diff --git a/studio/backend/core/inference/diffusion.py b/studio/backend/core/inference/diffusion.py index 0ededa6614..17ae1c5687 100644 --- a/studio/backend/core/inference/diffusion.py +++ b/studio/backend/core/inference/diffusion.py @@ -68,14 +68,6 @@ from .diffusion_attention import ( ) from . import diffusion_compile_cache as compile_cache from . import diffusion_gguf_compile as gguf_compile -from .diffusion_eager_patches import ( - install_compile_safe_patches, - uninstall_patches, -) -from .diffusion_arch_patches import ( - install_arch_patches, - uninstall_arch_patches, -) from .diffusion_cache import apply_step_cache from .diffusion_precision import quantize_text_encoders from .diffusion_prequant import ( @@ -149,11 +141,6 @@ def _decode_b64_image(data: str, *, mode: str = "RGB") -> Any: blob = base64.b64decode(raw, validate = False) except (binascii.Error, ValueError) as exc: raise ValueError(f"Invalid base64 image data: {exc}") from exc - try: - img = Image.open(io.BytesIO(blob)) - img.load() - except Exception as exc: # noqa: BLE001 — surfaced as a 400 to the client - raise ValueError(f"Could not decode image: {exc}") from exc # Bound the decoded size. Every image-conditioned workflow (img2img / inpaint / upscale / # reference / edit) decodes through here, so this single guard protects init, mask, and # each reference image uniformly. PIL only WARNS in its 89-178MP "decompression bomb" soft @@ -161,9 +148,19 @@ def _decode_b64_image(data: str, *, mode: str = "RGB") -> Any: # well below that. 4096px covers txt2img's 2048 max, upscales, and normal outpaint canvases; # anything larger is rejected with a clear 400 instead of risking an OOM. max_side = 4096 - w, h = img.size - if w > max_side or h > max_side: - raise ValueError(f"Image is too large ({w}x{h}); maximum is {max_side}px per side.") + try: + img = Image.open(io.BytesIO(blob)) + # Read the declared dimensions from the header (Image.open is lazy) and reject an + # over-limit image BEFORE img.load() decompresses its pixels, so a crafted + # small-payload/huge-dimension file can't spike memory before the guard runs. + w, h = img.size + if w > max_side or h > max_side: + raise ValueError(f"Image is too large ({w}x{h}); maximum is {max_side}px per side.") + img.load() + except ValueError: + raise # the size guard's own message; don't wrap it as a decode error + except Exception as exc: # noqa: BLE001 — surfaced as a 400 to the client + raise ValueError(f"Could not decode image: {exc}") from exc return img.convert(mode) @@ -476,6 +473,23 @@ class DiffusionBackend: if kind in ("gguf", "single_file"): if not gguf_filename: raise ValueError(f"a single-file checkpoint name is required for a '{kind}' load.") + # Fail a kind/extension mismatch here (before the route evicts chat and grabs the + # GPU), instead of deep in the background from_single_file: a "gguf" load needs a + # .gguf file, and a "single_file" load must not be handed a .gguf. + is_gguf_name = gguf_filename.lower().endswith(".gguf") + if kind == "gguf" and not is_gguf_name: + raise ValueError("a 'gguf' load requires a .gguf checkpoint name.") + if kind == "single_file" and is_gguf_name: + raise ValueError("a .gguf checkpoint needs model_kind 'gguf', not 'single_file'.") + # A single-file load must name an actual checkpoint: an arbitrary repo file + # (README.md, config.json) would pass preflight, evict the chat model, and + # only fail in the background from_single_file -- the eviction this + # validation exists to prevent. + if kind == "single_file" and not gguf_filename.lower().endswith(".safetensors"): + raise ValueError( + f"'{gguf_filename}' is not a loadable single-file checkpoint " + f"(expected a .safetensors name; use a .gguf name for a GGUF load)." + ) if local_root.exists(): resolve_local_gguf_child(local_root, gguf_filename) elif path_shaped: @@ -492,6 +506,16 @@ class DiffusionBackend: ) elif path_shaped: raise FileNotFoundError(f"Local model path does not exist: {repo_id}") + elif repo_id.upper().endswith("-GGUF"): + # A remote "*-GGUF" id is a single-file GGUF repo, not a full diffusers + # pipeline: loading it as a pipeline passes the trusted-repo check, evicts + # chat, then fails in the background when from_pretrained finds no + # model_index.json. Reject the certain case here (no network round-trip) + # so the bad pick fails before the GPU handoff, as the route expects. + raise ValueError( + f"'{repo_id}' is a single-file GGUF repo; load it with model_kind 'gguf' " + f"and a .gguf filename, not as a full pipeline." + ) return fam # ── Background load + progress ───────────────────────────────────────── @@ -730,6 +754,30 @@ class DiffusionBackend: return 0 # repo not in cache yet return total + @staticmethod + def _local_dir_weight_bytes(path: Path, *, exclude_transformer: bool) -> int: + """Sum the on-disk weight files under a local diffusers directory. The HF blob + cache