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: make Gemma 4 call a forced tool (#17432)

Nilaybb15e2cde930
diff --git a/studio/backend/core/inference/passthrough_healing.py b/studio/backend/core/inference/passthrough_healing.py
index 09cc789850a..eaf7ea4923e 100644
--- a/studio/backend/core/inference/passthrough_healing.py
+++ b/studio/backend/core/inference/passthrough_healing.py
@@ -328,6 +328,7 @@ _NATIVE_TOOL_FORMATS = (
_FORCED_CALL_OPENERS = {
"hermes": ('<tool_call>\n{"name":', lambda name: f' {json.dumps(name)}, "arguments":'),
"qwen3_coder": ("<tool_call>\n<function=", lambda name: f"{name}>\n"),
+ "gemma4": ("<|tool_call>call:", lambda name: f"{name}{{"),
}
@@ -337,7 +338,7 @@ def forced_call_prefix(names: Optional[set], *templates: Any) -> str:
from core.inference.engine_adapters import tool_parser_for_template
parsers = {
- tool_parser_for_template(body, "vllm")
+ "gemma4" if "<|tool_call>call:" in body else tool_parser_for_template(body, "vllm")
for template in templates
for body in _template_strings(template)
} - {None}
diff --git a/studio/backend/tests/test_sf_client_tools_passthrough.py b/studio/backend/tests/test_sf_client_tools_passthrough.py
index 12d66455568..7ec9b79b535 100644
--- a/studio/backend/tests/test_sf_client_tools_passthrough.py
+++ b/studio/backend/tests/test_sf_client_tools_passthrough.py
@@ -561,11 +561,22 @@ _QWEN_XML_TEMPLATE = (
"{% for t in tools %}{{ t.arguments }}{% endfor %}{% endif %}"
"{% for m in messages %}<|im_start|>{{ m.role }}\n{{ m.content }}<|im_end|>\n{% endfor %}"
)
+_GEMMA4_TEMPLATE = (
+ "{% if tools %}<|tool>{{ tools | tojson }}<tool|>{% endif %}"
+ "{% for m in messages %}<|turn>{{ m.role }}\n{{ m.content }}"
+ "{% for c in m.tool_calls %}{{ '<|tool_call>call:' + c.function.name + '{}<tool_call|>' }}"
+ "{% endfor %}<turn|>\n{% endfor %}"
+)
_ARGS = {"lookup": ("q", "cats"), "search": ("query", "dogs")}
def _model_that_answers_in_prose(messages, tools):
opener = messages[-1]["content"] if messages[-1]["role"] == "assistant" else ""
+ if opener.startswith("<|tool_call>call:"):
+ name = opener[len("<|tool_call>call:") :].rstrip("{") or "lookup"
+ key, value = _ARGS[name]
+ body = f'{key}:<|"|>{value}<|"|>}}<tool_call|>'
+ return [body if opener.endswith("{") else f"{name}{{{body}"]
if opener.startswith("<tool_call>\n<function="):
name = opener[len("<tool_call>\n<function=") :].rstrip(">\n") or "lookup"
key, value = _ARGS[name]
@@ -580,7 +591,9 @@ def _model_that_answers_in_prose(messages, tools):
@pytest.mark.parametrize(
- "template", [_HERMES_TEMPLATE, _QWEN_XML_TEMPLATE], ids = ["hermes", "qwen_xml"]
+ "template",
+ [_HERMES_TEMPLATE, _QWEN_XML_TEMPLATE, _GEMMA4_TEMPLATE],
+ ids = ["hermes", "qwen_xml", "gemma4"],
)
@pytest.mark.parametrize("stream", [False, True])
@pytest.mark.parametrize(
@@ -633,7 +646,9 @@ def test_auto_tool_choice_leaves_the_reply_to_the_model(monkeypatch):
@pytest.mark.parametrize("stream", [False, True])
@pytest.mark.parametrize(
- "template", [_HERMES_TEMPLATE, _QWEN_XML_TEMPLATE], ids = ["hermes", "qwen_xml"]
+ "template",
+ [_HERMES_TEMPLATE, _QWEN_XML_TEMPLATE, _GEMMA4_TEMPLATE],
+ ids = ["hermes", "qwen_xml", "gemma4"],
)
def test_an_empty_reply_to_a_forced_call_is_not_a_call(monkeypatch, template, stream):
backend = _ScriptedBackend(_fixed(""))
@@ -673,6 +688,8 @@ def test_forced_call_prefix_is_added_after_the_context_fit(monkeypatch):
({"lookup", "search"}, [_HERMES_TEMPLATE, None], '<tool_call>\n{"name":'),
({"lookup"}, [_QWEN_XML_TEMPLATE], "<tool_call>\n<function=lookup>\n"),
({"lookup", "search"}, [{"tool_use": _QWEN_XML_TEMPLATE}], "<tool_call>\n<function="),
+ ({"lookup"}, [_GEMMA4_TEMPLATE], "<|tool_call>call:lookup{"),
+ ({"lookup", "search"}, [_GEMMA4_TEMPLATE, None], "<|tool_call>call:"),
# Bodies disagreeing on the format, or a format with no opener, are not forced.
({"lookup"}, [_HERMES_TEMPLATE, _QWEN_XML_TEMPLATE], ""),
({"lookup"}, ["{% if tools %}[TOOL_CALLS]{% endif %}"], ""),