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
Studio diffusion: store request peaks for Wan pipelines with an empty transformer_2 slot so measured residency runs (#14864)
Daniel Hanca7b167160c0
diff --git a/studio/backend/core/inference/diffusion_memory.py b/studio/backend/core/inference/diffusion_memory.pyindex e1cefbfe711..1649df23558 100644--- a/studio/backend/core/inference/diffusion_memory.py+++ b/studio/backend/core/inference/diffusion_memory.py@@ -3127,6 +3127,8 @@ def stream_window_pad_mib(pipe: Any) -> Optional[int]:for name, module in (getattr(pipe, "components", {}) or {}).items():if str(name) not in ("transformer", "transformer_2", "unconditional_transformer"):continue+ if module is None:+ continue # an empty slot (Wan2.2-TI2V-5B and Wan2.1 register transformer_2 = None)state = _streaming_state(module)if state is None:return None@@ -3159,6 +3161,8 @@ def _streamed_group_window_mib(pipe: Any) -> int:for name, module in (getattr(pipe, "components", {}) or {}).items():if str(name) not in ("transformer", "transformer_2", "unconditional_transformer"):continue+ if module is None:+ continuefor group in _offload_groups(module):if getattr(group, "_unsloth_resident", False):continuediff --git a/studio/backend/tests/test_diffusion_peak_pad_empty_slot.py b/studio/backend/tests/test_diffusion_peak_pad_empty_slot.pynew file mode 100644index 00000000000..7f6d0ec7270--- /dev/null+++ b/studio/backend/tests/test_diffusion_peak_pad_empty_slot.py@@ -0,0 +1,56 @@+# SPDX-License-Identifier: AGPL-3.0-only+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0++"""A ``transformer_2 = None`` slot (Wan2.1, Wan2.2-TI2V-5B) must not make the stream pad None. CPU only."""++from __future__ import annotations++import types++import pytest++import core.inference.diffusion_memory as dm++torch = pytest.importorskip("torch")+++class _Group:+ def __init__(self, module, nbytes):+ self.modules = [module]+ self.parameters = []+ self.buffers = []+ self._nbytes = nbytes+++def streamed(monkeypatch):+ dit = torch.nn.Linear(64, 64)+ groups = {id(dit): [_Group(dit, 64 * 64 * 4)]}+ real = dm._offload_groups # raises on None, as for any non-module++ monkeypatch.setattr(+ dm,+ "_offload_groups",+ lambda module: list(groups[id(module)]) if id(module) in groups else real(module),+ )+ return dit+++def test_pad_skips_an_empty_denoiser_slot(streamed):+ pipe = types.SimpleNamespace(components = {"transformer": streamed, "transformer_2": None})+ assert dm.stream_window_pad_mib(pipe) == 0+ assert (+ dm.stream_window_pad_mib(types.SimpleNamespace(components = {"transformer": streamed})) == 0+ )+++def test_streamed_group_window_skips_an_empty_denoiser_slot(streamed):+ pipe = types.SimpleNamespace(components = {"transformer": streamed, "transformer_2": None})+ assert dm._streamed_group_window_mib(pipe) > 0+++def test_an_unhooked_off_card_denoiser_still_reads_unknown(streamed):+ # unhooked off-card weights stay unknown+ other = torch.nn.Linear(8, 8)+ pipe = types.SimpleNamespace(components = {"transformer": streamed, "transformer_2": other})+ assert dm.stream_window_pad_mib(pipe) is None