Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
Studio: route krea2 GGUFs to Images instead of opening them as chat models (#14477)
Indrajit swain5ff3fa4bf601
diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.pyindex 09644a72e00..916c8e30bc6 100644--- a/studio/backend/core/inference/llama_cpp.py+++ b/studio/backend/core/inference/llama_cpp.py@@ -22052,6 +22052,8 @@ class LlamaCppBackend:"qwen_image","qwen_image21","flux",+ # Krea 2 GGUFs from some converters (molbal/krea2-gguf), #14474.+ "krea2",))# Archs shared by more than one family, so the header alone does NOT say whether the@@ -22145,11 +22147,19 @@ class LlamaCppBackend:return True@staticmethod- def _image_arch_is_pickable(gguf_path: Optional[str], model_identifier: Optional[str]) -> bool:+ def _image_arch_is_pickable(+ gguf_path: Optional[str],+ model_identifier: Optional[str],+ arch: Optional[str] = None,+ ) -> bool:"""The Images picker's fail-open buildability gate (Qwen-Image-2.1 needs diffusers 0.41)."""try:- from routes.models import _gguf_family_buildable- return _gguf_family_buildable((model_identifier, os.path.basename(gguf_path or "")))+ from routes.models import _diffusion_gguf_family_hints, _gguf_family_buildable+ return _gguf_family_buildable(+ _diffusion_gguf_family_hints(+ arch, (model_identifier, os.path.basename(gguf_path or ""))+ )+ )except Exception as e: # noqa: BLE001 -- never lose the page over a probe failurelogger.debug("Family probe failed for image arch: %s", e)return True@@ -22262,7 +22272,7 @@ class LlamaCppBackend:"from a GGUF, so the Video page does not list it either.")if arch in self._IMAGE_ARCHES:- if self._image_arch_is_pickable(gguf_path, self._model_identifier):+ if self._image_arch_is_pickable(gguf_path, self._model_identifier, arch):return (f"This is an image-generation GGUF (architecture '{arch}'), which ""cannot run as a chat model. Open it from the Images page instead."@@ -23082,7 +23092,7 @@ class LlamaCppBackend:)if (arch in LlamaCppBackend._IMAGE_ARCHES- and LlamaCppBackend._image_arch_is_pickable(gguf_path, model_identifier)+ and LlamaCppBackend._image_arch_is_pickable(gguf_path, model_identifier, arch)) or (arch in LlamaCppBackend._AMBIGUOUS_IMAGE_ARCHESand LlamaCppBackend._ambiguous_image_arch_is_pickable(gguf_path, model_identifier)diff --git a/studio/backend/hub/services/models/catalog_classification.py b/studio/backend/hub/services/models/catalog_classification.pyindex 30d3b7afbd1..60eb39022dd 100644--- a/studio/backend/hub/services/models/catalog_classification.py+++ b/studio/backend/hub/services/models/catalog_classification.py@@ -25,8 +25,20 @@ from utils.paths.path_utils import file_contents_available_locally_DIFFUSION_GGUF_ARCHS = frozenset(- {"flux", "flux2", "qwen_image", "qwenimage", "qwen_image21", "qwenimage21", "z_image", "zimage"}+ {+ "flux",+ "flux2",+ "qwen_image",+ "qwenimage",+ "qwen_image21",+ "qwenimage21",+ "z_image",+ "zimage",+ "krea2",+ })+# Archs that alone name the family. Not "flux" / "qwen_image": several DiTs ship under those (#14474).+_FAMILY_NAMING_DIFFUSION_GGUF_ARCHS = frozenset({"krea2"})_UNSUPPORTED_DIFFUSION_GGUF_ARCHS = frozenset({"sd1", "sd3", "sdxl", "aura", "hidream", "cosmos", "hyvid"})@@ -117,6 +129,16 @@ def _gguf_family_buildable(name_hints: tuple[Optional[str], ...]) -> bool:return True+def _diffusion_gguf_family_hints(+ arch: Optional[str], name_hints: tuple[Optional[str], ...]+) -> tuple[Optional[str], ...]:+ """``name_hints`` led by the architecture when it alone names the family."""+ normalized = (arch or "").strip().lower()+ if normalized in _FAMILY_NAMING_DIFFUSION_GGUF_ARCHS:+ return (normalized, *name_hints)+ return tuple(name_hints)++def _video_family_buildable(family) -> bool:try:from core.inference.diffusion_families import family_pipeline_available@@ -302,7 +324,9 @@ def _arch_to_task(arch: Optional[str], name_hints: tuple[Optional[str], ...] = (return _SPEECH_TASKif normalized in _DIFFUSION_GGUF_ARCHS:return (- "text-to-image" if _gguf_family_buildable(name_hints) else _UNSUPPORTED_DIFFUSION_TASK+ "text-to-image"+ if _gguf_family_buildable(_diffusion_gguf_family_hints(normalized, name_hints))+ else _UNSUPPORTED_DIFFUSION_TASK)if normalized in _VIDEO_GGUF_ARCHS:from