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: per-family diffusion tables move into the family registries (#13536)
Daniel Hane227a5321b5f
diff --git a/studio/backend/core/inference/diffusion.py b/studio/backend/core/inference/diffusion.pyindex be9fedbae43..ede7912097b 100644--- a/studio/backend/core/inference/diffusion.py+++ b/studio/backend/core/inference/diffusion.py@@ -809,60 +809,12 @@ def _layered_canvas(family: Any, size: tuple[int, int]) -> tuple[int, int]:return max(32, int(round(width / 32)) * 32), max(32, int(round(height / 32)) * 32)-# Official non-unsloth pipeline bases: safetensors-only, no remote code, exact lowercased match-_TRUSTED_NON_GGUF_REPOS = frozenset(- {- "stabilityai/stable-diffusion-xl-base-1.0",- "stabilityai/sdxl-turbo",- # Vendor safetensors-only bases: LoRA training bases + the BF16 artifact per group. FLUX.1 is Hub-gated;- # Qwen/Z-Image open.- "black-forest-labs/flux.1-dev",- "black-forest-labs/flux.1-schnell",- "black-forest-labs/flux.1-kontext-dev",- # Krea: guidance-distilled FLUX.1-dev finetune, same arch and gating; detected via "flux.1"- "black-forest-labs/flux.1-krea-dev",- # FLUX.2 LoRA training bases (dev gated, klein-4B open); reloaded as a pipeline by "Deploy to Create".- "black-forest-labs/flux.2-dev",- "black-forest-labs/flux.2-klein-4b",- # The other klein sizes and variants. One family covers both sizes and defaults to 4B, on the understanding- # that the real base comes from the base_model card tag. Leaving these out made that tag untrusted, so a- # klein-9B GGUF silently resolved to the 4B config and died in the GGUF quantizer with a bare shape mismatch.- "black-forest-labs/flux.2-klein-9b",- "black-forest-labs/flux.2-klein-base-4b",- "black-forest-labs/flux.2-klein-base-9b",- "tongyi-mai/z-image-turbo",- # The undistilled Z-Image: the base the upstream DreamBooth recipe trains on, and the base_model tag of- # unsloth/Z-Image-GGUF. Leaving it out made that tag untrusted, so the GGUF fell back to the Turbo companions- # and denoised on its shift 3.0 scheduler.- "tongyi-mai/z-image",- "qwen/qwen-image",- "qwen/qwen-image-2512",- "qwen/qwen-image-edit-2511",- "qwen/qwen-image-layered",- # Qwen-Image-2.1: the family's own base_repo, and a family whose base is not listed here is- # not a family at all. Detection resolves it, the version gate passes once the pinned main- # build is in, and then validate_load_request refuses it as a non-unsloth repo before the- # pipeline is ever built.- "qwen/qwen-image-2.1",- "qwen/qwen-image-2.1-turbo",- # Krea 2: assembled per-component. Turbo = inference; Raw = the LoRA training base.- "krea/krea-2-turbo",- "krea/krea-2-raw",- # Lumina Image 2.0: standard diffusers layout, generic from_pretrained path.- "alpha-vllm/lumina-image-2.0",- # HunyuanImage 2.1: open community diffusers mirror, safetensors-only.- "hunyuanvideo-community/hunyuanimage-2.1-diffusers",- # HiDream-I1: open MIT weights, all three variants one family; Llama TE from the unsloth mirror.- "hidream-ai/hidream-i1-full",- "hidream-ai/hidream-i1-dev",- "hidream-ai/hidream-i1-fast",- # Ideogram 4: no bf16 ships. -fp8 stores both DiTs as raw float8 (the family base); both nf4 repos are- # identical.- "ideogram-ai/ideogram-4-fp8",- "ideogram-ai/ideogram-4-nf4",- "ideogram-ai/ideogram-4-nf4-diffusers",- }-)+# Official non-unsloth bases allowed to load as full (non-GGUF) artifacts: safetensors-only, no remote code, exact+# lowercased match. Derived from the registry rows (base_repo, variant / training bases, FamilyProfile.trusted_repos).