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: fix structured columns in Data Recipes on non-GGUF models (#11049)

Nilay9a127b6c118b
diff --git a/studio/backend/routes/data_recipe/jobs.py b/studio/backend/routes/data_recipe/jobs.py
index c778f7feea7..cef96a3d6ba 100644
--- a/studio/backend/routes/data_recipe/jobs.py
+++ b/studio/backend/routes/data_recipe/jobs.py
@@ -216,6 +216,11 @@ def _single_used_local_model_selection(
return next(iter(selections))
+def _local_chat_serves_gguf() -> bool:
+ from routes.inference import get_llama_cpp_backend
+ return bool(get_llama_cpp_backend().is_loaded)
+
+
def _loaded_local_model_identity() -> tuple[bool, str, str]:
from routes.inference import get_llama_cpp_backend
from core.inference import get_inference_backend
@@ -429,9 +434,10 @@ def _inject_local_providers(
extra_body["chat_template_kwargs"] = tpl_kwargs
params["extra_body"] = extra_body
- # Forward each llm-structured column's output_format as a response_format so
- # llama-server uses grammar-constrained sampling instead of broken JSON.
- _inject_local_structured_response_format(recipe, local_names)
+ # Only llama.cpp carries a grammar engine; /v1 refuses response_format on every other local
+ # backend, so those fall back to the prompt-level JSON llm-judge columns already rely on.
+ if _local_chat_serves_gguf():
+ _inject_local_structured_response_format(recipe, local_names)
return internal_key_id
diff --git a/studio/backend/tests/test_data_recipe_structured_response_format.py b/studio/backend/tests/test_data_recipe_structured_response_format.py
new file mode 100644
index 00000000000..fa761fae68a
--- /dev/null
+++ b/studio/backend/tests/test_data_recipe_structured_response_format.py
@@ -0,0 +1,111 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import copy
+import sys
+from types import SimpleNamespace
+
+import pytest
+from fastapi import HTTPException
+
+from auth import storage
+from models.inference import ChatCompletionRequest, ChatMessage
+from routes.data_recipe import jobs as route
+
+from .test_sf_client_tools_passthrough import _call, _fixed, _ScriptedBackend
+
+MODEL = "unsloth/Qwen3-0.6B"
+SCHEMA = {
+ "type": "object",
+ "required": ["name", "score"],
+ "properties": {"name": {"type": "string"}, "score": {"type": "integer"}},
+}
+
+
+def structured_recipe():
+ return {
+ "model_providers": [{"name": "local", "is_local": True}],
+ "model_configs": [{"alias": "local_model", "model": MODEL, "provider": "local"}],
+ "columns": [
+ {"column_type": "llm-text", "name": "blurb", "model_alias": "local_model"},
+ {
+ "column_type": "llm-structured",
+ "name": "rating",
+ "model_alias": "local_model",
+ "output_format": copy.deepcopy(SCHEMA),
+ },
+ ],
+ }
+
+
+def inject_with_loaded_model(monkeypatch, *, gguf):
+ llama = SimpleNamespace(is_loaded = gguf, model_identifier = MODEL, hf_variant = "")
+ backend = SimpleNamespace(active_model_name = "" if gguf else MODEL)
+ monkeypatch.setitem(
+ sys.modules, "routes.inference", SimpleNamespace(get_llama_cpp_backend = lambda: llama)
+ )
+ monkeypatch.setitem(
+ sys.modules, "core.inference", SimpleNamespace(get_inference_backend = lambda: backend)
+ )
+ monkeypatch.setattr(
+ route, "_resolve_local_v1_endpoint", lambda request: "http://127.0.0.1:8888/v1"
+ )
+ monkeypatch.setattr(storage, "create_api_key", lambda **kwargs: ("sk-unsloth-test", {"id": 7}))
+ recipe = structured_recipe()
+ assert route._inject_local_providers(recipe, SimpleNamespace()) == 7
+ return recipe
+
+
+def response_formats(recipe):
+ return {
+ mc["alias"]: mc["inference_parameters"]["extra_body"]["response_format"]
+ for mc in recipe["model_configs"]
+ if "response_format" in mc["inference_parameters"]["extra_body"]
+ }
+
+
+def column_aliases(recipe):
+ return {column["name"]: column["model_alias"] for column in recipe["columns"]}
+
+
+def test_gguf_model_gets_grammar_response_format(monkeypatch):
+ recipe = inject_with_loaded_model(monkeypatch, gguf = True)
+
+ assert column_aliases(recipe) == {
+ "blurb": "local_model",
+ "rating": "local_model__rating_structured",
+ }
+ assert response_formats(recipe) == {
+ "local_model__rating_structured": {"type": "json_schema", "schema": SCHEMA}
+ }
+
+
+def test_non_gguf_model_keeps_prompt_level_json(monkeypatch):
+ recipe = inject_with_loaded_model(monkeypatch, gguf = False)
+
+ # Whole columns, not just aliases: output_format feeds the fallback's prompt-level schema.
+ assert recipe["columns"] == structured_recipe()["columns"]
+ assert response_formats(recipe) == {}
+ assert [mc["alias"] for mc in recipe["model_configs"]] == ["local_model"]
+ extra_body = recipe["model_configs"][0]["inference_parameters"]["extra_body"]
+ assert extra_body == {"chat_template_kwargs": {"enable_thinking": False}}
+
+
+def test_the_injected_response_format_is_what_a_non_gguf_backend_refuses(monkeypatch):
+ """The gate's premise: without llama.cpp, /v1 refuses the payload the GGUF path injects.
+ Teaching the safetensors/MLX backend guided decoding must break this, not go unnoticed."""
+ # Its own context so the fake backend modules unwind before the route is called for real.
+ with pytest.MonkeyPatch.context() as injection:
+ recipe = inject_with_loaded_model(injection, gguf = True)
+ extra_body = recipe["model_configs"][-1]["inference_parameters"]["extra_body"]
+ payload = ChatCompletionRequest(
+ model = "default",
+ messages = [ChatMessage(role = "user", content = "rate it")],
+ **extra_body,
+ )
+
+ with pytest.raises(HTTPException) as caught:
+ _call(payload, monkeypatch, _ScriptedBackend(_fixed('{"name": "Riverside", "score": 8}')))
+
+ assert caught.value.status_code == 400
+ assert caught.value.detail["error"]["param"] == "response_format"