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
"""One persistent denoise thread per backend per card: cuDNN caches conv/SDPA plans thread_local (ATen Conv_v8.cpp, MHA.cpp)."""
from __future__ import annotations
import contextvars
import os
import threading
from concurrent.futures import ThreadPoolExecutor
from typing import Any, Callable
from .rocm_bf16 import is_rocm_torch
_ENV = "UNSLOTH_DIFFUSION_RENDER_THREAD"
_EXECUTORS: dict[str, ThreadPoolExecutor] = {}
_RENDER_THREAD_IDS: set[int] = set()
_LOCK = threading.Lock()
def enabled() -> bool:
if os.environ.get(_ENV, "").strip().lower() in ("0", "false", "no", "off"):
return False
try:
import torch # noqa: PLC0415
return bool(torch.cuda.is_available()) and not is_rocm_torch(torch)
except Exception: # noqa: BLE001
return False
def _mark_render_thread() -> None:
_RENDER_THREAD_IDS.add(threading.get_ident())
def _executor(name: str) -> ThreadPoolExecutor:
with _LOCK:
ex = _EXECUTORS.get(name)
if ex is None:
ex = ThreadPoolExecutor(
max_workers = 1,
thread_name_prefix = f"unsloth-{name}-render",
initializer = _mark_render_thread,
)
_EXECUTORS[name] = ex
return ex
def _apply_compile_config() -> None:
try:
from . import diffusion_compile_config # noqa: PLC0415
diffusion_compile_config.apply()
except Exception: # noqa: BLE001 - never fail a render over a config write
pass
def run(name: str, fn: Callable[[], Any]) -> Any:
if threading.get_ident() in _RENDER_THREAD_IDS or not enabled():
_apply_compile_config()
return fn()
import torch # noqa: PLC0415
device = torch.cuda.current_device()
inference = torch.is_inference_mode_enabled()
grad = torch.is_grad_enabled()
ctx = contextvars.copy_context()
def call() -> Any:
torch.cuda.set_device(device)
_apply_compile_config()
if inference:
with torch.inference_mode():
return fn()
with torch.set_grad_enabled(grad):
return fn()
return _executor(f"{name}-cuda{device}").submit(ctx.run, call).result()