+from .diffusion_families import _FAMILIES as _IMAGE_REGISTRY # noqa: E402+from .family_profile import trusted_repo_ids as _trusted_repo_ids # noqa: E402++_TRUSTED_NON_GGUF_REPOS = _trusted_repo_ids(_IMAGE_REGISTRY)def _is_trusted_diffusion_repo(repo_id: str) -> bool:diff --git a/studio/backend/core/inference/diffusion_autoopt/apply.py b/studio/backend/core/inference/diffusion_autoopt/apply.pyindex 23c64d06b4e..b5e9535833e 100644--- a/studio/backend/core/inference/diffusion_autoopt/apply.py+++ b/studio/backend/core/inference/diffusion_autoopt/apply.py@@ -121,12 +121,17 @@ class FamilyHook:backends: tuple[str, ...] = ("video",)-def _is_ltx(family: Any) -> bool:- try:- from ..video_families import LTX_FAMILIES # noqa: PLC0415- except Exception: # noqa: BLE001- return False- return getattr(family, "name", None) in LTX_FAMILIES+def _profile_flag(attr: str) -> Callable[[Any], bool]:+ """Gate a hook on a ``FamilyProfile`` flag (registry data), so a new family opts in by data."""++ def applies(family: Any) -> bool:+ try:+ from ..family_profile import profile_for # noqa: PLC0415+ except Exception: # noqa: BLE001+ return False+ return bool(getattr(profile_for(family), attr, False))++ return appliesdef _ltx_static_pre(view: Any, family: Any) -> None:@@ -157,8 +162,8 @@ def _cudnn_off_post(view: Any, family: Any, applied: dict, logger: Any) -> OptioFAMILY_HOOKS: list[FamilyHook] = [- FamilyHook("ltx_static_compile", _is_ltx, pre = _ltx_static_pre),- FamilyHook("ltx_stg_compile_adapter", _is_ltx, post = _ltx_stg_post),+ FamilyHook("ltx_static_compile", _profile_flag("compile_static"), pre = _ltx_static_pre),+ FamilyHook("ltx_stg_compile_adapter", _profile_flag("stg_compile_adapter"), post = _ltx_stg_post),FamilyHook("cudnn_benchmark_off",lambda fam: not getattr(fam, "cudnn_benchmark", True),diff --git a/studio/backend/core/inference/diffusion_cache.py b/studio/backend/core/inference/diffusion_cache.pyindex e7797fa59ed..e1e83d58350 100644--- a/studio/backend/core/inference/diffusion_cache.py+++ b/studio/backend/core/inference/diffusion_cache.py@@ -50,18 +50,18 @@ _TIER_MAX = "max"# Per-model interval on each tier ("default" also covers max), measured vs the no-skip render (PR #12652): default# needs mean LPIPS <= 0.05, max <= 0.10. "steps" = the step count it was measured at: fewer steps compute every step.# Keyed by the UPSTREAM repo id so distilled / unmeasured siblings never match.-AUTO_STATIC_SKIP: dict = {- "qwen/qwen-image-2.1": {"default": 3, "max": 3, "steps": 25},- "qwen/qwen-image": {"default": 2, "max": 3, "steps": 20},- "black-forest-labs/flux.1-krea-dev": {"default": 2, "max": 3, "steps": 28},- "black-forest-labs/flux.2-klein-base-4b": {"default": 2, "max": 3, "steps": 50},- "wan-ai/wan2.2-ti2v-5b-diffusers": {"default": 2, "max": 3, "steps": 50},- # every 2 is 0.29 LPIPS on a dense-texture prompt, so max only.- "black-forest-labs/flux.1-dev": {"max": 3, "steps": 28},- "hunyuanvideo-community/hunyuanimage-2.1-diffusers": {"max": 2, "steps": 50},- "hunyuanvideo-community/hunyuanvideo-1.5-diffusers-480p_t2v": {"max": 2, "steps": 50},- "minimaxai/minimax-h3": {"max": 2, "steps": 30},-}+# The rows are ``FamilyProfile.auto_static_skip`` on the registry; the registries are torch-free, so this stays light.