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: keep the base URL query string on connection requests (#16951)

Nilay7256cb4f3bc2
diff --git a/studio/backend/core/inference/external_provider.py b/studio/backend/core/inference/external_provider.py
index 256dfeffc2f..a0969c9575c 100644
--- a/studio/backend/core/inference/external_provider.py
+++ b/studio/backend/core/inference/external_provider.py
@@ -190,8 +190,10 @@ def _merge_concatenated_wav_segments(
def _append_provider_path(base_url: str, endpoint: str) -> str:
"""Append an API path without moving it behind a base URL's query string."""
parts = urlsplit(base_url)
- path = f"{parts.path.rstrip('/')}/{endpoint.lstrip('/')}"
- return urlunsplit((parts.scheme, parts.netloc, path, parts.query, parts.fragment))
+ endpoint_path, _, endpoint_query = endpoint.partition("?")
+ path = f"{parts.path.rstrip('/')}/{endpoint_path.lstrip('/')}"
+ query = "&".join(q for q in (parts.query, endpoint_query) if q)
+ return urlunsplit((parts.scheme, parts.netloc, path, query, parts.fragment))
def caches_at_the_last_block(
@@ -1984,7 +1986,7 @@ class ExternalProviderClient:
if response_format is not None:
body["response_format"] = response_format
- url = f"{self.base_url}/chat/completions"
+ url = _append_provider_path(self.base_url, "/chat/completions")
logger.info(
"Proxying chat completion to %s (provider=%s, model=%s)",
url,
@@ -2227,7 +2229,7 @@ class ExternalProviderClient:
reasoning_effort: Optional[str] = None,
) -> AsyncGenerator[str, None]:
"""Two requests: first is internal; second echoes tool args verbatim as the server runs the search."""
- url = f"{self.base_url}/chat/completions"
+ url = _append_provider_path(self.base_url, "/chat/completions")
body: dict[str, Any] = {
"model": model,
"messages": messages,
@@ -3174,7 +3176,7 @@ class ExternalProviderClient:
if fast_mode_active:
body["speed"] = "fast"
- url = f"{self.base_url}/messages"
+ url = _append_provider_path(self.base_url, "/messages")
completion_id = f"chatcmpl-anthropic-{model.replace('/', '-')}"
# Log outgoing config keys (not messages) to prove which thinking / effort fields reached the wire.
@@ -4889,7 +4891,7 @@ class ExternalProviderClient:
body["cachedContent"] = enable_prompt_caching
# Model id is validated at the top of _stream_gemini, so the URL path cannot be traversed.
- url = f"{self.base_url}/models/{model}:streamGenerateContent?alt=sse"
+ url = _append_provider_path(self.base_url, f"/models/{model}:streamGenerateContent?alt=sse")
completion_id = f"chatcmpl-gemini-{model.replace('/', '-')}"
logger.info(
@@ -7247,7 +7249,7 @@ class ExternalProviderClient:
else:
body["max_tokens"] = max_tokens
- url = f"{self.base_url}/chat/completions"
+ url = _append_provider_path(self.base_url, "/chat/completions")
headers = self._chat_headers(None)
response = await _client().post(
url,
@@ -7377,7 +7379,7 @@ class ExternalProviderClient:
headers = self._auth_headers()
headers.pop("Content-Type", None)
response = await _client().post(
- f"{self.base_url}/audio/transcriptions",
+ _append_provider_path(self.base_url, "/audio/transcriptions"),
headers = headers,
files = {"file": (filename, audio, content_type)},
data = data,
@@ -7468,7 +7470,7 @@ class ExternalProviderClient:
return cached[1]
try:
response = await _client().get(
- f"{self.base_url}/models",
+ _append_provider_path(self.base_url, "/models"),
headers = self._auth_headers(),
timeout = self._timeout,
)
@@ -7494,7 +7496,7 @@ class ExternalProviderClient:
async def _models_payload(self, timeout: Any) -> tuple[Any, list[dict[str, Any]]]:
response = await _client().get(
- f"{self.base_url}/models",
+ _append_provider_path(self.base_url, "/models"),
# Anthropic pages 20 models and native Gemini 50 by default; both cap a page at 1000. A Gemini
# connection through a proxy gets no native parameter it may reject.
params = (
@@ -7531,7 +7533,7 @@ class ExternalProviderClient:
break
seen.add(token)
response = await _client().get(
- f"{self.base_url}/models",
+ _append_provider_path(self.base_url, "/models"),
params = params.set("pageToken", token),
headers = self._auth_headers(),
timeout = timeout,
@@ -7764,7 +7766,7 @@ class ExternalProviderClient:
"""Confirm GET /models returns 200 without buffering the full response body. Used for
providers with enormous catalogs (e.g. OpenRouter, Hugging Face router) where downloading
the full JSON would be prohibitive."""
- url = f"{self.base_url}/models"
+ url = _append_provider_path(self.base_url, "/models")
try:
async with _client().stream(
"GET",
@@ -7789,7 +7791,7 @@ class ExternalProviderClient:
router answer /models without one, so that check alone passes any key. Errors other than
401/403 count as not rejected."""
if self.provider_type == "openrouter":
- url = f"{self.base_url}/key"
+ url = _append_provider_path(self.base_url, "/key")
elif (
self.provider_type == "huggingface"
and (urlparse(self.base_url).hostname or "").lower() == "router.huggingface.co"
@@ -7818,7 +7820,7 @@ class ExternalProviderClient:
them). Only valid against api.openai.com (caller guards is_openai_cloud).
