Tighten sticky VAE load preflight accounting

Handle sticky tiled fallback for inpaint VAE encodes

patches.py

@@ -686,6 +686,56 @@ def _sticky_vae_free_memory(*, device: Any, patcher: Any) -> int:
     return max(0, int(patcher.get_free_memory(device)))
 
 
+def _sticky_model_load_requirement(*, device: Any, models: tuple[Any, ...]) -> int | None:
+    import comfy.model_management as model_management
+
+    loaded_model_cls = getattr(model_management, "LoadedModel", None)
+    is_device_cpu = getattr(model_management, "is_device_cpu", None)
+    if device is None or loaded_model_cls is None:
+        return None
+    if callable(is_device_cpu) and is_device_cpu(device):
+        return 0
+
+    requested_models: set[Any] = set()
+    for model in models:
+        if model is None:
+            continue
+        requested_models.add(model)
+        additional_models = getattr(model, "model_patches_models", None)
+        if callable(additional_models):
+            try:
+                for additional in additional_models():
+                    if additional is not None:
+                        requested_models.add(additional)
+            except Exception as exc:
+                _LOG.debug(
+                    "GPU Resident Loader: failed to enumerate patched models for sticky VAE preflight: %s",
+                    exc,
+                )
+                return None
+
+    total_required = 0
+    for model in requested_models:
+        try:
+            loaded_model = loaded_model_cls(model)
+            loaded_device = getattr(loaded_model, "device", None)
+            if loaded_device != device:
+                continue
+            if callable(is_device_cpu) and is_device_cpu(loaded_device):
+                continue
+            total_required += max(0, int(loaded_model.model_memory_required(loaded_model.device)))
+        except Exception:
+            return None
+    return total_required
+
+
+def _sticky_model_load_target(model_load_required: int | None) -> int | None:
+    if model_load_required is None:
+        return None
+    required = max(0, int(model_load_required))
+    return (required * 11 + 9) // 10
+
+
 def _prepare_sticky_vae_batch(
     *,
     device: Any,
@@ -694,9 +744,12 @@ def _prepare_sticky_vae_batch(
     total_batch_count: int,
 ) -> tuple[int, int, bool]:
     free_memory = _sticky_vae_free_memory(device=device, patcher=patcher)
+    model_load_required = _sticky_model_load_requirement(device=device, models=(patcher,))
+    model_load_target = _sticky_model_load_target(model_load_required)
+    batch_budget = 0 if model_load_target is None else max(0, free_memory - model_load_target)
     batch_number = _sticky_safe_batch_number(
         batch_count=total_batch_count,
-        free_memory=free_memory,
+        free_memory=batch_budget,
         memory_used=total_memory_used,
         device=device,
     )
@@ -705,7 +758,11 @@ def _prepare_sticky_vae_batch(
     if REGISTRY.get_policy() != "sticky_gpu" or device is None:
         return batch_number, batch_memory_used, False
 
-    target_free = _sticky_protection_target(batch_memory_used, device)
+    target_free = (
+        free_memory + 1
+        if model_load_target is None
+        else model_load_target + _sticky_protection_target(batch_memory_used, device)
+    )
     if free_memory < target_free:
         try:
             trim_resident_vram(
@@ -717,25 +774,36 @@ def _prepare_sticky_vae_batch(
                 keep_models=(patcher,),
             )
         except Exception as exc:
-            _LOG.debug("GPU Resident Loader: proactive VAE trim failed for %s bytes: %s", batch_memory_used, exc)
+            _LOG.debug(
+                "GPU Resident Loader: proactive VAE trim failed for batch=%s load=%s bytes: %s",
+                batch_memory_used,
+                model_load_required,
+                exc,
+            )
 
