Merge remote-tracking branch 'origin/diffusion-controlnet' into diffusion-sdxl

studio/backend/tests/test_diffusion_backend.py

@@ -170,11 +170,13 @@ class _FakePipe:
         self.moved_to = device
         return self
 
-    def enable_model_cpu_offload(self) -> None:
+    def enable_model_cpu_offload(self, device = None) -> None:
         self.offloaded = True
+        self.offload_device = device
 
-    def enable_sequential_cpu_offload(self) -> None:
+    def enable_sequential_cpu_offload(self, device = None) -> None:
         self.sequential_offloaded = True
+        self.offload_device = device
 
     def enable_vae_tiling(self) -> None:
         self.vae_tiled = True
@@ -1214,6 +1216,29 @@ def test_unload_cancels_in_flight_load(fake_runtime):
         )
 
 
+def test_superseded_load_does_not_cancel_live_generation(fake_runtime):
+    # A superseded background load (its token was bumped by a newer load/unload) that
+    # finally reaches load_pipeline must bail WITHOUT signalling the current model's
+    # in-flight generation: the token check has to run before the cancel is set, or a
+    # stale worker aborts an unrelated, still-live denoise.
+    import threading as _threading
+
+    backend = DiffusionBackend()
+    fam = detect_family("unsloth/Z-Image-Turbo-GGUF")
+    live_cancel = _threading.Event()
+    backend._active_generate_cancel = live_cancel  # a generation from the CURRENT model
+    token = 11
+    backend._load_token = token + 1  # this load has already been superseded
+    with pytest.raises(RuntimeError, match = "cancelled"):
+        backend.load_pipeline(
+            "unsloth/Z-Image-Turbo-GGUF",
+            gguf_filename = "z-image-turbo-Q4_K_S.gguf",
+            base_repo = fam.base_repo,
+            _load_token = token,
+        )
+    assert not live_cancel.is_set()  # the live generation was left untouched
+
+
 def test_pick_dtype_bf16_only_on_ampere(fake_runtime, monkeypatch):
     # BF16 only on Ampere+ (cc >= 8); pre-Ampere cards must fall back to FP16.
     torch = sys.modules["torch"]
@@ -1863,3 +1888,41 @@ def test_transformer_quant_skipped_when_plan_offloads(fake_runtime, tmp_path, mo
     assert status["transformer_quant"] is None
     assert status["offload_policy"] == "model"
     assert _FakeTransformer.last["path"]  # GGUF path used
+
+
+def test_reset_step_cache_helper_is_best_effort():
+    # Calls the transformer's reset hook when present.
+    calls = []
+    pipe = types.SimpleNamespace(
+        transformer = types.SimpleNamespace(reset_stateful_hooks = lambda: calls.append(True))
+    )
+    DiffusionBackend._reset_step_cache(pipe)
+    assert calls == [True]
+    # No transformer, or a transformer without the hook -> silent no-op (never raises).
+    DiffusionBackend._reset_step_cache(types.SimpleNamespace())
+    DiffusionBackend._reset_step_cache(types.SimpleNamespace(transformer = object()))
+
+
+def test_generate_resets_step_cache_only_when_engaged(fake_runtime, tmp_path):
+    # FBCache residuals live on the resident transformer across generations, so each
+    # generate() must reset the stateful cache first -- but only when a cache is engaged.
+    (tmp_path / "model.gguf").write_bytes(b"weights")
+    backend = DiffusionBackend()
+    backend.load_pipeline(
+        str(tmp_path),
+        gguf_filename = "model.gguf",
+        base_repo = "base/repo",
+        family_override = "z-image",
+    )
+    resets = []
+    backend._state.pipe.transformer = types.SimpleNamespace(
+        reset_stateful_hooks = lambda: resets.append(True)
+    )
+    # No cache engaged (transformer_cache is None) -> reset must NOT run.
+    backend.generate(prompt = "a sloth")
+    assert resets == []
+    # Engage a cache; every subsequent generation resets the stateful cache first.
+    object.__setattr__(backend._state, "transformer_cache", "fbcache")
+    backend.generate(prompt = "a sloth")
+    backend.generate(prompt = "another sloth")
+    assert resets == [True, True]

studio/backend/tests/test_diffusion_memory.py

@@ -327,16 +327,19 @@ def test_snapshot_never_raises_on_probe_failure(monkeypatch):
 class _RecordingPipe:
     def __init__(self) -> None:
         self.calls: list[str] = []
+        self.offload_device = None
 
     def to(self, device):
         self.calls.append(f"to:{device}")
         return self
 
-    def enable_model_cpu_offload(self):
+    def enable_model_cpu_offload(self, device = None):
         self.calls.append("model_offload")
+        self.offload_device = device
 
