"""The diffusion decoder is built once, at start-up, with the other parts (#17). On the Space the default decode is the diffusion decoder. It used to be built lazily inside each GPU worker and lost with it, so every run loaded its 0.83 GB and resolved its NATTEN kernel again. ``app.py`` now builds it at start-up, after the pipeline and before ``demo.launch()``, so its ``.to("cuda")`` is registered for packing with the rest; each worker's runner finds that instance through its decoder factory. A failed start-up build logs one line and leaves today's per-run build in place. Two tiers: the CUDA path against the fake diffusers tree the stage-graph spec uses (``.to("cuda")``, the NATTEN processor), and the worker path end to end on the real tiny checkpoint, through ``generate()``. """ from __future__ import annotations import ast import logging import pytest from conftest import REPO_ROOT from test_stage_graph_contract import FakeModel, LoadRecorder, install_fake_stack from ltx25_studio import config, model_files, pipelines DECODER = "LTX2VideoDiffusionDecoderModel" @pytest.fixture(autouse=True) def _no_cached_pipelines(): pipelines.clear_cache() yield pipelines.clear_cache() # -------------------------------------------------------------------------- # the CUDA path, on the fake diffusers tree # -------------------------------------------------------------------------- class DecodePipelineRecorder: """``LTX2VideoDiffusionDecodePipeline``: records each build, returns a marked object.""" def __init__(self): self.built = [] def __call__(self, **kwargs): made = FakeModel() made.parts = kwargs self.built.append(made) return made class NattenOnce: """The NATTEN processor: the first construction fails as a Hub outage would.""" def __init__(self): self.calls = 0 def __call__(self): self.calls += 1 if self.calls == 1: raise RuntimeError("Cannot reach the Hub to resolve shi-labs/natten") return "natten-processor" @pytest.fixture def space(tmp_path, monkeypatch): """The ZeroGPU start-up's pieces on the fake tree: a cache holding the decoder at its own revision, the two runners ``app.py`` builds, and the recorders.""" def build(**classes): monkeypatch.delenv(config.ENV_MODEL_SNAPSHOT_DIR, raising=False) monkeypatch.setenv("HF_HUB_CACHE", str(tmp_path)) repo = tmp_path / f"models--{model_files.REPO_ID.replace('/', '--')}" / "snapshots" snapshot = repo / model_files.REVISION snapshot.mkdir(parents=True) (repo / model_files.DIFFUSION_DECODER_REVISION / "diffusion_decoder").mkdir(parents=True) model = FakeModel() decode = classes.pop("LTX2VideoDiffusionDecodePipeline", None) or DecodePipelineRecorder() base = FakeModel() base.components = {} # what `derive_pipeline` shares into the text2video runner recorders = install_fake_stack( monkeypatch, snapshot, LTX2ConditionPipeline=LoadRecorder(result=base), LTX2VideoDiffusionDecoderModel=LoadRecorder(result=model), LTX2VideoDiffusionDecodePipeline=decode, **classes, ) runners = { kind: pipelines.get_pipeline(pipeline=kind, device="cuda") for kind in ("condition", "text2video") } return model, decode, recorders, runners return build def test_the_startup_build_loads_the_decoder_once_and_moves_it_to_cuda(space, caplog): model, decode, recorders, _runners = space() assert not recorders[DECODER].calls, "building the runners must still not load the decoder" with caplog.at_level(logging.INFO, logger="ltx25_studio"): built = pipelines.build_decoder_at_startup(device="cuda") assert built is True assert len(recorders[DECODER].calls) == 1 assert model.moved_to == "cuda", "the .to('cuda') that ZeroGPU registers for packing" assert model.processor is not None, "the NATTEN processor is set at start-up too" assert model.tiled is True assert len(decode.built) == 1 lines = [r.getMessage() for r in caplog.records if "diffusion decoder" in r.getMessage()] assert any("built at start-up" in line and "GiB" in line for line in lines), lines def test_each_runner_finds_the_startup_decoder_instead_of_building_one(space): _model, decode, recorders, runners = space() pipelines.build_decoder_at_startup(device="cuda") startup_decoder = decode.built[0] for runner in runners.values(): decoder, why_conv = runner._single_stage_decoder("auto", {}) assert (decoder, why_conv) == ("diffusion", None) assert runner.decoder is startup_decoder assert len(recorders[DECODER].calls) == 1, "a run loaded the