is empty for a local path, so this is the only size signal for auto memory + planning; without it a large local model folds to zero and the planner skips + offload and OOMs. ``exclude_transformer`` drops the ``transformer/`` subfolder + for GGUF/single-file loads (their transformer is the single file, not resident + here); a full pipeline load keeps it (the whole repo is resident).""" + total = 0 + for f in path.rglob("*"): + if f.suffix.lower() not in (".safetensors", ".bin", ".pt", ".ckpt"): + continue + try: + rel = f.relative_to(path) + except ValueError: + continue + if exclude_transformer and rel.parts and rel.parts[0] == "transformer": + continue + try: + total += f.stat().st_size + except OSError: + continue + return total + @staticmethod def _companion_cache_bytes(base: str) -> int: """Resident companion (VAE + text-encoder) size for the memory plan. @@ -741,21 +789,7 @@ class DiffusionBackend: weights to zero and auto planning can pick a resident placement that OOMs.""" local = Path(base).expanduser() if local.is_dir(): - total = 0 - for f in local.rglob("*"): - if f.suffix.lower() not in (".safetensors", ".bin", ".pt", ".ckpt"): - continue - try: - rel = f.relative_to(local) - except ValueError: - continue - if rel.parts and rel.parts[0] == "transformer": - continue # supplied by the GGUF single-file; not resident here - try: - total += f.stat().st_size - except OSError: - continue - return total + return DiffusionBackend._local_dir_weight_bytes(local, exclude_transformer = True) return DiffusionBackend._cache_bytes(base) # ── Synchronous load / generate / unload ─────────────────────────────── @@ -1006,6 +1040,18 @@ class DiffusionBackend: eager_patched = False compile_ctx = None state_committed = False + # Lazy import: these patch modules import torch at module level, so + # importing them here (not at module load) keeps diffusion.py torch-free + # to import, letting get_diffusion_backend() run on a torchless native install. + from .diffusion_eager_patches import ( + install_compile_safe_patches, + uninstall_patches, + ) + from .diffusion_arch_patches import ( + install_arch_patches, + uninstall_arch_patches, + ) + try: if effective_speed != SPEED_OFF: install_compile_safe_patches() @@ -1256,7 +1302,13 @@ class DiffusionBackend: if kind == "pipeline": # The whole repo (transformer + companions) is one cached download; the # cached bytes are the resident estimate (bnb-4bit / fp8 stay compressed). - cached = self._cache_bytes(repo_id) if repo_id else 0 + # A LOCAL pipeline path isn't in the HF blob cache, so sum its on-disk weights + # (transformer included) instead of folding to zero and skipping offload. + local_repo = Path(repo_id).expanduser() if repo_id else None + if local_repo is not None and local_repo.is_dir(): + cached = self._local_dir_weight_bytes(local_repo, exclude_transformer = False) + else: + cached = self._cache_bytes(repo_id) if repo_id else 0 cached_mib = int(cached // (1024 * 1024)) if cached else None model_dense_mib = estimate_safetensors_dense_mib(cached_mib) companion_mib = None @@ -1326,7 +1378,14 @@ class DiffusionBackend: # reuse the resident modules AT THEIR LOADED dtype, which is the whole point of # from_pipe (component reuse, no reload, no extra VRAM). pipe = getattr(diffusers, class_name).from_pipe(state.pipe, torch_dtype = None) - self._aux_pipes[class_name] = pipe + # Only publish to the shared aux cache if THIS load is still current. from_pipe runs + # under _generate_lock but NOT _lock, so an unload()/superseding load can clear + # _aux_pipes and null _state while it builds; caching unconditionally would re-insert + # a wrapper over now-stale modules that a later same-workflow load would reuse (or + # keep the old VRAM pinned). This generation still uses the returned pipe. + with self._lock: + if self._state is state: + self._aux_pipes[class_name] = pipe return pipe def _controlnet_pipe(self, state: _LoadState, resolved_cn: Any, cancel: threading.Event) -> Any: @@ -1443,11 +1502,13 @@ class DiffusionBackend: family = getattr(state.family, "name", None), model_kind = state.kind, transformer_quant = state.transformer_quant, + compiled = "compiled" in (getattr(state, "speed_optims", ()) or ()), ): raise ValueError( "LoRA is not supported for this model/quantisation on the diffusers engine " - "(GGUF-via-diffusers or torchao fp8/int8). Use a bf16 or bnb-4bit load, or the " - "native engine for GGUF models." + "(GGUF-via-diffusers, torchao fp8/int8, or a torch.compile'd Speed=default/max " + "load). Use a bf16 or bnb-4bit load at Speed=off/eager, or the native engine " + "for GGUF models." ) resolved = diffusion_lora.resolve_specs(specs, hf_token = state.hf_token, cancel_event = cancel) @@ -1640,6 +1701,15 @@ class DiffusionBackend: fit = min(1.0, max_side / max(tw_f, th_f)) tw = max(16, int(round(tw_f * fit / 16.0)) * 16) th = max(16, int(round(th_f * fit / 