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 tests: plan as Linux does whatever the host can compile (#17135)
Daniel Hanb7dc64faacf6
diff --git a/studio/backend/tests/test_diffusion_backend.py b/studio/backend/tests/test_diffusion_backend.pyindex 3f792908b7d..539b6633866 100644--- a/studio/backend/tests/test_diffusion_backend.py+++ b/studio/backend/tests/test_diffusion_backend.py@@ -8,6 +8,7 @@ from __future__ import annotationsimport contextlibimport dataclassesimport json+import osimport reimport sysimport threading@@ -63,6 +64,15 @@ def _unmeasured_torchao(monkeypatch):monkeypatch.setattr(diffusion_memory, "_installed_torchao_version", lambda: None)# Studio's diffusers pin, so tests that opt into a measured torchao do not depend on the runner.monkeypatch.setattr(diffusion_memory, "_installed_diffusers_version", lambda: (0, 40))+ # Whether this host can torch.compile (Triton and MSVC on Windows) must not decide the plans either: answer as+ # Linux does, TORCHDYNAMO_DISABLE still honoured. Tests of the declined path patch compile_eligible themselves.+ from core.inference import diffusion_speed++ monkeypatch.setattr(+ diffusion_speed,+ "torch_compile_runtime_available",+ lambda: os.environ.get("TORCHDYNAMO_DISABLE", "").strip() in ("", "0"),+ )def test_clamp_max_side_bounds_oversized_init():@@ -11286,7 +11296,11 @@ def test_generate_allow_oversized_runs_a_refused_request(fake_runtime, tmp_path,def test_generate_guard_leaves_a_large_batch_to_the_oom_backoff(fake_runtime, tmp_path, monkeypatch):- # The chunk loop halves failed batches to singletons, so the guard budgets one image.+ # The chunk loop halves failed batches to singletons, so the guard budgets one image. Off Windows only: WDDM+ # never raises an OOM, so there _activation_guard_batch budgets the whole chunk instead.+ from core.inference import diffusion as dmod++ monkeypatch.setattr(dmod.sys, "platform", "linux")backend = _loaded_backend_on_a_16g_card(tmp_path, monkeypatch)out = backend.generate(prompt = "a sloth", width = 1024, height = 1024, steps = 4, batch_size = 8)assert len(out["images"]) == 8diff --git a/studio/backend/tests/test_video_backend.py b/studio/backend/tests/test_video_backend.pyindex ccac2a8a42d..405ffec85e2 100644--- a/studio/backend/tests/test_video_backend.py+++ b/studio/backend/tests/test_video_backend.py@@ -2776,7 +2776,7 @@ def test_wan_a14b_gguf_pair_loads_high_as_transformer_and_low_as_transformer_2(pick = high if picked == "high" else lowVideoBackend().load_pipeline(str(tmp_path),- gguf_filename = str(pick.relative_to(tmp_path)),+ gguf_filename = pick.relative_to(tmp_path).as_posix(), # a hub path, '/' on every OSbase_repo = "Wan-AI/Wan2.2-T2V-A14B-Diffusers",family_override = "wan2.2-t2v-a14b",)