ci: make multi-GPU jit test hangs attributable from the CI log (#29925)

python/sglang/jit_kernel/mp.py

@@ -168,6 +168,10 @@ def multigpu_launch(
     os.environ[pid_key] = str(os.getpid())
     os.environ.setdefault("OMP_NUM_THREADS", "1")
     os.environ.setdefault("GLOO_SOCKET_IFNAME", "lo")  # single-machine setup
+    # Unbuffered child stdout: when a worker is killed on timeout, pytest's
+    # block-buffered progress output is otherwise lost or flushed out of
+    # order into the CI log, making it impossible to tell which test hung.
+    os.environ.setdefault("PYTHONUNBUFFERED", "1")
     signal.signal(signal.SIGINT, signal.default_int_handler)
     runnable: List[int] = []
     for N in sorted(num_gpus):

python/sglang/jit_kernel/tests/utils.py

@@ -43,7 +43,14 @@ def multigpu_pytest_main(
         # CI's run_unittest_files invokes `python3 <file> -f` (legacy
         # unittest failfast). Translate to pytest's `-x` so it survives.
         pytest_args = ["-x" if a == "-f" else a for a in sys.argv[1:]]
-        return pytest.main([file] + pytest_args)
+        # Dump all thread stacks (every rank; stderr is not redirected) if a
+        # single test exceeds half the harness budget, so a hung collective
+        # is attributable from the CI log before the outer timeout kills the
+        # process group. Non-fatal: the test keeps running after the dump.
+        dump_after = (timeout // 2) if timeout else 300
+        return pytest.main(
+            [file, "-o", f"faulthandler_timeout={dump_after}"] + pytest_args
+        )
 
     return multigpu_launch(
         name,