16.0)) * 16) + # After the absolute cap, the target must still exceed the input, or + # "upscale" would shrink it (e.g. a 3000px source at 2x clamps to 2048). + # Reject rather than silently return a smaller image than uploaded. + if max(tw, th) <= max(iw, ih): + raise ValueError( + f"Upscale would not enlarge this image: its longest side " + f"({max(iw, ih)}px) already meets the {max_side}px output limit. " + f"Use a smaller source image." + ) init_pil = init_pil.resize((tw, th), Image.LANCZOS) if strength is None: # Hires-fix default: low enough to preserve content, high enough to @@ -1909,6 +1979,10 @@ class DiffusionBackend: # bit-identical dequant. Idempotent. gguf_compile.uninstall_all() if state.eager_patched: + # Lazy import (torch at module level) to keep diffusion.py torch-free to import. + from .diffusion_eager_patches import uninstall_patches + from .diffusion_arch_patches import uninstall_arch_patches + uninstall_patches() uninstall_arch_patches() # NOTE: we deliberately do NOT call state.pipe.unload_lora_weights() here. unload() @@ -1982,6 +2056,7 @@ class DiffusionBackend: family = state.family.name, model_kind = state.kind, transformer_quant = state.transformer_quant, + compiled = "compiled" in (getattr(state, "speed_optims", ()) or ()), ), "supports_controlnet": diffusion_controlnet.supports_controlnet( engine = "diffusers", diff --git a/studio/backend/core/inference/diffusion_families.py b/studio/backend/core/inference/diffusion_families.py index db4285da60..94f6a1e7cd 100644 --- a/studio/backend/core/inference/diffusion_families.py +++ b/studio/backend/core/inference/diffusion_families.py @@ -262,14 +262,26 @@ _FAMILIES: tuple[DiffusionFamily, ...] = ( _EDIT_KEYWORDS = ("edit", "kontext", "inpaint", "layered") +def _token_in_needle(token: str, needle: str) -> bool: + """True when ``token`` appears in ``needle`` as a whole path/name segment, i.e. + delimited by a separator (``- _ . / \\``) or a string boundary, not merely as a + raw substring. This keeps multi-part tokens matching where they should + ('qwen-image-edit' in 'qwen-image-edit-2511') while preventing a short token from + matching inside an unrelated word ('kontext' must not match 'kontextual', 'edit' + must not match 'edition').""" + return re.search(r"(?:^|[-_./\\])" + re.escape(token) + r"(?:$|[-_./\\])", needle) is not None + + def _best_family_match(needle: str) -> Optional[DiffusionFamily]: - """The family whose name/alias is the LONGEST substring of ``needle``. Longest = - most specific, so an edit checkpoint ('...qwen-image-edit-2511...') matches the - 'qwen-image-edit' family rather than the generic 'qwen-image' one.""" + """The family whose name/alias is the LONGEST whole-segment token of ``needle``. + Longest = most specific, so an edit checkpoint ('...qwen-image-edit-2511...') + matches the 'qwen-image-edit' family rather than the generic 'qwen-image' one. + Segment matching (not raw substring) stops a short alias like 'kontext' from + hijacking an unrelated path such as '.../kontextual/z-image-...gguf'.""" best: Optional[tuple[DiffusionFamily, int]] = None for fam in _FAMILIES: for token in (fam.name, *fam.aliases): - if token in needle and (best is None or len(token) > best[1]): + if _token_in_needle(token, needle) and (best is None or len(token) > best[1]): best = (fam, len(token)) return best[0] if best else None @@ -295,9 +307,16 @@ def detect_family(repo_id: str, override: Optional[str] = None) -> Optional[Diff # Don't let a generic base family (e.g. qwen-image) swallow a variant it can't run # (qwen-image-LAYERED, ...-Inpaint): if the id still carries a reject keyword the # matched family does not itself declare, reject so the load fails fast + clearly. + # Scope the keyword check to the LAST path component (the model id or + # filename), not arbitrary parent directories: a valid file selected as + # repo_id `/models/edit` + filename `Z-Image-Turbo-Q4.gguf` must not be + # rejected because a parent folder happens to be named `edit`. The + # combined `repo_id/gguf_filename` fallback passes the filename last. + basename = re.split(r"[/\\]+", needle)[-1] matched_tokens = (match.name, *match.aliases) if any( - kw in needle and not any(kw in tok for tok in matched_tokens) for kw in _EDIT_KEYWORDS + _token_in_needle(kw, basename) and not any(kw in tok for tok in matched_tokens) + for kw in _EDIT_KEYWORDS ): return None return match diff --git a/studio/backend/core/inference/diffusion_lora.py b/studio/backend/core/inference/diffusion_lora.py index 20b40aaf55..7db07dcd60 100644 --- a/studio/backend/core/inference/diffusion_lora.py +++ b/studio/backend/core/inference/diffusion_lora.py @@ -25,6 +25,8 @@ from typing import Optional from utils.hf_xet_fallback import hf_hub_download_with_xet_fallback from utils.paths.storage_roots import studio_root +from .diffusion_families import DIFFUSION_CANCELLED_MSG + # LoRA file formats