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
Fix the diffusion scheduler, parity, formatter and bundle reds on main (#13612)
Daniel Hanb8d76e1696cd
diff --git a/studio/backend/core/inference/diffusion.py b/studio/backend/core/inference/diffusion.pyindex 29a58a19d1b..729bb65280a 100644--- a/studio/backend/core/inference/diffusion.py+++ b/studio/backend/core/inference/diffusion.py@@ -10742,7 +10742,11 @@ class DiffusionBackend:# reconciles.self._make_vae_encode_dtype_safe(pipe)- swapped, resolved_schedule = build_scheduler(pipe.scheduler, sampler, schedule_type)+ # getattr like the status poll: a pipeline without a scheduler offers no samplers, so the default+ # request keeps it as is and any other sampler is refused by build_scheduler.+ swapped, resolved_schedule = build_scheduler(+ getattr(pipe, "scheduler", None), sampler, schedule_type+ )if swapped is not None:restore_scheduler = (pipe, pipe.scheduler)pipe.scheduler = swapped@@ -11944,7 +11948,6 @@ def _pipeline_components_from_index(raise ValueError(f"{repo_id}/model_index.json is not an object")if explicit_components is None:from .diffusion_component_sizes import manifest_entry_loaded-selected = {str(name) for name, spec in payload.items() if manifest_entry_loaded(name, spec)}diff --git a/studio/backend/tests/test_diffusion_autoopt_parity.py b/studio/backend/tests/test_diffusion_autoopt_parity.pyindex a4547f8e599..d26fdf34e5e 100644--- a/studio/backend/tests/test_diffusion_autoopt_parity.py+++ b/studio/backend/tests/test_diffusion_autoopt_parity.py@@ -513,6 +513,9 @@ def test_fixture_covers_every_registered_family(captured):def test_autoopt_resolves_the_same_optimizations_as_before(captured):+ # The fixture was captured with diffusers installed; without it nothing compiles and the diffusers patches+ # self-disable, so every case differs. Backend CI has no diffusers.+ pytest.importorskip("diffusers")got = snapshot(autoopt_entry)want = captured["cases"]assert set(got) == set(want)@@ -522,6 +525,7 @@ def test_autoopt_resolves_the_same_optimizations_as_before(captured):def test_legacy_sequence_still_matches_fixture(captured):"""The fixture is not stale: the old sequence over today's primitives still produces it."""+ pytest.importorskip("diffusers") # see test_autoopt_resolves_the_same_optimizations_as_beforegot = snapshot(legacy_entry)want = captured["cases"]diffs = [case for case in want if got.get(case) != want[case]]diff --git a/studio/backend/tests/test_diffusion_component_sizes.py b/studio/backend/tests/test_diffusion_component_sizes.pyindex 8642c23f733..14f6cedd05c 100644--- a/studio/backend/tests/test_diffusion_component_sizes.py+++ b/studio/backend/tests/test_diffusion_component_sizes.py@@ -948,8 +948,6 @@ def test_transformer_2_counts_only_for_moe(tmp_path):assert dcs.video_header_terms(_fam("wan2.2-t2v-a14b"), sizes)[0] == pytest.approx(4.0)--def test_pytorch_model_stem_and_its_index_are_recognised(tmp_path):t = tmp_path / "transformer"for i in (1, 2):@@ -1105,8 +1103,6 @@ def test_sub_byte_floats_count_parameters_by_bit_width(tmp_path):assert dcs.FLOAT_BITS["F4"] == 4 and dcs.FLOAT_BITS["F6_E2M3"] == dcs.FLOAT_BITS["F6_E3M2"] == 6--def test_unresolved_ref_probes_the_hub_and_falls_back_to_local(tmp_path, monkeypatch):fam = _fam("wan2.2-ti2v-5b")snap = tmp_path / ("a" * 40)@@ -1175,8 +1171,6 @@ def test_same_stem_order_follows_the_loader(tmp_path):assert