+def _auto_static_skip_table() -> dict:+ from .family_profile import profile_table++ return {+ repo.strip().lower(): dict(entry)+ for rows in profile_table("auto_static_skip").values()+ for repo, entry in rows+ }+++AUTO_STATIC_SKIP: dict = _auto_static_skip_table()# Load gate on the model's default steps (keeps distilled siblings out); explicit "static" keeps STATIC_MIN_STEPS (12).AUTO_STATIC_MIN_STEPS = 20diff --git a/studio/backend/core/inference/diffusion_content.py b/studio/backend/core/inference/diffusion_content.pyindex 761355b9c1e..892c848cb49 100644--- a/studio/backend/core/inference/diffusion_content.py+++ b/studio/backend/core/inference/diffusion_content.py@@ -43,19 +43,14 @@ SAME_ARCH_VARIANTS: dict[str, frozenset[str]] = {"wan2.2-t2v-a14b": frozenset({"wan2.2-t2v-a14b", "wan2.1-t2v-14b"}),}-_VIDEO_FAMILIES = frozenset(- {- "ltx-2",- "ltx-2.5",- "wan2.2-ti2v-5b",- "wan2.2-t2v-a14b",- "wan2.1-t2v-1.3b",- "wan2.1-t2v-14b",- "hunyuanvideo-1.5",- "hunyuanvideo-1.5-720p",- "minimax-h3",- }-)+def _video_family_names() -> frozenset[str]:+ from .video_families import _FAMILIES++ return frozenset(fam.name for fam in _FAMILIES)+++# Every video registry name; the registries are torch-free.+_VIDEO_FAMILIES = _video_family_names()_MAX_SAFETENSORS_HEADER = 128 * 1024 * 1024_MAX_GGUF_TENSORS = 1 << 20diff --git a/studio/backend/core/inference/diffusion_families.py b/studio/backend/core/inference/diffusion_families.pyindex 92b8c5afd81..9caa2817351 100644--- a/studio/backend/core/inference/diffusion_families.py+++ b/studio/backend/core/inference/diffusion_families.py@@ -29,6 +29,7 @@ from .family_name_match import (token_length,)from .diffusion_nvfp4_flag import nvfp4_blocked+from .family_profile import AutoPrefer, ConvRotSpec, FamilyProfile# Runtime->route contract: the /images/generate route matches these messages EXACTLY for a 409 (vs a 500), so both@@ -151,7 +152,9 @@ class DiffusionFamily:supports_torch_compile: bool = True# False keeps cudnn.benchmark off (as VideoFamily.cudnn_benchmark): its per-process conv pick changes pixels.cudnn_benchmark: bool = True- # (major, minor) archs on which the compile pins inductor's reduction-config filter; empty = inductor's pick.+ # Pin inductor's reduction-config filter on every arch (``filter_reduction_configs``) or only on the listed (major,+ # minor) archs (``filter_reduction_configs_archs``); both unset = inductor's pick. Same two fields on VideoFamily.+ filter_reduction_configs: bool = Falsefilter_reduction_configs_archs: tuple[tuple[int, int], ...] = field(default_factory = tuple)# Optional pre-quantized transformer checkpoints as (scheme, repo_id): fetched instead of the dense bf16 (lower# load VRAM + download).@@ -221,6 +224,11 @@ class DiffusionFamily:# the adapter on the matching 4-step checkpoint, so its 4B and 9B variants cannot share the single family-wide# deploy_base_repo above.deploy_base_repos: tuple[tuple[str, str], ...] = field(default_factory = tuple)+ # Display label (Train UI); None = the family name.+ label: Optional[str] = None+ # Per-family policy data (quant excludes / ConvRot / denies, TE int8 schedule, preview, step skip, trusted+ # repos); see family_profile.FamilyProfile.+ profile: FamilyProfile = field(default_factory = FamilyProfile)def deploy_base_for(self, trained_base: str) -> str:"""The inference checkpoint paired with ``trained_base``, or the input unchanged."""