373,7 +375,7 @@ def _run_cpu_process_probe(model_filter):
     env["LD_LIBRARY_PATH"] = os.path.dirname(cli)
     env["CUDA_VISIBLE_DEVICES"] = ""
     env["GGML_CUDA"] = "0"
-    command = [cli, "-m", path, "-t", "1", "-ngl", "0", "-c", "256", "-n", "1", "-p", "Say hi.", "--no-display-prompt", "--no-warmup", "-fa", "0"]
+    command = [cli, "-m", path, "-t", "1", "-ngl", "0", "-c", "256", "-n", "1", "-p", "Say hi.", "--no-display-prompt", "--no-warmup", "--log-disable", "-fa", "0"]
     started = time.time()
     child = subprocess.Popen(command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, env=env)
     time.sleep(8)
@@ -389,8 +391,8 @@ def _run_cpu_process_probe(model_filter):
     finally:
         child.kill()
         stdout, stderr = child.communicate(timeout=10)
-        probe["stdout"] = stdout[-2000:]
-        probe["stderr"] = stderr[-4000:]
+        probe["stdout"] = (stdout or "")[-2000:]
+        probe["stderr"] = (stderr or "")[-4000:]
         probe["returncode"] = child.returncode
     return probe