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: build each Qwen-Image-Edit variant on its own transformer config (2509 / original Edit no longer get 2511's zero_cond_t) (#12815)
Daniel Han4b00d49577ab
diff --git a/studio/backend/core/inference/diffusion.py b/studio/backend/core/inference/diffusion.pyindex 1649e3afe7f..820f0d0a573 100644--- a/studio/backend/core/inference/diffusion.py+++ b/studio/backend/core/inference/diffusion.py@@ -69,6 +69,8 @@ from .diffusion_families import (resolve_base_repo,resolve_local_gguf_child,supported_family_names,+ transformer_config_overrides_for,+ transformer_variant_differs_from_base,)from .diffusion_compat import (assert_flux2_pick_compatible,@@ -2586,6 +2588,14 @@ class DiffusionBackend:speed = kwargs.get("speed_mode")if speed is not None and str(speed).strip().lower() == SPEED_OFF:return False+ if transformer_variant_differs_from_base(+ fam,+ kwargs.get("base_repo"),+ kwargs.get("gguf_filename"),+ kwargs.get("repo_id"),+ kwargs.get("display_repo_id"),+ ):+ return Falsetry:# A definite-offload policy skips the dense build, so widening wastes a multi-GB pull with no GGUF# fallback.@@ -4364,7 +4374,13 @@ class DiffusionBackend:(lambda companions, transformer_files: self._dense_quant_prefetch_needed(fam,- {**load_kwargs, "base_repo": base, "hf_token": hf_token},+ {+ **load_kwargs,+ "repo_id": repo_id,+ "gguf_filename": gguf_filename,+ "base_repo": base,+ "hf_token": hf_token,+ },companion_files = companions,transformer_files = transformer_files,)@@ -5715,6 +5731,25 @@ class DiffusionBackend:# repo's transformer/ shards, since the fallback would pull them HERE, inside the load lock, after# eviction, where unload cannot preempt it and progress already reported 100%.dense_fallback_allowed = bool(_transformer_prefetched)+ # 2509 / original Edit GGUFs resolve to the 2511 base, whose transformer/ is another model.+ if (+ kind == "gguf"+ and normalize_transformer_quant(transformer_quant) is not None+ and transformer_variant_differs_from_base(+ fam, base, gguf_filename, repo_id, display_repo_id+ )+ ):+ dense_declined = True+ if transformer_quant_decline is None:+ transformer_quant_decline = (+ f"{base} has a different transformer than this GGUF, so only the GGUF "+ "itself can run it"+ )+ # Baking adapters needs the dense build; staying on auto lets the load fail loudly below.+ if transformer_quant_pinned is None and not _has_active_lora(loras):+ transformer_quant = "off"+ else:+ transformer_quant_decline_status = RESOLVED_UNSUPPORTED# Set when an offloading GGUF pick loads the pre-quantised checkpoint instead (diffusion_gguf_route).gguf_offload_placement: Optional[Any] = Nonegguf_offload_scheme: Optional[str] = None@@ -6663,6 +6698,10 @@ class DiffusionBackend:# it, so without the flag this branch reaches the Hub on a load nobody asked for. The# pipeline assembly below was already guarded; this call was not."local_files_only": local_files_only,+ # config is the family base's (2511); 2509 / original Edit lack its zero_cond_t+ **transformer_config_overrides_for(+ fam, gguf_filename, repo_id, display_repo_id, base+ ),}# Before the prefix shim below, which wraps whatever entry it finds and# finds nothing for a class that is not registered yet.