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
"""Local Git-only baseline materialization; no network or other worker checkout is read."""
from __future__ import annotations
import io
import json
import os
import subprocess
import sys
import tarfile
from pathlib import Path
REPO = Path(__file__).resolve().parents[4]
PROBE = REPO / "tests/studio/multi_account/perf/probe.py"
SCRATCH = REPO / "temp" / "multi_account_perf"
def baseline_ref() -> str | None:
"""The pre-account commit to measure against: UNSLOTH_STUDIO_PERF_BASE_REF, else the merge base with upstream main."""
explicit = os.environ.get("UNSLOTH_STUDIO_PERF_BASE_REF")
candidates = [explicit] if explicit else []
for upstream in ("origin/main", "upstream/main", "main"):
candidates.append(f"merge-base:{upstream}")
for candidate in candidates:
if candidate.startswith("merge-base:"):
command = ["git", "merge-base", "HEAD", candidate.split(":", 1)[1]]
else:
command = ["git", "rev-parse", "--verify", f"{candidate}^{{commit}}"]
result = subprocess.run(command, cwd = REPO, capture_output = True, text = True)
revision = result.stdout.strip()
if result.returncode == 0 and revision and revision != _head():
return revision
return None
def _head() -> str:
return subprocess.check_output(["git", "rev-parse", "HEAD"], cwd = REPO, text = True).strip()
def materialize_revision(ref: str, destination: Path) -> Path:
destination = destination.resolve()
assert destination.is_relative_to(REPO), "Revision snapshots must stay in this clone"
revision = subprocess.check_output(
["git", "rev-parse", "--verify", f"{ref}^{{commit}}"], cwd = REPO, text = True
).strip()
payload = subprocess.check_output(
["git", "archive", "--format=tar", revision, "studio/backend"], cwd = REPO
)
destination.mkdir(parents = True, exist_ok = False)
with tarfile.open(fileobj = io.BytesIO(payload)) as archive:
for member in archive.getmembers():
assert not member.issym() and not member.islnk()
assert (destination / member.name).resolve().is_relative_to(destination)
archive.extractall(destination, filter = "data")
return destination / "studio/backend"
def run_probe(
backend: Path,
home: Path,
*,
mode: str,
samples: int = 2000,
history_samples: int = 200,
) -> dict:
home = home.resolve()
assert home.is_relative_to(REPO), "Probe data must stay in this clone"
home.parent.mkdir(parents = True, exist_ok = True)
runtime = home.parent / (home.name + "-runtime")
for leaf in ("tmp", "hf", "cache", "Documents"):
(runtime / leaf).mkdir(parents = True, exist_ok = True)
env = dict(os.environ)
env.update(
PYTHONDONTWRITEBYTECODE = "1",
PYTHONPATH = str(backend),
UNSLOTH_STUDIO_HOME = str(home),
HF_HOME = str(runtime / "hf"),
HF_HUB_CACHE = str(runtime / "hf" / "hub"),
HF_HUB_OFFLINE = "1",
TMPDIR = str(runtime / "tmp"),
XDG_CACHE_HOME = str(runtime / "cache"),
UNSLOTH_STUDIO_DOCUMENTS_HOME = str(runtime / "Documents"),
UNSLOTH_ALLOW_CPU = "1",
UNSLOTH_IS_PRESENT = "1",
UNSLOTH_STUDIO_DISABLE_DEVICE_PROBE = "1",
UNSLOTH_DIFFUSION_ATTENTION_INSTALL = "0",
)
result = subprocess.run(
[
sys.executable,
str(PROBE),
"--backend",
str(backend),
"--helpers",
str(Path(__file__).parent),
"--mode",
mode,
"--samples",
str(samples),
"--history-samples",
str(history_samples),
],
cwd = backend,
env = env,
capture_output = True,
text = True,
timeout = 300,
)
assert result.returncode == 0, result.stdout + result.stderr
return json.loads(result.stdout.splitlines()[-1])