we accept. sd-cli probes .safetensors/.gguf/.pt; diffusers loads # .safetensors. We expose safetensors + gguf (pt is legacy/pickled -> excluded for safety). _NATIVE_EXTS = (".safetensors", ".gguf") @@ -241,7 +243,10 @@ def resolve_specs( A stale / unknown id raises FileNotFoundError inside resolve_one; convert it to ValueError so the route (which maps only ValueError to a 400) reports bad client - input instead of a generic 500.""" + input instead of a generic 500. A Hub download can also raise + ``RuntimeError("Cancelled")`` when the user unloads / starts a superseding load + mid-download; convert that to the diffusion cancellation sentinel so the route + maps it to a 409 instead of a generic server error toast.""" out: list[ResolvedLora] = [] try: for spec_id, weight in specs: @@ -250,6 +255,10 @@ def resolve_specs( out.append(resolve_one(spec_id, weight, hf_token = hf_token, cancel_event = cancel_event)) except FileNotFoundError as exc: raise ValueError(str(exc)) from exc + except RuntimeError as exc: + if str(exc) == "Cancelled": + raise RuntimeError(DIFFUSION_CANCELLED_MSG) from exc + raise return out @@ -297,14 +306,15 @@ def inject_prompt_tags(prompt: str, resolved: list[ResolvedLora]) -> str: sd-cli strips these tags before they reach the model, so appending them is safe and deterministic. A selected adapter's weight is validated (0-2) and recorded in the request/gallery, so the injected tag must WIN over any `` the user - typed for that same alias: strip a user tag whose alias matches a selected adapter, - then append the validated one. Tags for aliases the user typed that are NOT selected - are left untouched (free-form use). + typed. Strip ALL user-typed tags first: only the selected adapters are materialized in + the managed `--lora-model-dir`, so a tag for an unselected alias can never resolve + anyway (sd-cli's extract_and_remove_lora silently removes unresolved tags), and a tag + for a selected alias must not override the validated weight. Then append the validated + tags for the selected adapters. """ - selected = {r.alias for r in resolved} - # Drop any user-typed tag whose alias is one of the selected adapters, so the typed - # weight can't override the validated weight (or slip outside the 0-2 bounds). - cleaned = _TAG_RE.sub(lambda m: "" if m.group(1) in selected else m.group(0), prompt) + # Drop every user-typed tag: unselected ones are dead (not in the managed dir) and + # selected ones must not override the validated weight / 0-2 bounds. + cleaned = _TAG_RE.sub("", prompt) # Collapse whitespace left by stripped tags without disturbing the user's text. cleaned = re.sub(r"[ \t]{2,}", " ", cleaned).strip() tags = [f"" for r in resolved] @@ -343,12 +353,17 @@ def supports_lora( family: Optional[str], model_kind: Optional[str], transformer_quant: Optional[str], + compiled: bool = False, ) -> bool: """Single gate for whether the current load can apply LoRA (used by status + backends). Native (sd_cpp): GGUF via sd-cli, for the LoRA-capable families only (Qwen excluded). Diffusers: bf16 or bnb-4bit transformers, but NOT the dense torchao fp8/int8 fast path - (tensor-subclass weights) and NOT GGUF-via-diffusers. + (tensor-subclass weights) and NOT GGUF-via-diffusers. A diffusers transformer that was + torch.compile'd at load (Speed=default/max) also can't take a non-hotswap adapter: + diffusers requires the adapter to be loaded BEFORE compilation, so applying one to the + already-compiled module fails with adapter-key mismatches. ``compiled`` is diffusers-only + (the native sd-cli path has no torch compile). """ fam = (family or "").lower() if engine == "sd_cpp": @@ -358,4 +373,6 @@ def supports_lora( return False # GGUF diffusers transformer: use the native engine for LoRA if transformer_quant and transformer_quant.lower() in _DIFFUSERS_LORA_BLOCKED_QUANT: return False + if compiled: + return False # can't load an adapter onto an already-compiled transformer return True diff --git a/studio/backend/core/inference/sd_cpp_backend.py b/studio/backend/core/inference/sd_cpp_backend.py index ee8a677096..43df6aa12b 100644 --- a/studio/backend/core/inference/sd_cpp_backend.py +++ b/studio/backend/core/inference/sd_cpp_backend.py @@ -181,7 +181,7 @@ def _map_guidance( classifier-free ``--cfg-scale``. A distilled 0/1 means CFG off (sd-cli's 1.0); a value > 1 is real CFG. Mirrors the engine mapping validated in the CPU benchmark. """ - if fam.name in ("flux.1", "flux.2-klein"): + if fam.name in ("flux.1", "flux.2-klein", "flux.2-dev"): return None, (float(guidance) if guidance is not None else None) cfg = float(guidance) if (guidance is not None and guidance > 1.0) else 1.0 return cfg, None @@ -509,10 +509,18 @@ class SdCppDiffusionBackend: from core.inference import diffusion_lora - if init_image is not None or mask_image is not None or reference_images: + if ( + init_image is not None + or mask_image is