@@ -7230,7 +7269,9 @@ class DiffusionBackend:self._raise_if_load_cancelled(_load_token)# Before from_pipe copies the scheduler.apply_comfy_flow_shift(- pipe, comfy_flow_shift_for(fam, gguf_filename, repo_id, base), logger+ pipe,+ comfy_flow_shift_for(fam, gguf_filename, repo_id, display_repo_id, base),+ logger,)# Before the speed optims, so the fused batched tile decode does not replace it.try:diff --git a/studio/backend/core/inference/diffusion_families.py b/studio/backend/core/inference/diffusion_families.pyindex eaa890ff5ff..ca494d715dd 100644--- a/studio/backend/core/inference/diffusion_families.py+++ b/studio/backend/core/inference/diffusion_families.py@@ -121,6 +121,10 @@ class DiffusionFamily:comfy_flow_shift: Optional[float] = None# (lowercased id substring, shift) for checkpoints whose template differs; first match wins.comfy_flow_shift_variants: tuple[tuple[str, float], ...] = field(default_factory = tuple)+ # (lowercased id substring, ((key, value), ...)) overriding ``base_repo``'s transformer config; first match wins.+ transformer_config_variants: tuple[tuple[str, tuple[tuple[str, Any], ...]], ...] = field(+ default_factory = tuple+ )# Activation-guard cost of one condition pixel relative to one output pixel.condition_pixel_weight: float = 1.0# Extra lowercased substrings (besides ``name``) that map a repo id here.@@ -352,8 +356,17 @@ _FAMILIES: tuple[DiffusionFamily, ...] = (name = "qwen-image-edit",filter_reduction_configs_archs = _REDUCTION_RACE_ARCHS,comfy_flow_shift = 3.1, # ComfyUI ModelSamplingAuraFlow 3.1 (Qwen-Image-Edit 2511 template)- # The 2509 template samples at ModelSamplingAuraFlow 3.- comfy_flow_shift_variants = (("qwen-image-edit-2509", 3.0),),+ comfy_flow_shift_variants = (+ ("qwen-image-edit-2511", 3.1),+ ("qwen-image-edit-2509", 3.0),+ ("qwen-image-edit", 3.0),+ ),+ # Only the 2511 config sets zero_cond_t; on 2509 / original Edit it renders oversaturated, off-identity images.+ transformer_config_variants = (+ ("qwen-image-edit-2511", ()),+ ("qwen-image-edit-2509", (("zero_cond_t", False),)),+ ("qwen-image-edit", (("zero_cond_t", False),)),+ ),pipeline_class = "QwenImageEditPlusPipeline",transformer_class = "QwenImageTransformer2DModel",base_repo = "Qwen/Qwen-Image-Edit-2511",@@ -1284,15 +1297,33 @@ _GENERATION_DEFAULTS: tuple[tuple[str, int, float], ...] = (_GENERATION_DEFAULT_FALLBACK = (9, 0.0)-def comfy_flow_shift_for(fam: Any, *identifiers: Optional[str]) -> Optional[float]:- """ComfyUI's static shift for the loaded checkpoint: the first family variant whose key is in- an identifier (repo id, GGUF file, base repo), else the family default."""+def _first_variant(rows: Any, identifiers: tuple[Optional[str], ...]) -> Optional[tuple]:+ """First ``(key, ...)