         free_memory = _sticky_vae_free_memory(device=device, patcher=patcher)
+        batch_budget = 0 if model_load_target is None else max(0, free_memory - model_load_target)
         batch_number = _sticky_safe_batch_number(
             batch_count=total_batch_count,
-            free_memory=free_memory,
+            free_memory=batch_budget,
             memory_used=total_memory_used,
             device=device,
         )
         batch_memory_used = _scaled_batch_memory(total_memory_used, total_batch_count, batch_number)
-        target_free = _sticky_protection_target(batch_memory_used, device)
+        target_free = (
+            free_memory + 1
+            if model_load_target is None
+            else model_load_target + _sticky_protection_target(batch_memory_used, device)
+        )
 
     should_tile = free_memory < target_free and batch_number <= 1
     if should_tile:
         _LOG.info(
-            "GPU Resident Loader: skipping regular VAE pass and switching directly to tiled mode; free=%s target=%s batch_memory=%s",
+            "GPU Resident Loader: skipping regular VAE pass and switching directly to tiled mode; free=%s target=%s batch_memory=%s model_load=%s",
             free_memory,
             target_free,
             batch_memory_used,
+            model_load_required,
         )
     return batch_number, batch_memory_used, should_tile
 
@@ -994,6 +1062,108 @@ def _call_tiled_vae(
     return func(self, data, **kwargs)
 
 
+def _should_prefer_tiled_vae_encode(vae: Any, pixel_samples: Any) -> bool:
+    if REGISTRY.get_policy() != "sticky_gpu":
+        return False
+    if vae is None or pixel_samples is None:
+        return False
+
+    try:
+        vae.throw_exception_if_invalid()
+        prepared = vae.vae_encode_crop_pixels(pixel_samples)
+        prepared = prepared.movedim(-1, 1)
+        if int(getattr(vae, "latent_dim", 2)) == 3 and prepared.ndim < 5:
+            if not getattr(vae, "not_video", False):
+                prepared = prepared.movedim(1, 0).unsqueeze(0)
+            else:
+                prepared = prepared.unsqueeze(2)
+
+        memory_used = vae.memory_used_encode(prepared.shape, vae.vae_dtype)
+        _, _, should_tile = _prepare_sticky_vae_batch(
+            device=getattr(vae, "device", None),
+            patcher=getattr(vae, "patcher", None),
+            total_memory_used=memory_used,
+            total_batch_count=prepared.shape[0],
+        )
+        return bool(should_tile)
+    except Exception as exc:
+        _LOG.debug("GPU Resident Loader: failed to preflight sticky VAE encode preference: %s", exc)
+        return False
+
+
+def _call_bound_tiled_vae(func: Callable[..., Any], pixel_samples: Any, *args: Any, **kwargs: Any) -> Any:
+    supported_kwargs = _tiled_vae_supported_kwargs(getattr(func, "__func__", func))
+    filtered_kwargs = {key: value for key, value in kwargs.items() if key in supported_kwargs}
+    return func(pixel_samples, *args, **filtered_kwargs)
+
+
+@contextlib.contextmanager
+def _temporary_prefer_tiled_vae_encode(vae: Any):
+    original_encode = getattr(vae, "encode", None)
+    encode_tiled = getattr(vae, "encode_tiled", None)
+    if not callable(original_encode) or not callable(encode_tiled):
+        yield
+        return
+
+    had_instance_attr = "encode" in getattr(vae, "__dict__", {})
+    lock = getattr(vae, _TILED_VAE_MEMORY_LOCK_ATTR, None)
+    if lock is None:
+        with _TILED_VAE_LOCK_INIT:
+            lock = getattr(vae, _TILED_VAE_MEMORY_LOCK_ATTR, None)
+            if lock is None:
+                lock = threading.RLock()
+                setattr(vae, _TILED_VAE_MEMORY_LOCK_ATTR, lock)
+
+    def prefer_encode(pixel_samples, *args, **kwargs):
+        if _should_prefer_tiled_vae_encode(vae, pixel_samples):
+            return _call_bound_tiled_vae(encode_tiled, pixel_samples, *args, **kwargs)
+        return original_encode(pixel_samples, *args, **kwargs)
+
+    with lock:
+        setattr(vae, "encode", prefer_encode)
+        try:
+            yield
+        finally:
+            if had_instance_attr:
+                setattr(vae, "encode", original_encode)
+            else:
+                delattr(vae, "encode")
+
+
+def _wrap_vae_encode_for_inpaint_node(func: Callable[..., Any]) -> Callable[..., Any]:
+    supported_kwargs = frozenset(inspect.signature(func).parameters)
+
+    @functools.wraps(func)
+    def wrapper(self, *args, **kwargs):
+        # Filter kwargs to only include those supported by the wrapped function
+        filtered_kwargs = {key: value for key, value in kwargs.items() if key in supported_kwargs}
+
+        # Supply default grow_mask_by if not present and supported
+        if "grow_mask_by" in supported_kwargs and "grow_mask_by" not in filtered_kwargs:
+            filtered_kwargs["grow_mask_by"] = 6
+
+        if REGISTRY.get_policy() != "sticky_gpu":
+            return func(self, *args, **filtered_kwargs)
+
+        # Extract vae from args for the context manager
+        vae = args[0] if args else kwargs.get("vae")
+        with _temporary_prefer_tiled_vae_encode(vae):
+            return func(self, *args, **filtered_kwargs)
+
+    return wrapper
+
+
+def _wrap_inpaint_model_conditioning_node(func: Callable[..., Any]) -> Callable[..., Any]:
+    @functools.wraps(func)
+    def wrapper(self, positive, negative, pixels, vae, mask, noise_mask=True):
+        if REGISTRY.get_policy() != "sticky_gpu":
+            return func(self, positive, negative, pixels, vae, mask, noise_mask=noise_mask)
+        with _temporary_prefer_tiled_vae_encode(vae):
+            return func(self, positive, negative, pixels, vae, mask, noise_mask=noise_mask)
+
+    return wrapper
+
+
 def _wrap_vae_encode_tiled(func: Callable[..., Any]) -> Callable[..., Any]:
     @functools.wraps(func)
     def wrapper(self, pixel_samples, tile_x=None, tile_y=None, overlap=None, tile_t=None, overlap_t=None):
@@ -1441,6 +1611,7 @@ def install_patches() -> None:
     import comfy.model_patcher as model_patcher
     import comfy.sd as comfy_sd
     import comfy.utils as comfy_utils
+    import nodes as comfy_nodes
 
