diff --git a/studio/backend/core/inference/diffusion.py b/studio/backend/core/inference/diffusion.py index fa88706e25..bb9da0d59e 100644 --- a/studio/backend/core/inference/diffusion.py +++ b/studio/backend/core/inference/diffusion.py @@ -1216,11 +1216,30 @@ class DiffusionBackend: if cn_model is None: if cancel.is_set(): raise RuntimeError(DIFFUSION_CANCELLED_MSG) - cn_model = ( - getattr(diffusers, model_cls_name) - .from_pretrained(resolved_cn.path, torch_dtype = state.dtype, token = state.hf_token) - .to(state.device) + import torch + + # state.dtype is the display string saved at load ("bfloat16"), NOT a + # torch.dtype; pass the real dtype so diffusers loads the ControlNet at the + # base compute dtype instead of silently defaulting to float32 (extra VRAM). + cn_dtype = getattr(torch, str(state.dtype).replace("torch.", ""), None) + cn_model = getattr(diffusers, model_cls_name).from_pretrained( + resolved_cn.path, + torch_dtype = cn_dtype, + # An empty / malformed token means anonymous access; the HF client can + # raise on a blank credential instead of falling back, so coerce to None. + token = state.hf_token or None, ) + # Placement must follow the base model's offload policy. A resident base moves + # the ControlNet resident too; an offloaded (low-VRAM) base streams it through + # the device with group offloading instead of forcing the whole module onto the + # GPU, which would defeat the offload and risk an OOM. Best-effort: any failure + # falls back to the resident placement (the prior behaviour). + if getattr(state, "offload_policy", OFFLOAD_NONE) != OFFLOAD_NONE and ( + _offload_controlnet_module(cn_model, state.device, logger) + ): + pass + else: + cn_model = cn_model.to(state.device) if cancel.is_set(): # An unload raced the blocking download above and already cleared the # ControlNet caches; caching now would pin the module past the unload. @@ -1401,7 +1420,7 @@ class DiffusionBackend: pipe = state.pipe init_pil = mask_pil = None control_pil = None - cn_scale = cn_gstart = cn_gend = None + cn_scale = cn_gstart = cn_gend = cn_mode = None ref_extra: list = [] if getattr(state.family, "edit", False): # Instruction editing: the loaded pipe is the edit pipeline. It always @@ -1472,40 +1491,55 @@ class DiffusionBackend: if controlnet is not None: from core.inference import diffusion_controlnet - if workflow != "txt2img": - raise ValueError( - "ControlNet currently combines with plain text-to-image only, not the " - f"{workflow} workflow." - ) - if not diffusion_controlnet.supports_controlnet( - engine = "diffusers", - family = state.family.name, - has_controlnet_pipeline = bool( - getattr(state.family, "controlnet_pipeline_class", None) - ), - model_kind = state.kind, - transformer_quant = state.transformer_quant, - ): - raise ValueError( - "ControlNet is not supported for this model/quantisation on the " - "diffusers engine (needs a bf16 or bnb-4bit load of a family with a " - "ControlNet pipeline; not GGUF-via-diffusers or torchao fp8/int8)." - ) cn_id, cn_image_b64, cn_type, cn_strength, cn_gs, cn_ge = controlnet - resolved_cn = diffusion_controlnet.resolve_controlnet( - cn_id, - family = state.family.name, - hf_token = state.hf_token, - cancel_event = cancel, - ) - pipe = self._controlnet_pipe(state, resolved_cn, cancel) - workflow = "controlnet" - src = _decode_b64_image(cn_image_b64, mode = "RGB") - # Control map at the OUTPUT size so it aligns with the generated latents. - control_pil = diffusion_controlnet.preprocess_control(src, cn_type).resize( - (width, height), Image.LANCZOS - ) - cn_scale, cn_gstart, cn_gend = cn_strength, cn_gs, cn_ge + # strength 0 disables ControlNet (documented on the request model, and the + # frontend slider allows it): skip the whole path so a no-op selection never + # pays the multi-GB ControlNet download / VRAM cost. + if cn_strength in (None, 0, 0.0): + controlnet = None + else: + if workflow != "txt2img": + raise ValueError( + "ControlNet currently combines with plain text-to-image only, not " + f"the {workflow} workflow." + ) + if not diffusion_controlnet.supports_controlnet( + engine = "diffusers", + family = state.family.name, + has_controlnet_pipeline = bool( + getattr(state.family, "controlnet_pipeline_class", None) + ), + model_kind = state.kind, + transformer_quant = state.transformer_quant, + ): + raise ValueError( + "ControlNet is not supported for this model/quantisation on the " + "diffusers engine (needs a bf16 or bnb-4bit load of a family with a " + "ControlNet pipeline; not GGUF-via-diffusers or torchao fp8/int8)." + ) + # Decode + preprocess the control image FIRST so a malformed / unsupported + # image fails as a clean 