-    def enable_sequential_cpu_offload(self):
+    def enable_sequential_cpu_offload(self, device = None):
         self.calls.append("sequential_offload")
+        self.offload_device = device
 
     def enable_vae_tiling(self):
         self.calls.append("vae_tiling")
@@ -385,6 +388,16 @@ def test_apply_model_offload_engages_offload_and_tiling():
     assert "to:cuda" not in pipe.calls  # offload owns placement; never both
     assert "vae_tiling" in pipe.calls and "vae_slicing" in pipe.calls
     assert effective == OFFLOAD_MODEL and tiled is True
+    assert pipe.offload_device == "cuda"  # device threaded to enable_model_cpu_offload
+
+
+def test_apply_model_offload_passes_target_device():
+    # enable_model_cpu_offload defaults to CUDA in diffusers; on a non-CUDA accelerator
+    # (e.g. Intel XPU, which this backend supports) the target device must be forwarded
+    # or diffusers offloads to the wrong backend and the load fails.
+    pipe = _RecordingPipe()
+    apply_memory_plan(pipe, _plan(OFFLOAD_MODEL, tiling = False), device = "xpu")
+    assert pipe.offload_device == "xpu"
 
 
 def test_apply_vae_tiling_falls_back_to_vae_submodule():
@@ -404,7 +417,7 @@ def test_apply_vae_tiling_falls_back_to_vae_submodule():
         def _slice(self):
             self.vae.sliced = True
 
-        def enable_model_cpu_offload(self):
+        def enable_model_cpu_offload(self, device = None):
             self.offloaded = True
 
     pipe = _VaeOnly()
@@ -439,13 +452,14 @@ def test_apply_sequential_offload():
     )
     assert "sequential_offload" in pipe.calls and "to:cuda" not in pipe.calls
     assert effective == OFFLOAD_SEQUENTIAL
+    assert pipe.offload_device == "cuda"  # device threaded to sequential offload too
 
 
 def test_apply_sequential_falls_back_to_model_offload_when_unsupported():
     # Sequential offload is unreliable for GGUF on some diffusers versions; the
     # applier must fall back to whole-module offload and report what actually ran.
     class _NoSeqPipe(_RecordingPipe):
-        def enable_sequential_cpu_offload(self):
+        def enable_sequential_cpu_offload(self, device = None):
             raise RuntimeError("sequential offload not supported for this transformer")
 
     pipe = _NoSeqPipe()

studio/backend/tests/test_diffusion_prequant.py

@@ -66,6 +66,7 @@ class _FakeTransformer:
     def __init__(self):
         self.assigned = None
         self.moved = None
+        self.eval_called = False
 
     @classmethod
     def load_config(cls, base, **kw):
@@ -101,6 +102,10 @@ class _FakeTransformer:
         self.moved = device
         return self
 
+    def eval(self):
+        self.eval_called = True
+        return self
+
 
 def _stub_torch_accelerate(
     monkeypatch,
@@ -182,6 +187,14 @@ def test_load_meta_init_and_assign(monkeypatch, tmp_path):
     assert t._unsloth_runtime_quant == "fp8"
 
 
+def test_load_puts_transformer_in_eval_mode(monkeypatch, tmp_path):
+    # Built via from_config (not from_pretrained), so the loader must eval() it to match
+    # the dense/GGUF paths; otherwise train-mode dropout makes inference nondeterministic.
+    t = _load(monkeypatch, tmp_path, _good_ckpt())
+    assert t is not None
+    assert t.eval_called is True
+
+
 def test_load_missing_file_is_none(monkeypatch, tmp_path):
     assert _load(monkeypatch, tmp_path, _good_ckpt(), exists = False) is None
 

studio/backend/tests/test_sd_cpp_args.py

@@ -222,6 +222,25 @@ def test_build_inpaint_adds_mask():
     assert _pair(cmd, "--init-img") == "/in/src.png"
 
 
+def test_build_inpaint_mask_without_init_img_rejected():
+    # sd-cli inpaint needs a source image; a --mask with no --init-img is invalid argv,
+    # so the builder must reject it up front instead of emitting a doomed command.
+    files = SdCppModelFiles(diffusion_model = "/m/z.gguf")
+    params = SdCppGenParams(prompt = "x", mask = "/in/mask.png")
+    with pytest.raises(ValueError, match = "init_img is required"):
+        build_sd_cpp_command("/bin/sd-cli", files, params, output_path = "/o.png")
+
+
+def test_build_rejects_none_prompt():
+    # A None prompt must be rejected, not coerced to the literal string "None" and
+    # forwarded into argv.
+    files = SdCppModelFiles(diffusion_model = "/m/z.gguf")
+    with pytest.raises(ValueError, match = "prompt is required"):
+        build_sd_cpp_command(
+            "/bin/sd-cli", files, SdCppGenParams(prompt = None), output_path = "/o.png"
+        )
+
+
 def test_build_edit_repeats_ref_image():
     files = SdCppModelFiles(diffusion_model = "/m/flux.gguf")
     params = SdCppGenParams(prompt = "add a hat", ref_images = ("/r/a.png", "/r/b.png"))