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.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
# 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 json
import httpx
import pytest
from fastmcp.exceptions import ToolError
from studio_mcp.errors import audio_code, raise_for_audio, raise_for_route
from .mcp_harness import openai_error
OOM = "Kokoro ran out of memory. Free GPU memory, or set the audio device to CPU."
VARIANTS = [{"variant": "Q4_K_M", "label": "Q4_K_M", "size_bytes": 1}, {"variant": "Q3_K_M"}]
def _api(status, detail, **headers):
content = json.dumps({"detail": detail}).encode()
return httpx.Response(status, content = content, headers = headers)
def _v1(status, message, **fields):
sent = openai_error(message, status, **fields)
return httpx.Response(status, content = sent.body, headers = sent.raw_headers)
def _message(resp, **kwargs):
with pytest.raises(ToolError) as caught:
raise_for_audio(resp, **kwargs)
return str(caught.value)
def test_out_of_memory_names_cpu_and_the_smaller_variants():
detail = {"code": "out_of_memory", "message": OOM, "smaller_variants": VARIANTS}
message = _message(_api(503, detail))
assert message.startswith("[out_of_memory] Kokoro ran out of memory.")
assert "device='cpu'" in message
assert message.endswith("load_model(variant=...): Q4_K_M, Q3_K_M.")
def test_the_v1_envelope_code_gets_the_tag_and_the_generic_hint():
message = _message(_v1(503, "Out of memory.", type = "server_error", code = "out_of_memory"))
assert message.startswith("[out_of_memory] Out of memory.")
assert message.endswith("Load the audio model again with device='cpu'.")
# The envelope's "(out_of_memory)" suffix is dropped; the tag names the code once.
assert message.count("out_of_memory") == 1
def test_gpu_unavailable_points_at_the_cpu():
detail = {"code": "gpu_unavailable", "message": "The audio runtime could not use the GPU."}
message = _message(_api(503, detail))
assert message.startswith("[gpu_unavailable] ")
assert message.endswith("Load the audio model again with device='cpu'.")
def test_runtime_stopped_says_retry_once():
detail = {"code": "runtime_stopped", "message": "The audio runtime stopped: exit 9."}
message = _message(_api(500, detail))
assert message.startswith("[runtime_stopped] The audio runtime stopped")
assert "Retry once" in message
def test_an_already_tagged_message_is_not_tagged_twice():
detail = {"code": "out_of_memory", "message": "[out_of_memory] Out of memory."}
message = _message(_api(503, detail))
assert message.startswith("[out_of_memory] Out of memory.")
assert message.count("[out_of_memory]") == 1
def test_a_gpu_busy_answer_keeps_its_retry_and_gets_no_audio_tag():
resp = _v1(409, "Busy.", code = "gpu_busy", headers = {"Retry-After": "5"})
assert _message(resp) == "[gpu_busy retry_after=5] GPU busy: Busy. Retry after 5 s."
@pytest.mark.parametrize(
"resp",
[
_api(400, "No model loaded."),
_api(500, {"code": "something_else", "message": "Broke."}),
_v1(400, "No model loaded."),
],
)
def test_other_errors_are_unchanged(resp):
message = _message(resp, hints = {400: "Load one."})
with pytest.raises(ToolError) as generic:
raise_for_route(resp, hints = {400: "Load one."})
# No audio code, so the audio path leaves the generic error as it is.
assert message == str(generic.value)
assert audio_code(json.loads(resp.content)) is None
def test_a_success_is_returned_as_is():
resp = httpx.Response(200, content = b'{"clips": []}')
assert raise_for_audio(resp) == {"clips": []}