Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Make page-in startup progress tests independent of host scheduling (#16476)

Daniel Hanb773d91a0017
diff --git a/studio/backend/tests/test_llama_cpp_wait_for_health.py b/studio/backend/tests/test_llama_cpp_wait_for_health.py
index 5f6a9f4b795..aa73fe95172 100644
--- a/studio/backend/tests/test_llama_cpp_wait_for_health.py
+++ b/studio/backend/tests/test_llama_cpp_wait_for_health.py
@@ -5,7 +5,6 @@
from __future__ import annotations
-import os
import subprocess
import sys
import threading
@@ -1517,22 +1516,6 @@ class TestHealthWaitMeasuresStalls:
"time.sleep(60)\n"
)
- _MMAP_WORKER = (
- "import mmap, os, sys, time\n"
- "fd = os.open(sys.argv[1], os.O_RDONLY | getattr(os, 'O_BINARY', 0))\n"
- "view = 16 << 20\n"
- "end = time.monotonic() + float(sys.argv[2])\n"
- "offset = 0\n"
- "while time.monotonic() < end:\n"
- " if hasattr(os, 'posix_fadvise'):\n"
- " os.posix_fadvise(fd, 0, 0, os.POSIX_FADV_DONTNEED)\n"
- " with mmap.mmap(fd, view, access = mmap.ACCESS_READ, offset = offset) as m:\n"
- " sum(m[p] for p in range(0, view, mmap.PAGESIZE))\n"
- " offset = view - offset\n"
- " time.sleep(0.05)\n"
- "time.sleep(60)\n"
- )
-
def _wait_on_child(
self,
monkeypatch,
@@ -1605,21 +1588,74 @@ class TestHealthWaitMeasuresStalls:
assert ok is True
assert elapsed >= 2.5
- def test_mmap_page_ins_with_flat_resident_memory_count(self, monkeypatch, tmp_path):
- monkeypatch.setattr(LlamaCppBackend, "_STARTUP_PROGRESS_MIN_CPU_FRACTION", 100.0)
- monkeypatch.setattr(LlamaCppBackend, "_STARTUP_PROGRESS_MIN_BYTES", 1 << 40)
- model = tmp_path / "model.gguf"
- # Small writes avoid collapsing the file into 2 MiB cache folios.
- with open(model, "wb") as f:
- for _ in range(512):
- f.write(os.urandom(64 << 10))
- # Longer stall window: page faults use a flat threshold, so CPU starvation hits this test.
- # Without fault progress the wait still dies at 1.5s, before the 2.5s health flip.
- b, ok, elapsed = self._wait_on_child(
- monkeypatch, [self._MMAP_WORKER, str(model), "3.0"], healthy_after = 2.5, timeout = 1.5
+ @pytest.mark.parametrize("faults_progress", [True, False])
+ def test_mmap_page_ins_with_flat_resident_memory_count(self, monkeypatch, faults_progress):
+ import contextlib
+ import psutil
+
+ module = sys.modules[LlamaCppBackend.__module__]
+ clock = _types.SimpleNamespace(now = 0.0)
+ fake_time = _types.SimpleNamespace(monotonic = lambda: clock.now)
+ monkeypatch.setattr(module, "time", fake_time)
+ monkeypatch.setattr(module, "_local_ssl_context", lambda: None)
+ monkeypatch.setattr(LlamaCppBackend, "_STARTUP_PROGRESS_SAMPLE_S", 0.1)
+ monkeypatch.setattr(
+ LlamaCppBackend, "_health_listener_is_ours", staticmethod(lambda p, port: True)
)
- assert ok is True
- assert elapsed >= 2.5
+
+ class Event:
+ def clear(self):
+ pass
+
+ def wait(self, interval):
+ clock.now = round(clock.now + interval, 6)
+
+ fake_pid = (1 << 63) - 1
+ member = mock.Mock(pid = fake_pid)
+ member.children.return_value = []
+ member.oneshot.side_effect = contextlib.nullcontext
+ member.cpu_times.return_value = _types.SimpleNamespace(user = 1.0, system = 0.0)
+ member.io_counters.return_value = _types.SimpleNamespace(read_bytes = 100, read_chars = 100)
+ member.memory_info.side_effect = lambda: _types.SimpleNamespace(
+ rss = 1000, pfaults = 1000 + (round(clock.now * 10) * 128 if faults_progress else 0)
+ )
+ monkeypatch.setattr(psutil, "Process", lambda pid: member)
+ monkeypatch.setattr(
+ httpx, "get", lambda *a, **kw: mock.Mock(status_code = 200 if clock.now >= 2.5 else 503)
+ )
+ b = _make_backend()
+ b._process.pid = fake_pid
+ b._process.poll.return_value = None
+ b._health_probe_event = Event()
+ ok = b._wait_for_health(timeout = 1.5, interval = 0.1)
+ assert ok is faults_progress
+ if faults_progress:
+ assert clock.now >= 2.5
+ assert not b._health_wait_stalled
+ else:
+ assert clock.now == pytest.approx(1.5)
+ assert b._health_wait_stalled
+ assert any("no startup progress for 1.5s" in ln for ln in b._stdout_lines)
+
+ def test_page_fault_counters_from_platform_and_linux_stat(self, monkeypatch):
+ import io
+
+ member = _types.SimpleNamespace(pid = (1 << 63) - 1)
+ assert LlamaCppBackend._page_faults(member, _types.SimpleNamespace(pfaults = 71)) == 71
+ assert (
+ LlamaCppBackend._page_faults(member, _types.SimpleNamespace(num_page_faults = 93)) == 93
+ )
+ opened = []
+
+ def fake_open(path, mode):
+ opened.append((path, mode))
+ return io.BytesIO(
+ f"{member.pid} (name with ) in it) R 1 2 3 4 5 6 123 8 45 10".encode()
+ )
+
+ monkeypatch.setattr("builtins.open", fake_open)
+ assert LlamaCppBackend._page_faults(member, _types.SimpleNamespace()) == 168
+ assert opened == [(f"/proc/{member.pid}/stat", "rb")]
def test_unreadable_counters_keep_the_fixed_deadline(self, monkeypatch):
monkeypatch.setattr(LlamaCppBackend, "_startup_work_sample", staticmethod(lambda p: None))
diff --git a/tests/studio/install/test_rocm_support.py b/tests/studio/install/test_rocm_support.py
index 7ad50d767a6..bf6981aea9d 100644
--- a/tests/studio/install/test_rocm_support.py
+++ b/tests/studio/install/test_rocm_support.py
@@ -5711,7 +5711,7 @@ class TestProgressStepCountMatchesTotal:
_no_torch_pair = MagicMock(return_value = ("2026.11.1", "2026.11.1", 0)),
),
patch.object(stack_mod, "_rocm_windows_torch_installed", False),
- patch.object(stack_mod, "_bootstrap_uv", return_value = False),
+ patch.object(stack_mod, "_bootstrap_uv", return_value = True),
patch.object(stack_mod, "_installed_torch_is_windows_rocm", return_value = False),
patch.object(stack_mod, "_has_usable_nvidia_gpu", return_value = True),
patch.object(stack_mod, "_repair_bad_anyio"),