not None + or reference_images + or (upscale is not None and upscale > 1) + ): + # upscale needs an input image, so a direct API call with upscale > 1 but no + # init_image must be rejected too rather than silently returning a plain, + # un-upscaled text-to-image result (the diffusers backend rejects the same). raise ValueError( - "img2img / inpaint / reference are not yet supported on the native sd.cpp " - "engine; run on a GPU (diffusers) for image-conditioned workflows." + "img2img / inpaint / reference / upscale are not yet supported on the native " + "sd.cpp engine; run on a GPU (diffusers) for image-conditioned workflows." ) if controlnet is not None: raise ValueError( @@ -536,8 +544,13 @@ class SdCppDiffusionBackend: cfg_scale, flux_guidance = _map_guidance(state.family, guidance) # Resolve any selected LoRA adapters up front (downloads land in the HF # cache; a bad id fails here as a clear 400 before we spawn sd-cli). + # Drop weight-0 rows BEFORE the support gate: LoraSpec documents weight 0 + # as disabling the adapter (the diffusers path treats it as empty), so a + # request carrying only disabled rows must stay a no-op even on a family + # where native LoRA is unsupported, rather than 400 on a dead selection. lora_resolved: list = [] - if loras: + active_loras = [(i, w) for (i, w) in (loras or []) if w != 0] + if active_loras: if not diffusion_lora.supports_lora( engine = "sd_cpp", family = state.family.name, @@ -549,7 +562,7 @@ class SdCppDiffusionBackend: "sd.cpp engine." ) lora_resolved = diffusion_lora.resolve_specs( - loras, hf_token = state.hf_token, cancel_event = cancel + active_loras, hf_token = state.hf_token, cancel_event = cancel ) extra_args: list[str] = [] if state.vae_format: @@ -695,6 +708,7 @@ class SdCppDiffusionBackend: "engine": "sd_cpp", "supports_lora": False, "supports_controlnet": False, + "workflows": [], } from core.inference import diffusion_lora @@ -728,6 +742,11 @@ class SdCppDiffusionBackend: ), # Native ControlNet (sd-cli --control-net) is a follow-up; off for now. "supports_controlnet": False, + # The native engine supports plain text-to-image only (generate() rejects + # img2img / inpaint / reference / upscale), so advertise just txt2img. Without + # this the status omits workflows, the UI reads [], and it disables the Create + # tab for a loaded native model, stranding the user on an image-only tab. + "workflows": ["txt2img"], } diff --git a/studio/backend/core/inference/sd_cpp_engine.py b/studio/backend/core/inference/sd_cpp_engine.py index 0d4d0cc47a..18a64d9ea0 100644 --- a/studio/backend/core/inference/sd_cpp_engine.py +++ b/studio/backend/core/inference/sd_cpp_engine.py @@ -75,6 +75,14 @@ def _terminate(proc: "subprocess.Popen") -> None: proc.kill() except Exception: # noqa: BLE001 -- best-effort teardown pass + # Reap the killed child so it does not linger as a zombie until the next Popen + # cleanup / interpreter exit. Callers raise immediately after _terminate (the + # cancellation and timeout paths), so without this a burst of image cancellations + # leaks process-table entries. SIGKILL is prompt, so a short bounded wait suffices. + try: + proc.wait(timeout = 5) + except Exception: # noqa: BLE001 -- best-effort reap; never block teardown + pass def _binary_name(stem: str) -> str: diff --git a/studio/backend/models/inference.py b/studio/backend/models/inference.py index a6cff30e9b..9348fa448d 100644 --- a/studio/backend/models/inference.py +++ b/studio/backend/models/inference.py @@ -1947,6 +1947,23 @@ class DiffusionGenerateRequest(BaseModel): "the loaded model or its quantisation can't apply ControlNet.", ) + @field_validator("loras") + @classmethod + def _unique_lora_ids(cls, value: Optional[list[LoraSpec]]) -> Optional[list[LoraSpec]]: + # Both apply paths break alias collisions by suffixing the adapter name/file, so a + # repeated id would load the SAME adapter as several distinct adapters and stack + # its effect past the per-adapter weight bound. The UI already prevents duplicates; + # reject them for API clients too so each adapter takes effect at most once. + if value: + seen: set[str] = set() + for spec in value: + if spec.id in seen: + raise ValueError( + f"duplicate LoRA id '{spec.id}'; list each adapter at most once" + ) + seen.add(spec.id) + return value + @field_validator("reference_images") @classmethod def _bounded_reference_items(cls, value: Optional[list[str]]) -> Optional[list[str]]: diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py index daf895072b..9176bd5cb3 100644 --- a/studio/backend/routes/inference.py +++ b/studio/backend/routes/inference.py @@ -11051,12 +11051,10 @@ async def load_diffusion_model( from core.inference.diffusion import get_diffusion_backend, resolve_model_kind from core.inference.diffusion_device import resolve_diffusion_device_target from core.inference.diffusion_engine_router import ( - active_engine_name, annotate_status, select_and_activate_engine, ) from core.inference.gpu_arbiter import acquire_for, release, DIFFUSION - from core.inference.sd_cpp_engine import ENGINE_SD_CPP from utils.native_path_leases import redact_native_paths backend = get_diffusion_backend() @@ -11084,12 +11082,15 @@ async def load_diffusion_model( engine = await asyncio.to_thread( select_and_activate_engine, fam, hf_token = request.hf_token, model_kind = kind ) - # Take the GPU from the chat backend only when this load will actually use it. - # diffusers always does; a *force-native* sd.cpp load on a CUDA/XPU/MPS box does - # too. But a native sd.cpp load on a pure-CPU host never touches the GPU, so - # acquiring would evict the resident chat model for nothing -- skip the handoff. + # Take the GPU from the chat backend only when this load will actually use it, + # which is exactly the resolved device being non-CPU. diffusers on an accelerator + # and a force-native sd.cpp load on CUDA/XPU/MPS both resolve to that device; a + # native sd.cpp load on a pure-CPU host does not. Crucially, a CPU-only host with + # no usable sd-cli falls back to diffusers ON CPU -- that also never touches GPU + # memory, so keying off the engine name (not the device) would wrongly evict a + # resident chat model for a load that cannot use the GPU. Gate on the device. device = await asyncio.to_thread(lambda: resolve_diffusion_device_target().device) - needs_gpu = active_engine_name() != ENGINE_SD_CPP or device != "cpu" + needs_gpu = device != "cpu" if needs_gpu: # Then kick the (slow) load onto a background thread and return at once -- # the client polls images/load-progress. @@ -11212,8 +11213,13 @@ async def generate_diffusion_image( { "prompt": request.prompt, "negative_prompt": request.negative_prompt, - "width": request.width, - "height": request.height, + # Persist the ACTUAL output size, not the request sliders: Transform/ + # Inpaint/Edit derive it from the uploaded image, Extend grows the + # canvas, and Upscale resizes it, so request.width/height would record + # (and later restore) the wrong dimensions for those workflows. For + # plain txt2img the image size equals the sliders anyway. + "width": getattr(image, "width", None) or request.width, + "height": getattr(image, "height", None) or request.height, "steps": request.steps, "guidance": request.guidance, "seed": seed, @@ -11257,6 +11263,8 @@ async def list_gallery_images( offset: int = 0, current_subject: str = Depends(get_current_subject), ): + from pydantic import ValidationError + from core.inference import image_gallery limit = max(1, min(limit, 200)) @@ -11264,10 +11272,18 @@ async def list_gallery_images( # Fetch one extra to learn whether more remain, without a second scan. records = await asyncio.to_thread(image_gallery.list_images, limit + 1, offset) has_more = len(records) > limit - return GalleryListResponse( - images = [GalleryImage(**r) for r in records[:limit]], - has_more = has_more, - ) + # Build the response per record and drop any that fail schema validation: a PNG + # whose recipe chunk has all required keys but a wrong value type (e.g. a + # hand-dropped or corrupted file) passes the presence-only read but would raise + # inside GalleryImage(**r). Skipping it keeps one bad file from 500-ing the whole + # gallery listing. + images = [] + for r in records[:limit]: + try: + images.append(GalleryImage(**r)) + except ValidationError: + continue + return GalleryListResponse(images = images, has_more = has_more) @studio_router.get("/images/gallery/{image_id}/file") diff --git a/studio/backend/tests/test_diffusion_backend.py b/studio/backend/tests/test_diffusion_backend.py index 214193d065..2b258df003 100644 --- a/studio/backend/tests/test_diffusion_backend.py +++ b/studio/backend/tests/test_diffusion_backend.py @@ -22,6 +22,14 @@ from core.inference.diffusion import ( _base_file_downloaded, _resolve_diffusion_compute_dtype, ) + +# diffusion.py imports the compile/arch patch modules LAZILY (they pull torch at module +# level, and diffusion.py must stay importable on a torchless native install). Import them +# here at collection time -- under the real torch -- so they are cached in sys.modules +# before the fake-torch fixtures swap it out; otherwise the lazy import inside load_pipeline +# would try to build them against the incomplete stub torch. +import core.inference.diffusion_eager_patches # noqa: E402,F401 +import core.inference.diffusion_arch_patches # noqa: E402,F401 from core.inference.diffusion_families import ( detect_family, resolve_base_repo, @@ -77,6 +85,34 @@ def test_detect_family_from_repo_id(): assert detect_family("meta-llama/Llama-3-8B") is None +def test_detect_family_matches_reject_and_alias_by_segment(): + # Reject keywords and short aliases must match whole path/name segments, not raw + # substrings, so an unrelated word that merely CONTAINS one does not