400 BEFORE any ControlNet download or pipe build, + # rather than after paying that cost. Control map at the OUTPUT size so it + # aligns with the generated latents. + src = _decode_b64_image(cn_image_b64, mode = "RGB") + control_pil = diffusion_controlnet.preprocess_control(src, cn_type).resize( + (width, height), Image.LANCZOS + ) + try: + resolved_cn = diffusion_controlnet.resolve_controlnet( + cn_id, family = state.family.name + ) + except FileNotFoundError as exc: + # An unknown / missing ControlNet id is a bad selection -> 400, not a + # generic 500 (the route maps ValueError, not FileNotFoundError). + raise ValueError(str(exc)) from exc + pipe = self._controlnet_pipe(state, resolved_cn, cancel) + workflow = "controlnet" + cn_scale, cn_gstart, cn_gend = cn_strength, cn_gs, cn_ge + # Flux Union ControlNet selects the active mode by an integer + # ``control_mode`` (canny/depth/pose/...); map the chosen control type so + # the union model applies the right head instead of a default/wrong one. + cn_mode = diffusion_controlnet.union_control_mode(cn_id, cn_type) # Auto-resize odd-sized inputs to a multiple of 16 for the workflows whose # OUTPUT size is taken from the input image (img2img / inpaint / extend / edit), # so an upload like 186px tall no longer fails the pipeline's divisibility check. @@ -1580,6 +1614,10 @@ class DiffusionBackend: kwargs["control_guidance_start"] = cn_gstart if "control_guidance_end" in call_params and cn_gend is not None: kwargs["control_guidance_end"] = cn_gend + # Union ControlNet mode index (Flux); only when the pipe accepts it and the + # selected control type maps to a known mode. + if "control_mode" in call_params and cn_mode is not None: + kwargs["control_mode"] = cn_mode gen = _GenState(total_steps = steps) @@ -1821,6 +1859,34 @@ def _hf_base_model(repo_id: str, hf_token: Optional[str]) -> Optional[str]: return base if isinstance(base, str) and base.strip() else None +def _offload_controlnet_module(cn_model: Any, device: str, logger: Any) -> bool: + """Stream a ControlNet module through ``device`` via diffusers group offloading. + + Used when the base model was loaded with an offload policy: forcing the ControlNet + fully resident with ``.to(device)`` would defeat that low-VRAM placement and can OOM. + Group offloading is applied to this single module (it does not touch the base pipe's + existing hooks), so it is isolated and reversible. Returns True on success; on any + failure the caller falls back to a resident placement, so this never blocks a load.""" + try: + import torch + from diffusers.hooks import apply_group_offloading + + onload = torch.device(device) + apply_group_offloading( + cn_model, + onload_device = onload, + offload_device = torch.device("cpu"), + offload_type = "block_level", + num_blocks_per_group = 1, + use_stream = onload.type == "cuda", + ) + return True + except Exception as exc: # noqa: BLE001 — offload is best-effort; resident is the fallback + if logger is not None: + logger.warning("diffusion.controlnet: group offload failed (%s); loading resident", exc) + return False + + def _base_file_downloaded(rfilename: str) -> bool: """True for base-repo files ``from_pretrained`` actually fetches. diff --git a/studio/backend/core/inference/diffusion_controlnet.py b/studio/backend/core/inference/diffusion_controlnet.py index bf103a9206..3ad0468c1a 100644 --- a/studio/backend/core/inference/diffusion_controlnet.py +++ b/studio/backend/core/inference/diffusion_controlnet.py @@ -19,7 +19,6 @@ backend read an arbitrary location. from __future__ import annotations import re -import threading from dataclasses import dataclass from pathlib import Path from typing import Any, Optional @@ -142,17 +141,28 @@ def resolve_controlnet( spec_id: str, *, family: Optional[str] = None, - hf_token: Optional[str] = None, - cancel_event: Optional[threading.Event] = None, ) -> ResolvedControlNet: """Resolve a ControlNet id to a loadable repo id / local dir. Accepts a catalog/local id, or a bare public HF repo id (``owner/name``). The backend loads the result with ``ControlNetModelClass.from_pretrained(path)`` (download + cache handled there, like the base pipeline). Raises on an unknown id -> the caller maps to 400. + + ``family`` (the loaded base family) enforces catalog compatibility: a ControlNet is + architecture-specific, so a catalog entry tagged for another family is rejected here + with a clear error rather than being loaded through the wrong pipeline class later. """ entry = _catalog_by_id().get(spec_id) if entry is not None: + # A curated/local entry may declare the families it is built for. A client that + # bypasses the UI filter (direct API