     original_load_torch_file = _remember_original("utils.load_torch_file", comfy_utils.load_torch_file)
     if comfy_utils.load_torch_file is original_load_torch_file:
@@ -1496,6 +1667,22 @@ def install_patches() -> None:
     if comfy_sd.VAE.decode_tiled is original_vae_decode_tiled:
         comfy_sd.VAE.decode_tiled = _wrap_vae_decode_tiled(original_vae_decode_tiled)
 
+    if hasattr(comfy_nodes, "VAEEncodeForInpaint") and hasattr(comfy_nodes.VAEEncodeForInpaint, "encode"):
+        original_vae_encode_for_inpaint = _remember_original(
+            "nodes.VAEEncodeForInpaint.encode",
+            comfy_nodes.VAEEncodeForInpaint.encode,
+        )
+        if comfy_nodes.VAEEncodeForInpaint.encode is original_vae_encode_for_inpaint:
+            comfy_nodes.VAEEncodeForInpaint.encode = _wrap_vae_encode_for_inpaint_node(original_vae_encode_for_inpaint)
+
+    if hasattr(comfy_nodes, "InpaintModelConditioning") and hasattr(comfy_nodes.InpaintModelConditioning, "encode"):
+        original_inpaint_model_conditioning = _remember_original(
+            "nodes.InpaintModelConditioning.encode",
+            comfy_nodes.InpaintModelConditioning.encode,
+        )
+        if comfy_nodes.InpaintModelConditioning.encode is original_inpaint_model_conditioning:
+            comfy_nodes.InpaintModelConditioning.encode = _wrap_inpaint_model_conditioning_node(original_inpaint_model_conditioning)
+
     original_clip_vision_load = _remember_original("clip_vision.load", clip_vision.load)
     if clip_vision.load is original_clip_vision_load:
         clip_vision.load = _wrap_with_load_context(