core.inference.video_families import detect_video_familydiff --git a/studio/backend/routes/models.py b/studio/backend/routes/models.pyindex 9b85dd5bd48..c6a93e288c6 100644--- a/studio/backend/routes/models.py+++ b/studio/backend/routes/models.py@@ -68,6 +68,7 @@ _MAX_TASK_CLASSIFY_GGUFS = _catalog_classification._MAX_TASK_CLASSIFY_GGUFS_PLACEHOLDER_DIFFUSION_GGUF_ARCHS = _catalog_classification._PLACEHOLDER_DIFFUSION_GGUF_ARCHS_TASK_CLASSIFY_READ_SECONDS = _catalog_classification._TASK_CLASSIFY_READ_SECONDS_gguf_family_buildable = _catalog_classification._gguf_family_buildable+_diffusion_gguf_family_hints = _catalog_classification._diffusion_gguf_family_hints_is_h3_bundle_gguf_hint = _catalog_classification._is_h3_bundle_gguf_hintSPEECH_GGUF_ARCHS = _gguf_archs.SPEECH_GGUF_ARCHSis_speech_gguf_architecture = _gguf_archs.is_speech_gguf_architecturediff --git a/studio/backend/tests/test_cached_gguf_routes.py b/studio/backend/tests/test_cached_gguf_routes.pyindex e247f2292f7..8795dfb90d6 100644--- a/studio/backend/tests/test_cached_gguf_routes.py+++ b/studio/backend/tests/test_cached_gguf_routes.py@@ -2683,6 +2683,34 @@ def test_legacy_delete_delegates_to_shared_service(monkeypatch):assert calls == [("org/repo", None, "token", "/data/hf/hub")]+def test_krea2_gguf_is_an_image_model_not_chat(monkeypatch):+ """krea2-arch GGUFs (molbal/krea2-gguf) are images, not chat RUN (#14474)."""+ for hints in (+ (),+ ("custom/my-mix-Q4_K_M.gguf",),+ ("molbal/krea2-gguf", "krea2_turbo-Q4_0.gguf"),+ ):+ assert models_route._arch_to_task("krea2", hints) == "text-to-image"+ assert models_route._arch_to_task("KREA2", hints) == "text-to-image"++ # The arch alone names the family, so an unbuildable host hides even an uninformative name.+ import core.inference.diffusion_engine_router as router++ seen = []++ def buildable(fam, *, model_kind):+ seen.append((fam.name, model_kind))+ return fam.name != "krea-2"++ monkeypatch.setattr(router, "family_buildable_here", buildable)+ assert models_route._arch_to_task("krea2", ("custom/my-mix-Q4_K_M.gguf",)) == (+ models_route._UNSUPPORTED_DIFFUSION_TASK+ )+ assert seen == [("krea-2", "gguf")]+ assert models_route._arch_to_task("flux", ("custom/my-mix-Q4_K_M.gguf",)) == "text-to-image"+ assert models_route._arch_to_task("qwen_image", ("org/Qwen-Image-GGUF",)) == "text-to-image"++def test_arch_to_task_hides_unsupported_diffusion_from_chat():assert models_route._arch_to_task("flux") == "text-to-image"assert models_route._arch_to_task("z_image") == "text-to-image"diff --git a/studio/backend/tests/test_llama_cpp_non_chat_gguf_preflight.py b/studio/backend/tests/test_llama_cpp_non_chat_gguf_preflight.pyindex bed693bbce2..64eedd64e82 100644--- a/studio/backend/tests/test_llama_cpp_non_chat_gguf_preflight.py+++ b/studio/backend/tests/test_llama_cpp_non_chat_gguf_preflight.py@@ -459,3 +459,32 @@ def test_an_image_arch_this_host_cannot_build_promises_no_page(tmp_path, monkeypidentifier = "unsloth/Qwen-Image-2.1-GGUF",)assert "Open it from the Images page" in message++[email protected]("name", ["krea2_turbo-Q4_0.gguf", "my-custom-mix-Q8_0.gguf"])+def test_a_krea2_gguf_names_the_images_page(tmp_path, name):+ # Was "does not recognise this GGUF's model architecture ('krea2')" from llama-server (#14474).+ _, message = _refusal(tmp_path, arch = "krea2", name = name)+ assert message is not None+ assert "Open it from the Images page" in message+ assert "Video page" not in message+++def test_a_krea2_gguf_this_host_cannot_build_promises_no_page(tmp_path, monkeypatch):+ import core.inference.diffusion_engine_router as router++ monkeypatch.setattr(+ router, "family_buildable_here", lambda fam, *, model_kind: fam.name != "krea-2"+ )+ _, message = _refusal(tmp_path, arch = "krea2", name = "my-custom-mix-Q8_0.gguf")+ assert message is not None+ assert "cannot run as a chat model" in message+ assert "Open it from the Images page" not in message+ text = LlamaCppBackend._classify_start_failure_text(+ "error loading model: unknown model architecture: 'krea2'",+ str(tmp_path / "my-custom-mix-Q8_0.gguf"),+ None,+ )+ assert "krea2" in text+ assert "Use Unsloth's Images page to generate" not in text+ assert "does not list it either" in text