misroute a + # valid base model (regression: substring matching broke these). + assert detect_family("/models/edited/z-image-turbo-Q4_K_M.gguf").name == "z-image" + assert detect_family("unsloth/Z-Image-Edition-GGUF").name == "z-image" + assert detect_family("/models/kontextual/z-image-turbo-Q4_K_M.gguf").name == "z-image" + # Supported edit families still resolve (edit / kontext are whole tokens there). + assert detect_family("unsloth/Qwen-Image-Edit-2511-GGUF").name == "qwen-image-edit" + assert detect_family("unsloth/FLUX.1-Kontext-dev-GGUF").name == "flux.1-kontext" + # Unsupported variants sharing only a base arch keyword are still rejected. + assert detect_family("unsloth/Qwen-Image-Layered-GGUF") is None + assert detect_family("unsloth/Qwen-Image-2512-Inpaint") is None + + +def test_detect_family_edit_keyword_scoped_to_basename(): + from core.inference.diffusion_families import detect_family_for_pick + + # A parent directory named `edit`/`inpaint` must NOT poison a valid pick: only + # the model id / filename basename is scanned for reject keywords. A direct + # local pick arrives as (parent_dir, filename). + assert detect_family("/models/edit") is None # the dir alone is ambiguous + assert detect_family_for_pick("/models/edit", "Z-Image-Turbo-Q4.gguf").name == "z-image" + assert detect_family_for_pick("/models/inpaint", "qwen-image-2512-Q4.gguf").name == "qwen-image" + # A genuinely unsupported variant keyword in the FILENAME still rejects. + assert detect_family_for_pick("/models/misc", "Qwen-Image-Layered-Q4.gguf") is None + + def test_detect_family_override(): assert detect_family("local/path", override = "z-image").name == "z-image" assert detect_family("local/path", override = "zimage").name == "z-image" @@ -1424,10 +1460,30 @@ def test_validate_load_request(tmp_path): backend.validate_load_request("some-org/Z-Image", gguf_filename = "model.safetensors") with pytest.raises(ValueError, match = "family"): backend.validate_load_request("meta/Llama-3", gguf_filename = "q.gguf") + # A family-looking repo paired with a non-GGUF single-file name is rejected here, + # BEFORE the route evicts chat and hands over the GPU (the background load would + # otherwise be the first to notice README.md is not a checkpoint). + with pytest.raises(ValueError, match = r"\.gguf"): + backend.validate_load_request("unsloth/Z-Image-Turbo-GGUF", gguf_filename = "README.md") assert ( backend.validate_load_request("unsloth/Z-Image-Turbo-GGUF", gguf_filename = "q.gguf").name == "z-image" ) + # A kind/extension mismatch fails fast here, before the route evicts chat + grabs the + # GPU only to fail in the background from_single_file path. + with pytest.raises(ValueError, match = ".gguf"): + backend.validate_load_request( + "unsloth/Z-Image-Turbo-GGUF", gguf_filename = "model.safetensors", model_kind = "gguf" + ) + with pytest.raises(ValueError, match = "gguf"): + backend.validate_load_request( + "unsloth/Qwen-Image-2512-FP8", gguf_filename = "q.gguf", model_kind = "single_file" + ) + # A remote "*-GGUF" repo loaded as a full pipeline (no single-file name) is a single-file + # GGUF repo, so from_pretrained would find no pipeline manifest and fail after chat is + # already evicted; reject it here before the GPU handoff. + with pytest.raises(ValueError, match = "GGUF"): + backend.validate_load_request("unsloth/Z-Image-Turbo-GGUF", model_kind = "pipeline") # A local path with a missing child fails here (before any GPU/network work). with pytest.raises(FileNotFoundError): backend.validate_load_request( diff --git a/studio/backend/tests/test_diffusion_lora.py b/studio/backend/tests/test_diffusion_lora.py index e7df3da936..f99739eff9 100644 --- a/studio/backend/tests/test_diffusion_lora.py +++ b/studio/backend/tests/test_diffusion_lora.py @@ -44,12 +44,14 @@ def test_inject_prompt_tags_validated_weight_overrides_user_typed(): assert dl.inject_prompt_tags("a cat ", [r]) == "a cat " -def test_inject_prompt_tags_keeps_unselected_user_tags(): +def test_inject_prompt_tags_strips_unselected_user_tags(): r = dl.ResolvedLora("id", "style", "/p", "safetensors", 0.8) - # A user tag for an alias that is NOT one of the selected adapters is left untouched. + # A user tag for an alias that is NOT selected is stripped: only selected adapters are + # materialized in the managed --lora-model-dir, so sd-cli would drop the dead tag anyway; + # removing it keeps the prompt clean and unambiguous. out = dl.inject_prompt_tags("a cat ", [r]) - assert "" in out - assert "" in out + assert "" not in out + assert out == "a cat " def test_inject_prompt_tags_empty_returns_prompt(): @@ -83,6 +85,45 @@ def test_supports_lora_matrix(): assert not dl.supports_lora( engine = "diffusers", family = "flux.1", model_kind = "gguf", transformer_quant = None ) + # A torch.compile'd diffusers transformer (Speed=default/max) can't take a non-hotswap + # adapter: diffusers needs