@@ -235,11 +243,64 @@ class DiffusionFamily:# near-tied norm-reduction configs, benchmarked per process, sum in different orders. B200 measured deterministic._REDUCTION_RACE_ARCHS: tuple[tuple[int, int], ...] = ((8, 0), (8, 9), (12, 0))+# --- Shared FamilyProfile data (see family_profile.FamilyProfile) ---++# int8 PER-FAMILY exclusions, on top of diffusion_transformer_quant._INT8_EXCLUDE_NAME_TOKENS. Qwen-Image MMDiT runs+# every TEXT-stream Linear at M = actual prompt tokens (unpadded, unlike FLUX's 512-token T5), so a short prompt falls+# under _int_mm's M floor of 16 and the denoise crashes. bf16 there costs ~nothing. Baked into the published int8+# artifacts of qwen-image / qwen-image-edit, so it moves only with a rebuild and republish.+_QWENIMAGE_INT8_EXCLUDES = (+ "txt_in",+ "add_q_proj",+ "add_k_proj",+ "add_v_proj",+ "to_add_out",+ "txt_mlp",+)+# Families whose activation ranges break specific schemes at the MODEL level (the smoke probe only proves the GEMM+# runs). Measured with the 28-pair prequant accuracy gate on B200: qwen-image + mxfp8 does semantic damage at 1024px,+# + nvfp4 is unusable (LPIPS 0.51). Neither has a known fix, so both stay denied and auto falls through to int8+# (excellent on Qwen). The deny also applies to an EXPLICIT request.+# fp8 was denied here too, for black frames. That cause is gone: it was torchao's fp8 scale chooser having no eps+# clamp, so qwen's all-zero text rows gave scale 0 and NaN, and ``_make_quant_config`` now floors it with+# ``activation_value_lb``. Re-measured on B200 with the floor, and BOTH paths the deny governs clear the 0.10 LPIPS+# bar: prequant checkpoint, full 28-pair gate -> 28/28 PASS (SSIM 0.87-0.99, LPIPS 0.027-0.228, CLIP delta <= 0.019);+# on-the-fly quantize_, gate's simple_object prompt at its own seeds -> LPIPS 0.048 / 0.033 / 0.027 / 0.063, against+# pre-floor plain fp8 at LPIPS 0.712 with SSIM 0.016 and mean luma 0. Keeping the deny cost real speed: it pinned qwen+# to int8 at 1.10x bf16 when compiled fp8 measures 1.21x.+_QWENIMAGE_SCHEME_DENY = frozenset({"mxfp8", "nvfp4"})+# Schemes denied for TRAINING on top of the inference deny. Training holds a stricter bar because the evidence above is+# rendering evidence: it says a frozen fp8 forward reconstructs the bf16 image, not that a LoRA converges when its+# frozen linears are fp8. Nobody has run that, so qwen fp8 stays out of the Train UI until someone does. Delete the+# entry once a training run is measured.+_QWENIMAGE_TRAIN_SCHEME_DENY = frozenset({"fp8"})+# Text-encoder int8 keep-bf16 schedule (skip_first, skip_last) from measured hidden-state fidelity: qwen-image+# (Qwen2.5-VL-7B) first+last 6 gives ~0.997 cosine; flux.2-dev (Mistral-Small-24B) first 3 gives ~0.98 (early-layer+# seeding). A family without one has no schedule clearing the bar, so int8 falls back to fp8.+_QWENIMAGE_TE_INT8_SKIP = (6, 6)+# Gated NVFP4 auto head (int8 first; nvfp4 only with a passing gate record for this base and device).+_GATED_AUTO_PREFER = AutoPrefer(floor = (10, 0), schemes = ("int8", "fp8", "nvfp4", "mxfp8"), gated = True)+# AUTO keeps bf16 while it fits for these (measured with both sides compiled: INT8 was at most ~5% faster and failed+# the default-on LPIPS bar); when bf16 would offload, INT8 is still picked.+_LUMINA2_BF16_WHEN_RESIDENT = "INT8 is no faster than compiled bf16 on a card that holds bf16 and changes image detail"+_HIDREAM_BF16_WHEN_RESIDENT = "INT8 is barely faster than compiled bf16 on a card that holds bf16 and changes image detail"+# Keyed by architecture, not per variant: the base repo is read from the HF base_model tag at load time, so one entry# covers Turbo/full, schnell/dev._FAMILIES: tuple[DiffusionFamily, ...] = (DiffusionFamily(name = "flux.1",+ label = "FLUX.1-dev",+ profile = FamilyProfile(+ auto_prefer = _GATED_AUTO_PREFER,+ preview_factors = "flux1",+ # Auto static step skip, measured vs the no-skip render (PR #12652); dev: every 2 is 0.29 LPIPS on a+ # dense-texture prompt, so max only.