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.

Studio smoke: report a transport failure instead of crashing on one (#10420)

Daniel Hanfb9bb4ea003e
diff --git a/tests/studio/playwright_loaded_models_indicator.py b/tests/studio/playwright_loaded_models_indicator.py
index 0642ffd604c..6c37ffd94af 100644
--- a/tests/studio/playwright_loaded_models_indicator.py
+++ b/tests/studio/playwright_loaded_models_indicator.py
@@ -366,7 +366,18 @@ def run(page, state: Runtime) -> None:
for route in ("/hub", "/train", "/images"):
page.goto(BASE + route, wait_until = "domcontentloaded")
- page.wait_for_timeout(3000)
+ # Wait for the card, not for the clock. This is a hard navigation: a
+ # full SPA reload plus a loaded-models poll, and 3000ms was the only
+ # fixed budget in this file that was not derived from SETTLE_MS. On a
+ # loaded runner the first route overran it and all three then failed
+ # together, which is what a fixed budget looks like when it is the
+ # thing that is wrong. The check below is unchanged and still fails if
+ # the card genuinely does not survive the navigation -- this only stops
+ # a slow render from being read as a missing card.
+ try:
+ page.wait_for_selector(CARD, timeout = SETTLE_MS)
+ except Exception:
+ pass
check(f"card survives {route}", page.locator(CARD).count() > 0)
# ── Hardware shapes a CUDA runner never produces ────────────────────
diff --git a/tests/studio/studio_api_smoke.py b/tests/studio/studio_api_smoke.py
index ab76297b2bb..d37403662d6 100644
--- a/tests/studio/studio_api_smoke.py
+++ b/tests/studio/studio_api_smoke.py
@@ -100,6 +100,23 @@ def http(
return exc.code, json.loads(raw)
except (json.JSONDecodeError, UnicodeDecodeError):
return exc.code, raw
+ except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
+ # A transport failure is a failed check, not a crashed run. Only
+ # HTTPError was caught here, so a socket timeout propagated out of
+ # http() and killed the script at module level: the traceback named
+ # urlopen and the line number, and nothing about which endpoint had
+ # been slow or what the rest of the suite would have said. Observed on
+ # three unrelated PRs whose /v1/embeddings call hit the 30s ceiling.
+ # Returned as status 0 so every existing `if code == 200 ... else
+ # fail(...)` site reports it with its own diagnosis and the run
+ # continues to the remaining sections.
+ #
+ # Emitted here as well as returned: the call sites report the status
+ # code and not the body, so status 0 on its own would say a request
+ # failed without saying that it timed out or after how long.
+ detail = f"{type(exc).__name__}: {exc}"
+ _emit(" NET ", f"{method} {path} failed after {timeout}s -- {detail}")
+ return 0, {"error": detail, "timeout": timeout}
def login(password: str) -> tuple[int, str | None]:
@@ -449,7 +466,14 @@ code, body = http(
"/v1/embeddings",
body = {"model": GGUF_REPO, "input": "hello"},
headers = AUTH_HEADER,
- timeout = 30,
+ # This is the first request in the section that forces a model load, so it
+ # pays the load on top of the embedding. Measured at 23.8s on a healthy
+ # macOS runner against a 30s ceiling, i.e. 79% of budget, which is not a
+ # margin: three unrelated PRs timed out here at exactly 30.0s on the same
+ # base while main passed. Raised to leave room for a loaded runner rather
+ # than to hide a slow endpoint -- if this starts taking 90s that is a real
+ # regression and it will still be reported, now as a failed check.
+ timeout = 90,
)
if code == 200 and isinstance(body, dict) and body.get("data"):
ok("/v1/embeddings -> 200 with data")