the adapter loaded before compilation. + assert not dl.supports_lora( + engine = "diffusers", + family = "flux.1", + model_kind = "pipeline", + transformer_quant = None, + compiled = True, + ) + # compiled is diffusers-only; the native path ignores it. + assert dl.supports_lora( + engine = "sd_cpp", + family = "flux.1", + model_kind = "gguf", + transformer_quant = None, + compiled = True, + ) + + +def test_resolve_specs_maps_cancelled_to_diffusion_sentinel(tmp_path, monkeypatch): + # A Hub download cancelled mid-flight raises RuntimeError("Cancelled"); resolve_specs + # must convert it to the diffusion cancellation sentinel so the route maps it to 409, + # not a generic 500 server-error toast. + def _boom(spec_id, weight, **kw): + raise RuntimeError("Cancelled") + + monkeypatch.setattr(dl, "resolve_one", _boom) + with pytest.raises(RuntimeError) as ei: + dl.resolve_specs([("a", 1.0)]) + assert str(ei.value) == dl.DIFFUSION_CANCELLED_MSG + + # A non-cancellation RuntimeError is left untouched. + def _other(spec_id, weight, **kw): + raise RuntimeError("disk full") + + monkeypatch.setattr(dl, "resolve_one", _other) + with pytest.raises(RuntimeError) as ei2: + dl.resolve_specs([("a", 1.0)]) + assert str(ei2.value) == "disk full" def test_materialize_native_dir_symlinks_and_breaks_collisions(tmp_path): @@ -192,6 +233,12 @@ def test_lora_spec_and_request_validation(): LoraSpec(id = "a", weight = -0.1) # default weight assert LoraSpec(id = "a").weight == 1.0 + # duplicate ids are rejected: repeating an id would load the same adapter as several + # distinct suffixed adapters and stack its effect past the per-adapter weight bound. + with pytest.raises(Exception): + DiffusionGenerateRequest( + prompt = "x", loras = [{"id": "a", "weight": 0.5}, {"id": "a", "weight": 1.0}] + ) # ── Diffusers apply manager ───────────────────────────────────────────────── diff --git a/studio/backend/tests/test_sd_cpp_engine.py b/studio/backend/tests/test_sd_cpp_engine.py index 2c10699be1..5c9d3256af 100644 --- a/studio/backend/tests/test_sd_cpp_engine.py +++ b/studio/backend/tests/test_sd_cpp_engine.py @@ -135,6 +135,25 @@ def test_runtime_env_handles_missing_lib_path(): assert env[var] == "/opt/sdcpp/bin" +def test_terminate_reaps_killed_child(): + # Cancellation/timeout paths call _terminate then immediately raise, so it must + # reap the killed child itself -- otherwise a burst of image cancellations leaves + # zombies until a later Popen cleanup. After _terminate the returncode is set + # (the child has been waited on), so nothing lingers. + import subprocess + proc = subprocess.Popen( + [sys.executable, "-c", "import time; time.sleep(30)"], + start_new_session = (os.name == "posix"), + ) + try: + eng._terminate(proc) + assert proc.returncode is not None + finally: + if proc.poll() is None: + proc.kill() + proc.wait() + + # ── generate (fake subprocess) ────────────────────────────────────────────── diff --git a/studio/frontend/src/app/routes/__root.tsx b/studio/frontend/src/app/routes/__root.tsx index 9b959c3939..220f66d277 100644 --- a/studio/frontend/src/app/routes/__root.tsx +++ b/studio/frontend/src/app/routes/__root.tsx @@ -83,6 +83,11 @@ const CHAT_ONLY_ALLOWED = new Set([ function isChatOnlyAllowed(pathname: string): boolean { if (CHAT_ONLY_ALLOWED.has(pathname)) return true; if (pathname === "/data-recipes" || pathname.startsWith("/data-recipes/")) return true; + // Images runs on CPU/MPS via the native sd.cpp engine, which is exactly the + // no-GPU (chat-only) setup it was added for. The generic chat-only flag is about + // training/export needing a GPU, so it must not redirect /images away here or the + // native image path is unreachable on the hosts that need it. + if (pathname === "/images" || pathname.startsWith("/images/")) return true; return false; } @@ -159,6 +164,13 @@ function RootLayout() { setImagesMounted(true); } const shouldMountImages = isImagesRoute || imagesMounted; + // Chat and Images both render their own full-height shell (a fixed top rail + an + // internally-scrolling body), so both want the chat-style layout: no outer pt-14 + // inset and no outer scroll. Keying the layout off isChatRoute alone gave /images + // the non-chat pt-14 + outer overflow, pushing its picker down and clipping the + // bottom gallery. Treat them the same for the container padding/overflow only; the + // keep-alive mounts below stay keyed to each specific route. + const isChatLike = isChatRoute || isImagesRoute; useTrainingUnloadGuard(); // Global export driver: streams worker logs and tracks status from any route @@ -251,10 +263,10 @@ function RootLayout() { className="!min-h-0 h-[calc(100dvh-var(--studio-titlebar-height,0px))] overflow-hidden" > - +
{/* Stays mounted across navigation so an in-flight generation is not cancelled when leaving /chat; hidden (not unmounted) off-route. @@ -280,7 +292,7 @@ function RootLayout() {