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: bring the Llasa top_p default inside the range the API accepts (#9194)

Daniel Hanef8f965f97d0
diff --git a/studio/backend/assets/configs/model_defaults/llasa/unsloth_Llasa-3B.yaml b/studio/backend/assets/configs/model_defaults/llasa/unsloth_Llasa-3B.yaml
index 30e4440afb5..c5c1e6b60f2 100644
--- a/studio/backend/assets/configs/model_defaults/llasa/unsloth_Llasa-3B.yaml
+++ b/studio/backend/assets/configs/model_defaults/llasa/unsloth_Llasa-3B.yaml
@@ -41,5 +41,5 @@ logging:
inference:
temperature: 1.2
- top_p: 1.2
-
+ # 1.0 disables nucleus sampling and satisfies the API's 0-to-1 constraint.
+ top_p: 1.0
diff --git a/studio/backend/tests/test_sampling_resolution.py b/studio/backend/tests/test_sampling_resolution.py
index 7c6c1dd0044..f685e68e17c 100644
--- a/studio/backend/tests/test_sampling_resolution.py
+++ b/studio/backend/tests/test_sampling_resolution.py
@@ -7,7 +7,10 @@ Precedence per field: operator UNSLOTH_SAMPLING_* pin -> client explicit value -
per-model recommendation (load_inference_config) -> static schema default.
"""
+from pathlib import Path
+
import pytest
+import yaml
from utils.inference.inference_config import resolve_effective_sampling, SAMPLING_FIELD_NAMES
from utils.inference import inference_config as ic
@@ -110,6 +113,20 @@ def test_recommendation_matches_ui_source(model):
assert rec.get(f) == cleaned, f"{model}:{f} rec={rec.get(f)} ui={ui.get(f)}"
+def test_model_recommended_sampling_values_are_in_range():
+ defaults_dir = Path(ic.__file__).resolve().parents[2] / "assets" / "configs" / "model_defaults"
+ invalid = []
+ for path in sorted(defaults_dir.rglob("*.yaml")):
+ inference = (yaml.safe_load(path.read_text(encoding = "utf-8")) or {}).get(
+ "inference", {}
+ ) or {}
+ for field in ic._UI_RECOMMENDED_FIELDS:
+ if field in inference and ic._clean_sampling_value(field, inference[field]) is None:
+ invalid.append(f"{path.relative_to(defaults_dir)}:{field}={inference[field]!r}")
+
+ assert not invalid, "Out-of-range model sampling defaults: " + ", ".join(invalid)
+
+
def test_qwen38_reuses_qwen36_sampling_defaults():
qwen36 = ic.load_inference_config("unsloth/Qwen3.6-27B-GGUF")
qwen38 = ic.load_inference_config("unsloth/Qwen3.8-27B-GGUF")