sizes.get("transformer").files == ("diffusion_pytorch_model-00001-of-00001.safetensors",)--def test_hub_skips_folders_the_manifest_does_not_load(monkeypatch):files = ["model_index.json",@@ -1288,8 +1282,6 @@ def test_local_index_read_is_capped(tmp_path, monkeypatch):assert sizes is None and "diffusion_pytorch_model.safetensors.index.json" not in reads--def test_snapshot_missing_a_loaded_role_is_incomplete(tmp_path, monkeypatch):import shutil@@ -1346,8 +1338,6 @@ def test_index_fetch_is_bounded_by_the_deadline_while_a_peer_trickles():release.set()--def test_ltx_bundled_connector_tensors_leave_the_denoiser_term(tmp_path):fam = _fam("ltx-2.5")_stub_for_table(tmp_path, fam)@@ -1439,8 +1429,6 @@ def test_moe_second_expert_is_optional_unless_shipped(tmp_path):] == pytest.approx(28.6)--def test_partial_term_keeps_its_known_sum_as_a_lower_bound(tmp_path, monkeypatch):import shutil@@ -1574,8 +1562,6 @@ def test_calibration_key_carries_the_revision():assert a != vsr.calibration_key("ltx-2", "r", "group", "cuda:0", revision = "b" * 40)--def test_manifest_predicate_matches_the_loader(tmp_path):import inspect@@ -1667,8 +1653,6 @@ def test_retagged_peak_is_recorded_under_the_mode_that_ran(monkeypatch):assert vsr.measured_extra_mib(predicted, 64) is None--def test_pre_download_plan_threads_the_dev_flag():import inspectdiff --git a/studio/backend/tests/test_h3_native_memory_fit.py b/studio/backend/tests/test_h3_native_memory_fit.pyindex ee50aa38fc7..b60d73d2d6b 100644--- a/studio/backend/tests/test_h3_native_memory_fit.py+++ b/studio/backend/tests/test_h3_native_memory_fit.py@@ -188,7 +188,9 @@ def test_gtt_used_on_another_amd_card_shrinks_the_selected_cards_pool(tmp_path):_card(tmp_path, "card1", "0000:c5:00.0", total = 24 * GIB, used = 12 * GIB)assert amdgpu_gtt_free_bytes("0000:03:00", sysfs_root = str(tmp_path)) == 10 * GIB# Only the selected card at work: the per-device bound still applies.- (tmp_path / "class" / "drm" / "card1" / "device" / "mem_info_gtt_used").write_text("0\n", encoding = "utf-8")+ (tmp_path / "class" / "drm" / "card1" / "device" / "mem_info_gtt_used").write_text(+ "0\n", encoding = "utf-8"+ )assert amdgpu_gtt_free_bytes("0000:03:00", sysfs_root = str(tmp_path)) == 22 * GIBdiff --git a/studio/frontend/scripts/check-bundle-budget.ts b/studio/frontend/scripts/check-bundle-budget.tsindex 23159e3a567..6b29382fa25 100644--- a/studio/frontend/scripts/check-bundle-budget.ts+++ b/studio/frontend/scripts/check-bundle-budget.ts@@ -70,8 +70,13 @@ export const BUDGET = {// /compact and the in-loop compaction divider add 3.6 KB raw / 1.4 KB transfer to the eager startup// JS (6,563.4 -> 6,567.0 KB raw, 1,973.0 -> 1,974.4 KB transfer against 280c45695, one machine// and build), raised by that.- transferBytes: 2_022_600,- rawBytes: 6_731_800,+ // Five merges since then add 2.1 KB raw / 0.8 KB transfer (6,567.3 -> 6,569.4 KB raw, 1,974.5 ->+ // 1,975.3 KB transfer from a062c04d51 to c35052e42f, one machine and build). The step that crossed+ // is the token activity Chart view (#13558, +0.7 KB raw / +0.3 KB transfer): its persisted store+ // pulls zustand/middleware into the startup set as a 99th chunk. Raised by that growth so main's+ // margin stays where it was.+ transferBytes: 2_023_500,+ rawBytes: 6_734_000,};// The chunk count is reported but not budgeted. Splitting a page out of the entry raises it while lowering the