`` row whose key is in an identifier; ``_`` reads as ``-`` (ComfyUI file names)."""for identifier in identifiers:- needle = (identifier or "").lower()- for key, shift in getattr(fam, "comfy_flow_shift_variants", ()) or ():- if key in needle:- return shift- return getattr(fam, "comfy_flow_shift", None)+ needle = (identifier or "").lower().replace("_", "-")+ for row in rows or ():+ if row[0] in needle:+ return row+ return None+++def comfy_flow_shift_for(fam: Any, *identifiers: Optional[str]) -> Optional[float]:+ row = _first_variant(getattr(fam, "comfy_flow_shift_variants", ()), identifiers)+ return row[1] if row else getattr(fam, "comfy_flow_shift", None)+++def transformer_config_overrides_for(fam: Any, *identifiers: Optional[str]) -> dict[str, Any]:+ row = _first_variant(getattr(fam, "transformer_config_variants", ()), identifiers)+ return dict(row[1]) if row else {}+++def transformer_variant_differs_from_base(+ fam: Any, base: Optional[str], *identifiers: Optional[str]+) -> bool:+ """Checkpoint and ``base`` name different variants; a base naming none (a local dir) is unknown."""+ rows = getattr(fam, "transformer_config_variants", ())+ base_row = _first_variant(rows, (base,))+ return base_row is not None and _first_variant(rows, identifiers) not in (None, base_row)def default_generation_params(*identifiers: Optional[str]) -> tuple[int, float]:diff --git a/studio/backend/tests/test_diffusion_backend.py b/studio/backend/tests/test_diffusion_backend.pyindex 034c180c24f..62f55d0a481 100644--- a/studio/backend/tests/test_diffusion_backend.py+++ b/studio/backend/tests/test_diffusion_backend.py@@ -2090,6 +2090,129 @@ def test_edit_family_uses_own_pipeline_and_requires_image(fake_runtime, tmp_pathbackend.generate(prompt = "make it night", steps = 8)+ "gguf_filename, expected",+ [+ ("qwen-image-edit-2509-Q6_K.gguf", {"zero_cond_t": False}),+ ("qwen_image_edit_2509_Q4_K_M.gguf", {"zero_cond_t": False}),+ ("qwen-image-edit-Q4_K_M.gguf", {"zero_cond_t": False}),+ ("qwen-image-edit-2511-Q4_K_M.gguf", {}),+ ("model.gguf", {}),+ ],+)+def test_qwen_edit_gguf_builds_on_its_variant_config(+ fake_runtime, tmp_path, gguf_filename, expected+):+ """2509 / original Edit override the 2511 companion's zero_cond_t; 2511 and unnamed files do not."""+ (tmp_path / gguf_filename).write_bytes(b"x")+ backend = DiffusionBackend()+ _load_into(+ backend,+ tmp_path,+ gguf_filename = gguf_filename,+ base_repo = "Qwen/Qwen-Image-Edit-2511",+ family_override = "qwen-image-edit",+ )+ assert _FakeTransformer.last["config"].endswith("/Qwen-Image-Edit-2511")+ assert {k: v for k, v in _FakeTransformer.last.items() if k == "zero_cond_t"} == expected+++def test_non_qwen_edit_gguf_gets_no_config_override(fake_runtime, tmp_path):+ (tmp_path / "qwen-image-2512-Q4_K_M.gguf").write_bytes(b"x")+ backend = DiffusionBackend()+ _load_into(+ backend, tmp_path, gguf_filename = "qwen-image-2512-Q4_K_M.gguf", family_override = "qwen-image"+ )+ assert "zero_cond_t" not in _FakeTransformer.last+++def test_qwen_edit_display_repo_id_names_the_variant(fake_runtime, tmp_path):+ (tmp_path / "model.gguf").write_bytes(b"x")+ _load_into(+ DiffusionBackend(),+ tmp_path,+ gguf_filename = "model.gguf",+ display_repo_id = "unsloth/Qwen-Image-Edit-2509-GGUF",+ base_repo = "Qwen/Qwen-Image-Edit-2511",+ family_override = "qwen-image-edit",+ )+ assert _FakeTransformer.last.get("zero_cond_t") is False+++def _qwen_edit_dense_route(backend, monkeypatch):+ from core.inference import diffusion as dmod++ _force_cuda_target(backend, monkeypatch)+ monkeypatch.setattr(dmod, "dense_transformer_supported", lambda target: True)+ monkeypatch.setattr(+ dmod, "select_transformer_quant_scheme", lambda target, mode, family = None, **_kw: "fp8"+ )+ attempted = []++ def fake_dense_load(self, *a, **k):+ attempted.append(k.get("base") if "base" in k else a[2])+ return None, None++ monkeypatch.setattr(DiffusionBackend, "_load_dense_quant_pipeline", fake_dense_load)+ return attempted+++ "gguf_filename, takes_dense",+ [("qwen-image-edit-2509-Q6_K.gguf", False), ("qwen-image-edit-2511-Q4_K_M.gguf", True)],+)+def test_qwen_edit_variant_gguf_never_runs_the_2511_dense_transformer(+ fake_runtime, tmp_path, monkeypatch, allow_precision_fallback, gguf_filename, takes_dense+):+ backend = DiffusionBackend()+ attempted = _qwen_edit_dense_route(backend, monkeypatch)+ (tmp_path / gguf_filename).write_bytes(b"x")+ _load_into(+ backend,+ tmp_path,+ gguf_filename = gguf_filename,+ base_repo = "Qwen/Qwen-Image-Edit-2511",+ family_override = "qwen-image-edit",+ transformer_quant = "fp8",+ )+ assert bool(attempted) is takes_dense+ assert _FakeTransformer.last["path"].endswith(gguf_filename)+++def test_qwen_edit_variant_gguf_refuses_a_pinned_quant(fake_runtime, tmp_path, monkeypatch):+ backend = DiffusionBackend()+ attempted = _qwen_edit_dense_route(backend, monkeypatch)+ (tmp_path / "qwen-image-edit-2509-Q6_K.gguf").write_bytes(b"x")+ with pytest.raises(RuntimeError, match = "different transformer"):+ _load_into(+ backend,+ tmp_path,+ gguf_filename = "qwen-image-edit-2509-Q6_K.gguf",+ base_repo = "Qwen/Qwen-Image-Edit-2511",+ family_override = "qwen-image-edit",+ transformer_quant = "fp8",+ )+ assert attempted == []+++def test_qwen_edit_variant_gguf_with_baked_loras_fails_instead_of_silent_drop(+ fake_runtime, tmp_path, monkeypatch+):+ backend = DiffusionBackend()+ attempted = _qwen_edit_dense_route(backend, monkeypatch)+ (tmp_path / "qwen-image-edit-2509-Q6_K.gguf").write_bytes(b"x")+ with pytest.raises(RuntimeError, match = "LoRA adapters could not be applied"):+ _load_into(+ backend,+ tmp_path,+ gguf_filename = "qwen-image-edit-2509-Q6_K.gguf",+ base_repo = "Qwen/Qwen-Image-Edit-2511",+ family_override = "qwen-image-edit",+ loras = [("adapter", 1.0)],+ )+ assert attempted == []++def test_load_pipeline_kind_uses_from_pretrained(fake_runtime):"""A full-pipeline (no single-file) load on an unsloth/* repo builds the pipe withpipeline_cls.from_pretrained(repo_id) -- NO single-file transformer build, NO GGUFdiff --git a/studio/backend/tests/test_diffusion_prequant.py b/studio/backend/tests/test_diffusion_prequant.pyindex b13e9508fcb..10f362605a0 100644--- a/studio/backend/tests/test_diffusion_prequant.py+++ b/studio/backend/tests/test_diffusion_prequant.py@@ -2857,7 +2857,7 @@ def test_the_download_plan_probes_with_the_user_token():from core.inference.diffusion import DiffusionBackendsrc = inspect.getsource(DiffusionBackend.download_plan)- assert '{**load_kwargs, "base_repo": base, "hf_token": hf_token}' in src+ assert '"base_repo": base,' in src and '"hf_token": hf_token,' in srcdef test_a_full_fingerprint_pass_is_remembered_per_unchanged_file(monkeypatch, tmp_path):diff --git a/studio/backend/tests/test_diffusion_qwen_edit_variants.py b/studio/backend/tests/test_diffusion_qwen_edit_variants.pynew file mode 100644index 00000000000..638ef4dcdab--- /dev/null+++ b/studio/backend/tests/test_diffusion_qwen_edit_variants.py@@ -0,0 +1,181 @@+# SPDX-License-Identifier: AGPL-3.0-only+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0++"""Qwen-Image-Edit variants: only 2511 sets zero_cond_t; 2509 and the original sample at shift 3."""++from __future__ import annotations++import pytest++from core.inference.diffusion_families import (+ comfy_flow_shift_for,+ detect_family,+ detect_family_for_pick,+ transformer_config_overrides_for,+ transformer_variant_differs_from_base,+)++_OFF = {"zero_cond_t": False}+++ "identifiers, expected",+ [+ (+ (+ "qwen-image-edit-2509-Q6_K.gguf",+ "unsloth/Qwen-Image-Edit-2509-GGUF",+ "Qwen/Qwen-Image-Edit-2511",+ ),+ _OFF,+ ),+ ((None, "unsloth/Qwen-Image-Edit-2509-GGUF", "Qwen/Qwen-Image-Edit-2511"), _OFF),+ ((None, "/models/local", "Qwen/Qwen-Image-Edit-2509"), _OFF),+ (("qwen_image_edit_2509_Q4_K_M.gguf", "/models/unet", "Qwen/Qwen-Image-Edit-2511"), _OFF),+ (+ (+ "qwen-image-edit-Q4_K_M.gguf",+ "unsloth/Qwen-Image-Edit-GGUF",+ "Qwen/Qwen-Image-Edit-2511",+ ),+ _OFF,+ ),+ ((None, "QuantStack/Qwen-Image-Edit-GGUF", "Qwen/Qwen-Image-Edit"), _OFF),+ (+ (+ "qwen-image-edit-2511-Q4_K_M.gguf",+ "unsloth/Qwen-Image-Edit-2511-GGUF",+ "Qwen/Qwen-Image-Edit-2511",+ ),+ {},+ ),+ (("qwen_image_edit_2511_bf16.gguf", "/models/unet", "Qwen/Qwen-Image-Edit-2511"), {}),+ (("model.gguf", "/models/unet", "Qwen/Qwen-Image-Edit-2511"), {}),+ ],+)+def test_qwen_edit_transformer_config_per_variant(identifiers, expected):+ fam = detect_family("qwen-image-edit")+ assert fam.base_repo == "Qwen/Qwen-Image-Edit-2511"+ assert transformer_config_overrides_for(fam, *identifiers) == expected+++def test_variant_lookup_returns_a_fresh_dict():+ fam = detect_family("qwen-image-edit")+ first = transformer_config_overrides_for(fam, "qwen-image-edit-2509-Q6_K.gguf")+ first["zero_cond_t"] = True+ assert transformer_config_overrides_for(fam, "qwen-image-edit-2509-Q6_K.gguf") == _OFF+++ "repo_id, gguf_filename",+ [+ ("unsloth/Qwen-Image-2512-GGUF", "qwen-image-2512-Q4_K_M.gguf"),+ ("unsloth/Qwen-Image-GGUF", "qwen-image-Q4_K_M.gguf"),+ ("unsloth/FLUX.1-Kontext-dev-GGUF", "flux1-kontext-dev-Q4_K_M.gguf"),+ ("unsloth/Z-Image-Turbo-GGUF", "z-image-turbo-Q4_K_M.gguf"),+ ],+)+def test_other_families_get_no_config_override(repo_id, gguf_filename):+ fam = detect_family_for_pick(repo_id, gguf_filename)+ assert fam is not None and fam.name != "qwen-image-edit"+ assert transformer_config_overrides_for(fam, gguf_filename, repo_id, fam.base_repo) == {}+++def test_qwen_edit_flow_shift_per_variant():+ # Comfy-Org image_qwen_image_edit{,_2509} templates: ModelSamplingAuraFlow 3; _2511: 3.1.