"""
response = await _client().get(
- f"{self.base_url}/containers",
+ _append_provider_path(self.base_url, "/containers"),
headers = self._container_headers(),
timeout = self._timeout,
)
@@ -7844,7 +7846,7 @@ class ExternalProviderClient:
},
}
response = await _client().post(
- f"{self.base_url}/containers",
+ _append_provider_path(self.base_url, "/containers"),
json = body,
headers = self._container_headers(),
timeout = self._timeout,
@@ -7854,7 +7856,7 @@ class ExternalProviderClient:
async def delete_openai_container(self, container_id: str) -> None:
"""Fresh client, since shared-pool DELETEs left containers alive; body must report deleted:true."""
- url = f"{self.base_url}/containers/{container_id}"
+ url = _append_provider_path(self.base_url, f"/containers/{container_id}")
headers = self._container_headers()
logger.info(
"openai_container_delete.outbound url=%s has_auth=%s openai_beta=%s",
diff --git a/studio/backend/tests/test_custom_provider_api_type.py b/studio/backend/tests/test_custom_provider_api_type.py
index f3a5cf70768..68b873d65e9 100644
--- a/studio/backend/tests/test_custom_provider_api_type.py
+++ b/studio/backend/tests/test_custom_provider_api_type.py
@@ -51,6 +51,7 @@ def test_invalid_api_type_is_rejected(schema):
("custom", "responses", "responses", ""),
("custom", "responses", "responses", "?api-version=2025-04-01-preview"),
("custom", "chat_completions", "chat/completions", ""),
+ ("custom", "chat_completions", "chat/completions", "?api-version=2025-04-01-preview"),
("openai", "chat_completions", "responses", ""),
],
)
@@ -114,6 +115,83 @@ def test_selected_endpoint_and_payload(monkeypatch, provider_type, api_type, end
assert body["messages"][0]["content"] == "Be brief"
+def test_base_url_query_survives_model_list_and_non_stream_calls(monkeypatch):
+ sent = []
+
+ def handle(request):
+ sent.append(request)
+ if request.url.path == "/v1/models":
+ return httpx.Response(200, json = {"data": [{"id": "gateway-model"}]})
+ if request.url.path == "/v1/chat/completions":
+ return httpx.Response(200, json = {"choices": [{"message": {"content": "pong"}}]})
+ if request.url.path == "/v1/audio/transcriptions":
+ return httpx.Response(200, json = {"text": "hi"})
+ return httpx.Response(404)
+
+ async def run():
+ async with httpx.AsyncClient(transport = httpx.MockTransport(handle)) as transport:
+ monkeypatch.setattr(ep, "_http_client", transport)
+ client = ExternalProviderClient(
+ "custom", "https://gateway.example/v1?api-version=preview", "test-key"
+ )
+ assert [m["id"] for m in await client.list_models()] == ["gateway-model"]
+ await client.verify_models_endpoint_lightweight()
+ await client.chat_completion(
+ messages = [{"role": "user", "content": "ping"}], model = "gateway-model"
+ )
+ await client.create_transcription(b"RIFF", "a.wav", "audio/wav", "whisper-1")
+
+ asyncio.run(run())
+ assert [(r.url.path, r.url.query.decode()) for r in sent] == [
+ ("/v1/models", "api-version=preview"),
+ ("/v1/models", "api-version=preview"),
+ ("/v1/chat/completions", "api-version=preview"),
+ ("/v1/audio/transcriptions", "api-version=preview"),
+ ]
+
+
+def test_base_url_query_survives_native_provider_calls(monkeypatch):
+ sent = []
+
+ def handle(request):
+ sent.append(request)
+ if request.url.path.endswith("/models"):
+ return httpx.Response(200, json = {"data": []})
+ return httpx.Response(200, json = {})
+
+ async def drain(client, model):
+ async for _ in client.stream_chat_completion(
+ messages = [{"role": "user", "content": "ping"}], model = model
+ ):
+ pass
+
+ async def run():
+ async with httpx.AsyncClient(transport = httpx.MockTransport(handle)) as transport:
+ monkeypatch.setattr(ep, "_http_client", transport)
+ monkeypatch.setattr(ep, "_openrouter_catalog", {})
+ await drain(
+ ExternalProviderClient("anthropic", "https://gw.example/v1?tenant=a", "k"),
+ "claude-sonnet-4-5",
+ )
+ await drain(
+ ExternalProviderClient(
+ "gemini", "https://generativelanguage.googleapis.com/v1beta?tenant=a", "k"
+ ),
+ "gemini-2.5-flash",
+ )
+ router = ExternalProviderClient("openrouter", "https://gw.example/api/v1?tenant=a", "k")
+ await router.list_models()
+ await router.api_key_rejected()
+
+ asyncio.run(run())
+ assert [(r.url.path, r.url.query.decode()) for r in sent] == [
+ ("/v1/messages", "tenant=a"),
+ ("/v1beta/models/gemini-2.5-flash:streamGenerateContent", "tenant=a&alt=sse"),
+ ("/api/v1/models", "tenant=a"),
+ ("/api/v1/key", "tenant=a"),
+ ]
+
+
@pytest.mark.parametrize(
"base_url,api_type,key,expected",
[