decoder again" assert len(decode.built) == 1 def test_a_second_startup_call_builds_nothing_more(space): _model, decode, recorders, _runners = space() assert pipelines.build_decoder_at_startup(device="cuda") is True assert pipelines.build_decoder_at_startup(device="cuda") is True assert len(recorders[DECODER].calls) == 1 and len(decode.built) == 1 def test_a_failed_startup_build_logs_one_line_and_the_run_builds_its_own(space, caplog): natten = NattenOnce() _model, decode, recorders, runners = space( LTX2VideoVaeNeighborhoodNattenProcessor=natten ) with caplog.at_level(logging.INFO, logger="ltx25_studio"): built = pipelines.build_decoder_at_startup(device="cuda") assert built is False warnings = [r for r in caplog.records if r.levelno >= logging.WARNING] assert len(warnings) == 1, [r.getMessage() for r in warnings] message = warnings[0].getMessage() assert "NOT built at start-up" in message and "each run builds its own" in message, message assert "Cannot reach the Hub" in message and "\n" not in message, message assert not decode.built, "a half-built decoder was kept" decoder, why_conv = runners["text2video"]._single_stage_decoder("auto", {}) assert (decoder, why_conv) == ("diffusion", None), "the per-run build is gone" assert len(recorders[DECODER].calls) == 2 and len(decode.built) == 1 class WrapperOnce(DecodePipelineRecorder): """``LTX2VideoDiffusionDecodePipeline`` whose first construction fails.""" def __call__(self, **kwargs): if not getattr(self, "failed", False): self.failed = True raise TypeError("LTX2VideoDiffusionDecodePipeline got an unexpected component") return super().__call__(**kwargs) def test_a_failure_after_the_load_leaves_the_decoder_unmoved_and_the_run_builds_it(space): """ZeroGPU packs whatever is moved at start-up and then prunes the file it came from. A decoder moved but never cached would leave the per-run build a pruned file, so the move comes last: a build that fails at any step after the load has moved nothing.""" wrapper = WrapperOnce() model, _decode, recorders, runners = space(LTX2VideoDiffusionDecodePipeline=wrapper) assert pipelines.build_decoder_at_startup(device="cuda") is False assert model.moved_to is None, "the decoder was moved to cuda by a build that then failed" assert "__decoder__" not in pipelines._CACHE decoder, why_conv = runners["condition"]._single_stage_decoder("auto", {}) assert (decoder, why_conv) == ("diffusion", None) assert runners["condition"].decoder is wrapper.built[0] assert model.moved_to == "cuda" and len(recorders[DECODER].calls) == 2 def test_a_failed_startup_build_never_raises(space, monkeypatch): """Whatever goes wrong, start-up goes on: the Space must come up as before.""" _model, _decode, _recorders, _runners = space() def broken(*args, **kwargs): raise ImportError("kernels is not installed") monkeypatch.setattr(pipelines, "_decoder_part", broken) assert pipelines.build_decoder_at_startup(device="cuda") is False def test_the_app_builds_the_decoder_at_module_scope_after_the_pipeline(): """Where ``app.py`` asks for it: in the CUDA branch, after both runners, at module scope -- inside ``@spaces.GPU`` it would never be packed.""" tree = ast.parse((REPO_ROOT / "app.py").read_text()) branch = next( node for node in tree.body if isinstance(node, ast.If) and "get_pipeline" in ast.unparse(node) ) calls = [ ast.unparse(statement.value) for statement in branch.body if isinstance(statement, ast.Expr) and isinstance(statement.value, ast.Call) ] assert calls[-1] == "pipelines.build_decoder_at_startup(device='cuda')", calls assert sum("get_pipeline" in call for call in calls[:-1]) == 2, calls def test_the_startup_report_names_the_decoder_beside_the_pipeline(monkeypatch): from ltx25_studio import device, zerogpu monkeypatch.setattr(zerogpu, "host_has_cuda", lambda: True) monkeypatch.setenv(config.ENV_DEVICE, "cuda") runtime_config = config.load_config() report = pipelines.startup_report( runtime_config, missing_files=(), resolution=device.resolve(None, runtime_config) ) line = next(line for line in report.splitlines() if line.strip().startswith("runner")) assert "diffusion decoder" in line and "start" in line, line # -------------------------------------------------------------------------- # the worker path, end to end, on the real tiny checkpoint # -------------------------------------------------------------------------- def _tiny_composite(tmp_path_factory): pytest.importorskip("torch") pytest.importorskip("diffusers") from test_real_pipeline_tiny import _tiny_snapshot_cached, load_smoke if not _tiny_snapshot_cached(): pytest.skip(f"{model_files.TINY_REPO_ID} is not in the HF cache") smoke = load_smoke() snapshot, missing = smoke.composite_snapshot( smoke.tiny_snapshot(fetch=False), tmp_path_factory.mktemp("tiny") / "snapshot", fetch=False ) if missing: pytest.skip(missing) return snapshot @pytest.fixture(scope="module") def tiny_composite_dir(tmp_path_factory): return _tiny_composite(tmp_path_factory) @pytest.fixture def tiny_space(tiny_composite_dir, monkeypatch): """``app.py``'s start-up on the tiny checkpoint, on the CPU: both runners, then the decoder. Records every decoder load and every decode call.""" from diffusers.pipelines.ltx2.pipeline_ltx2_diffusion_decode import ( LTX2VideoDiffusionDecodePipeline, ) monkeypatch.setenv(config.ENV_MODEL_SNAPSHOT_DIR, str(tiny_composite_dir)) # The process's device, as LTX25_DEVICE sets it: the worker's own runner lookup # (`get_pipeline` with no device) resolves it. monkeypatch.setenv(config.ENV_DEVICE, "cpu") loads, decodes = [], [] real_load = pipelines._load_part def counting_load(cls, component, **kwargs): if component == "diffusion_decoder": loads.append(component) return real_load(cls, component, **kwargs) monkeypatch.setattr(pipelines, "_load_part", counting_load) real_call = LTX2VideoDiffusionDecodePipeline.__call__ def recording_call(self, *args, **kwargs): decodes.append(self) return real_call(self, *args, **kwargs) monkeypatch.setattr(LTX2VideoDiffusionDecodePipeline, "__call__", recording_call) for kind in ("condition", "text2video"): pipelines.get_pipeline(pipeline=kind, device="cpu") return loads, decodes def _run(tmp_path, name): """One run as the GPU worker makes it: inside the worker context ``app.run_generation`` sets (so no dispatch another test left registered can intercept it), through ``generate()`` and its default runner lookup (``pipelines.get_pipeline``), no factory injected. 128x128x9 is the tiny decoder's minimum and within the CPU decode bound.""" from ltx25_studio import generate token = generate._IN_GPU_WORKER.set(True) try: return generate.generate( mode="text_to_video", prompt="a lighthouse at dusk", seed=7, width=128, height=128, num_frames=9, frame_rate=24.0, decoder="diffusion", output_root=tmp_path / name, retention_bytes=10**9, ) finally: generate._IN_GPU_WORKER.reset(token) def _decode_is_timed(result) -> None: """The run's time line still has its decode phase, after sampling, and adds up.""" from test_run_report_lines import _adds_up, took_lines (line,) = took_lines(result) labels = [label for label, _seconds in _adds_up(line)] assert "decoding the frames" in labels, line assert labels.index("decoding the frames") > labels.index("sampling"), line def test_tiny_the_decoder_is_built_once_and_every_run_uses_that_instance(tiny_space, tmp_path): loads, decodes = tiny_space assert pipelines.build_decoder_at_startup(device="cpu") is True startup_decoder = pipelines._CACHE["__decoder__"] first = _run(tmp_path, "one") second = _run(tmp_path, "two") assert loads == ["diffusion_decoder"], f"the decoder was loaded {len(loads)} times" assert decodes == [startup_decoder, startup_decoder] assert first.output_path.is_file() and second.output_path.is_file() _decode_is_timed(first) _decode_is_timed(second) def test_tiny_a_failed_startup_build_falls_back_to_the_per_run_build( tiny_space, tmp_path, monkeypatch, caplog ): loads, decodes = tiny_space real_part = pipelines._decoder_part failures = [] def fails_at_startup(*args, **kwargs): if not failures: failures.append(1) raise OSError("the decoder's file could not be read") return real_part(*args, **kwargs) monkeypatch.setattr(pipelines, "_decoder_part", fails_at_startup) with caplog.at_level(logging.WARNING, logger="ltx25_studio"): assert pipelines.build_decoder_at_startup(device="cpu") is False result = _run(tmp_path, "one") fell_back = [ r for r in caplog.records if r.getMessage().startswith("diffusion decoder NOT built") ] assert len(fell_back) == 1, [r.getMessage() for r in caplog.records] assert loads == ["diffusion_decoder"], "the run did not build its own decoder" assert len(decodes) == 1 and decodes[0] is pipelines._CACHE["__decoder__"] assert result.output_path.is_file() _decode_is_timed(result)