+ fam = detect_family("qwen-image-edit")+ base = fam.base_repo+ assert (+ comfy_flow_shift_for(+ fam, "qwen-image-edit-Q4_K_M.gguf", "unsloth/Qwen-Image-Edit-GGUF", base+ )+ == 3.0+ )+ assert comfy_flow_shift_for(fam, None, "Qwen/Qwen-Image-Edit", None) == 3.0+ assert (+ comfy_flow_shift_for(+ fam, "qwen-image-edit-2509-Q6_K.gguf", "unsloth/Qwen-Image-Edit-2509-GGUF", base+ )+ == 3.0+ )+ assert (+ comfy_flow_shift_for(+ fam, "qwen-image-edit-2511-Q4_K_M.gguf", "unsloth/Qwen-Image-Edit-2511-GGUF", base+ )+ == 3.1+ )+ assert comfy_flow_shift_for(fam, None, "unsloth/Qwen-Image-Edit-2511", None) == 3.1+ assert comfy_flow_shift_for(fam, "model.gguf", "/models/unet", base) == 3.1+ assert comfy_flow_shift_for(fam, None, None, None) == 3.1+++ "gguf_filename, shift",+ [+ ("qwen_image_edit_2509_Q4_K_M.gguf", 3.0),+ ("Qwen_Image_Edit-Q4_0.gguf", 3.0),+ ("qwen_image_edit_2511_Q4_K_M.gguf", 3.1),+ ],+)+def test_comfy_style_names_pick_their_shift(gguf_filename, shift):+ fam = detect_family("qwen-image-edit")+ assert comfy_flow_shift_for(fam, gguf_filename, "/models/unet", fam.base_repo) == shift+++ "base, gguf_filename, differs",+ [+ ("Qwen/Qwen-Image-Edit-2511", "qwen-image-edit-2509-Q6_K.gguf", True),+ ("Qwen/Qwen-Image-Edit-2511", "qwen_image_edit_Q4_K_M.gguf", True),+ ("Qwen/Qwen-Image-Edit-2511", "qwen-image-edit-2511-Q4_K_M.gguf", False),+ ("Qwen/Qwen-Image-Edit-2511", "model.gguf", False),+ ("Qwen/Qwen-Image-Edit-2509", "qwen-image-edit-2509-Q6_K.gguf", False),+ ("/models/qwen-edit-pipeline", "qwen-image-edit-2509-Q6_K.gguf", False),+ ],+)+def test_variant_differs_from_base(base, gguf_filename, differs):+ fam = detect_family("qwen-image-edit")+ assert (+ transformer_variant_differs_from_base(fam, base, gguf_filename, "/models/unet") is differs+ )+++def test_other_families_never_differ_from_base():+ fam = detect_family("qwen-image")+ assert not transformer_variant_differs_from_base(+ fam, fam.base_repo, "qwen-image-edit-2509-Q6_K.gguf"+ )+++ "gguf_filename, stages_transformer",+ [("qwen-image-edit-2509-Q4_K_M.gguf", False), ("qwen-image-edit-2511-Q4_K_M.gguf", True)],+)+def test_download_plan_skips_the_2511_transformer_for_a_variant(+ monkeypatch, gguf_filename, stages_transformer+):+ import core.inference.diffusion as dmod++ monkeypatch.setattr(dmod, "_resolve_base_repo", lambda *a, **k: "Qwen/Qwen-Image-Edit-2511")+ monkeypatch.setattr(dmod, "_assert_base_repo_accessible", lambda *a, **k: None)+ monkeypatch.setattr(+ dmod.DiffusionBackend, "_te_prequant_plan_files", staticmethod(lambda *a, **k: {})+ )+ monkeypatch.setattr(+ dmod.DiffusionBackend, "_dense_quant_prefetch_decision", lambda self, *a, **k: True+ )+ asked = []++ def estimate(+ *a,+ include_transformer = False,+ **k,+ ):+ asked.append(include_transformer((), ("transformer/x.safetensors",)))+ return 0, []++ monkeypatch.setattr(dmod.DiffusionBackend, "_estimate_download_bytes", staticmethod(estimate))+ dmod.DiffusionBackend().download_plan(+ "unsloth/Qwen-Image-Edit-GGUF", gguf_filename = gguf_filename, transformer_quant = "fp8"+ )+ assert asked == [stages_transformer]