Skip to content

Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.

C++

serve: engine_ready_s in the run config (DeepSeek-V4.1-Flash reads 300 GB at start; setup writes 3600)

Niko1221a4c933641674
diff --git a/serve/server.py b/serve/server.py
index f41d96dfc..6dc058a33 100644
--- a/serve/server.py
+++ b/serve/server.py
@@ -2902,6 +2902,18 @@ def layer_split_value(cfg: dict) -> str:
return ",".join(str(x) for x in vals)
+def apply_ready_timeout(cfg: dict) -> None:
+ """The config's "engine_ready_s": how long a start may take before READY (default ENGINE_READY_S, 900). A model of 300+ GB read
+ from a slow disk (DeepSeek-V4.1-Flash: setup writes 3600) needs more. ValueError for anything but a number > 0."""
+ global ENGINE_READY_S
+ v = cfg.get("engine_ready_s")
+ if v is None:
+ return
+ if isinstance(v, bool) or not isinstance(v, (int, float)) or not v > 0:
+ raise ValueError(f'"engine_ready_s" must be a number of seconds above 0, not {v!r}')
+ ENGINE_READY_S = float(v)
+
+
def layer_split_of(cfg: dict) -> bool:
"""Several GPUs in the config are a layer split, unless the args put the later card(s) to another use: with
--peer-device the second card is an expert-cache tier, and the engine refuses that beside --layer-split (#665)."""
@@ -8004,6 +8016,7 @@ def main(argv=None, *, lifecycle_event=None, interrupt_exit_code: int = 0) -> in
exe = cfg["exe"] if os.path.isabs(cfg["exe"]) else os.path.abspath(os.path.join(cfg.get("cwd") or ".", cfg["exe"]))
try:
silence = engine_silence_s(cfg) # #481: checked before the (minutes-long) start
+ apply_ready_timeout(cfg)
except ValueError as e:
raise SystemExit(f"[strata] config {e}")
try:
diff --git a/serve/test_dsv41.py b/serve/test_dsv41.py
index 71c6a12fe..ae9040473 100644
--- a/serve/test_dsv41.py
+++ b/serve/test_dsv41.py
@@ -280,6 +280,22 @@ class Parser(unittest.TestCase):
[("get_weather", {"city": "Rome"})])
+class ReadyTimeout(unittest.TestCase):
+ """A 300+ GB model read from a slow disk needs more than the default 900 s before READY: the config's "engine_ready_s"."""
+
+ def test_the_config_sets_it_and_a_bad_value_is_said(self):
+ from serve import server
+ before = server.ENGINE_READY_S
+ self.addCleanup(setattr, server, "ENGINE_READY_S", before)
+ server.apply_ready_timeout({})
+ self.assertEqual(server.ENGINE_READY_S, before)
+ server.apply_ready_timeout({"engine_ready_s": 3600})
+ self.assertEqual(server.ENGINE_READY_S, 3600.0)
+ for bad in (0, -5, "soon", True):
+ with self.assertRaises(ValueError):
+ server.apply_ready_timeout({"engine_ready_s": bad})
+
+
class Api(unittest.TestCase):
"""Both APIs against a fake engine that replays scripted DeepSeek text."""
diff --git a/setup.py b/setup.py
index 6f7eb0c0a..21fe40512 100644
--- a/setup.py
+++ b/setup.py
@@ -7350,7 +7350,7 @@ def install_dsv41(a, family, gpu, hip, amd, cuda_tk, multi, port, ram, adopted,
args = ["--pack", str(pack), "--max-context", str(ctx), "--eos", "1", *plan["args"]]
cfg = {"exe": str(exe), "args": args, "cwd": str(ROOT), "tokenizer": str(pack / "tokenizer"),
"model_name": fam["name"], "arch": fam["arch"], "log": str(home() / f"strata-{tag.lower()}.log"),
- "lib_dirs": lib_dirs, "port": port}
+ "lib_dirs": lib_dirs, "port": port, "engine_ready_s": 3600} # 300+ GB read from the files: a slow disk needs its time
if cuda_tk == 12:
cfg["cuda"] = 12
if hip:
diff --git a/tools/test_setup_dsv41.py b/tools/test_setup_dsv41.py
index 7b403c89b..fd9e86865 100644
--- a/tools/test_setup_dsv41.py
+++ b/tools/test_setup_dsv41.py
@@ -200,6 +200,7 @@ class Install(unittest.TestCase):
self.assertEqual(a[a.index("--tp") + 1], "2")
self.assertTrue(Path(a[a.index("--pack") + 1]).match("*packs/deepseek-v4.1-flash-ud-q8_k_xl"))
self.assertEqual(self.cfg["gpu"], [0, 1, 2, 3])
+ self.assertEqual(self.cfg["engine_ready_s"], 3600)
self.assertEqual(self.cfg["backend"], "hip")
self.assertTrue(self.cfg["tokenizer"].endswith("tokenizer"))
self.assertNotIn("--vision", a)