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: do not download PyTorch Clef for an over-long decision request (#16509)

Nilaya970757b5cbc
diff --git a/studio/backend/core/systemone/laya_runtime.py b/studio/backend/core/systemone/laya_runtime.py
index 6d4af4ac051..002f8f11b3f 100644
--- a/studio/backend/core/systemone/laya_runtime.py
+++ b/studio/backend/core/systemone/laya_runtime.py
@@ -2282,7 +2282,7 @@ def _route(
try:
return _decide(target, state, questions, images, audio)
except NativeContextOverflow as exc:
- # Auto answers on PyTorch only what its longer window can hold; at equal windows it would refuse too.
+ # Auto answers on PyTorch only what its longer window can hold, and only from a download already made.
from .catalog import clef_unsupported_reason
if (
@@ -2291,6 +2291,7 @@ def _route(
or images
or exc.limit >= MAX_LENGTH
or clef_unsupported_reason() is not None
+ or not is_cached(checkpoint)
):
raise Unavailable(422, "invalid_request_error", str(exc)) from None
_fallback_reason = "The request is longer than the llama.cpp context."
diff --git a/studio/backend/tests/test_decision_native.py b/studio/backend/tests/test_decision_native.py
index dd459326e2f..e5ab610f285 100644
--- a/studio/backend/tests/test_decision_native.py
+++ b/studio/backend/tests/test_decision_native.py
@@ -796,6 +796,28 @@ def test_auto_retries_an_over_long_state_on_pytorch_only_when_it_is_longer(
assert home.torch_agents == []
+def test_auto_refuses_an_over_long_state_when_pytorch_is_not_downloaded(
+ home, client, stub, monkeypatch
+):
+ downloads = []
+
+ def not_downloaded(checkpoint, local_only):
+ if not local_only:
+ downloads.append(checkpoint.name)
+ raise FileNotFoundError(checkpoint.name)
+
+ monkeypatch.setattr(laya_runtime, "_clef_dir", not_downloaded)
+ _cache_gguf(home.cache)
+ _put(client, enabled = True, model = "clef-flash", native_ctx = 1024)
+ response = _post(client, state = "word " * 1100)
+ if laya_runtime._loader is not None:
+ laya_runtime._loader.join(5)
+ assert response.status_code == 422 and "at most 1024" in response.json()["detail"]["message"]
+ assert downloads == [] and home.torch_agents == []
+ assert _post(client).headers["x-unsloth-decision-backend"] == "llama.cpp"
+ assert len(stub.records("start")) == 1
+
+
def test_other_native_failures_are_surfaced_not_retried(home, client, stub):
_cache_gguf(home.cache)
_put(client, enabled = True, model = "clef-flash")