call) could send an entry for another family; + # reject it before any download so it never reaches the wrong ControlNet pipeline. + fam = (family or "").strip().lower() + if entry.families and fam and fam not in {f.lower() for f in entry.families}: + raise ValueError( + f"ControlNet '{spec_id}' is for {', '.join(entry.families)}, not the loaded " + f"'{family}' model; pick a ControlNet built for this family." + ) if entry.source == "local": path = entry.local_path or "" if not path or not Path(path).is_dir(): @@ -174,6 +184,33 @@ def resolve_controlnet( ) +# Union ControlNet mode indices. A single "union" model covers several control modes and +# selects the active one via an integer ``control_mode`` argument; these are the standard +# indices used by the FLUX.1 / Qwen-Image union ControlNets. "passthrough" (an already-made +# map) carries no intrinsic mode, so it maps to nothing (the caller omits control_mode). +_UNION_CONTROL_MODES: dict[str, int] = { + "canny": 0, + "tile": 1, + "depth": 2, + "blur": 3, + "pose": 4, + "gray": 5, + "lq": 6, +} + + +def union_control_mode(spec_id: str, control_type: str) -> Optional[int]: + """The integer ``control_mode`` for a union ControlNet, or None. + + Returns a mode only for a curated *union* catalog entry AND a control type that maps to a + known index; otherwise None so the caller omits the kwarg (a non-union ControlNet has a + single fixed mode, and 'passthrough' does not name one). Pure lookup, no network.""" + entry = _catalog_by_id().get(spec_id) + if entry is None or not entry.is_union: + return None + return _UNION_CONTROL_MODES.get((control_type or "").strip().lower()) + + def preprocess_control(image: Any, control_type: str) -> Any: """Turn a source image into a control map. diff --git a/studio/backend/models/inference.py b/studio/backend/models/inference.py index 580545b766..979dad1252 100644 --- a/studio/backend/models/inference.py +++ b/studio/backend/models/inference.py @@ -1856,6 +1856,14 @@ class ControlNetSpec(BaseModel): 1.0, ge = 0.0, le = 1.0, description = "Fraction of steps at which ControlNet ends" ) + @model_validator(mode = "after") + def _check_guidance_range(self) -> "ControlNetSpec": + # An inverted range (start > end) means "act over no steps"; reject it as a clean + # 422 instead of letting the diffusers pipeline raise a 500 deep in the denoise. + if self.guidance_start > self.guidance_end: + raise ValueError("guidance_start must be <= guidance_end") + return self + class DiffusionGenerateRequest(BaseModel): """Request to generate one image from the loaded diffusion model.""" diff --git a/studio/backend/tests/test_diffusion_controlnet.py b/studio/backend/tests/test_diffusion_controlnet.py index f9f3d5bd36..e91abad053 100644 --- a/studio/backend/tests/test_diffusion_controlnet.py +++ b/studio/backend/tests/test_diffusion_controlnet.py @@ -45,6 +45,26 @@ def test_resolve_controlnet_rejects_filesystem_like_ids(): dc.resolve_controlnet(bad) +def test_resolve_controlnet_enforces_family_match(): + # A curated entry tagged for another family must be rejected before download so it + # never reaches the wrong ControlNet pipeline class. + with pytest.raises(ValueError, match = "not the"): + dc.resolve_controlnet("qwen-union", family = "flux.1") + # The matching family resolves fine, and no family (unfiltered) is permissive. + assert dc.resolve_controlnet("qwen-union", family = "qwen-image").path + assert dc.resolve_controlnet("qwen-union").path + + +def test_union_control_mode_maps_only_union_entries(): + # Union entries map a known control type to its integer mode; passthrough / unknown + # types and non-union ids return None so the caller omits control_mode. + assert dc.union_control_mode("flux-union-pro", "canny") == 0 + assert dc.union_control_mode("flux-union-pro", "depth") == 2 + assert dc.union_control_mode("flux-union-pro", "pose") == 4 + assert dc.union_control_mode("flux-union-pro", "passthrough") is None + assert dc.union_control_mode("some/bare-repo", "canny") is None + + def test_resolve_controlnet_local(tmp_path, monkeypatch): d = tmp_path / "controlnets" d.mkdir() diff --git a/studio/frontend/src/features/images/images-page.tsx b/studio/frontend/src/features/images/images-page.tsx index f58c1a565e..a51f75795a 100644 --- a/studio/frontend/src/features/images/images-page.tsx +++ b/studio/frontend/src/features/images/images-page.tsx @@ -2095,7 +2095,7 @@ export function ImagesPage({ active = true }: { active?: boolean }) { hint="Condition the image on a control map (edges / depth / pose). Union models cover many types. Use 'Canny' to trace edges from your image, or 'Passthrough' if it is already a control map." >
-