+ auto_static_skip = (+ ("black-forest-labs/flux.1-krea-dev", (("default", 2), ("max", 3), ("steps", 28))),+ ("black-forest-labs/flux.1-dev", (("max", 3), ("steps", 28))),+ ),+ ),checkpoint_variants = ("flux1-schnell", "flux1-krea-dev", "flux1-dev"),filter_reduction_configs_archs = _REDUCTION_RACE_ARCHS,cudnn_benchmark = False,@@ -293,6 +354,13 @@ _FAMILIES: tuple[DiffusionFamily, ...] = (# flux match.DiffusionFamily(name = "flux.2-klein",+ label = "FLUX.2 Klein",+ profile = FamilyProfile(+ preview_factors = "flux2",+ auto_static_skip = (+ ("black-forest-labs/flux.2-klein-base-4b", (("default", 2), ("max", 3), ("steps", 50))),+ ),+ ),pipeline_class = "Flux2KleinPipeline",uses_negative_prompt = False,transformer_class = "Flux2Transformer2DModel",@@ -346,6 +414,8 @@ _FAMILIES: tuple[DiffusionFamily, ...] = (# unsloth/FLUX.2-VAE.DiffusionFamily(name = "flux.2-dev",+ label = "FLUX.2-dev",+ profile = FamilyProfile(te_int8_skip = (3, 0), preview_factors = "flux2"),pipeline_class = "Flux2Pipeline",uses_negative_prompt = False,transformer_class = "Flux2Transformer2DModel",@@ -373,6 +443,7 @@ _FAMILIES: tuple[DiffusionFamily, ...] = (# FLUX instruction editing: FluxKontextPipeline takes an image + instruction. Specific aliases first so# detect_family prefers this over "flux.1".name = "flux.1-kontext",+ profile = FamilyProfile(preview_factors = "flux1"),filter_reduction_configs_archs = _REDUCTION_RACE_ARCHS,comfy_flow_shift = flux_mu_shift(1.15@@ -394,6 +465,14 @@ _FAMILIES: tuple[DiffusionFamily, ...] = (# Qwen instruction editing: the 2511 checkpoint ships as QwenImageEditPlusPipeline. Specific aliases first so# detect_family prefers this over "qwen-image".name = "qwen-image-edit",+ # Same DiT class + unpadded text stream as qwen-image.+ profile = FamilyProfile(+ int8_exclude_tokens = _QWENIMAGE_INT8_EXCLUDES,+ scheme_deny = _QWENIMAGE_SCHEME_DENY,+ train_scheme_deny = _QWENIMAGE_TRAIN_SCHEME_DENY,+ te_int8_skip = _QWENIMAGE_TE_INT8_SKIP,+ preview_factors = "qwen_image",+ ),filter_reduction_configs_archs = _REDUCTION_RACE_ARCHS,comfy_flow_shift = 3.1, # ComfyUI ModelSamplingAuraFlow 3.1 (Qwen-Image-Edit 2511 template)comfy_flow_shift_variants = (@@ -439,6 +518,12 @@ _FAMILIES: tuple[DiffusionFamily, ...] = (# qwen-image's DiT plus use_additional_t_cond / use_layer3d_rope (read from the base config for a GGUF) and an# RGBA VAE. Listed before qwen-image so the name outranks it.name = "qwen-image-layered",+ # Same DiT class + unpadded text stream as qwen-image.+ profile = FamilyProfile(+ int8_exclude_tokens = _QWENIMAGE_INT8_EXCLUDES,+ scheme_deny = _QWENIMAGE_SCHEME_DENY,+ te_int8_skip = _QWENIMAGE_TE_INT8_SKIP,+ ),filter_reduction_configs_archs = _REDUCTION_RACE_ARCHS,# ComfyUI's Image to Layers template: ModelSamplingAuraFlow 1, 2 layers, the input scaled to 640.comfy_flow_shift = 1.0,@@ -471,6 +556,16 @@ _FAMILIES: tuple[DiffusionFamily, ...] = (),DiffusionFamily(name = "qwen-image",+ label = "Qwen-Image",+ profile = FamilyProfile(+ int8_exclude_tokens = _QWENIMAGE_INT8_EXCLUDES,+ scheme_deny = _QWENIMAGE_SCHEME_DENY,+ train_scheme_deny = _QWENIMAGE_TRAIN_SCHEME_DENY,+ auto_prefer = _GATED_AUTO_PREFER,+ te_int8_skip = _QWENIMAGE_TE_INT8_SKIP,+ preview_factors = "qwen_image",+ auto_static_skip = (("qwen/qwen-image", (("default", 2), ("max", 3), ("steps", 20))),),+ ),filter_reduction_configs_archs = _REDUCTION_RACE_ARCHS,comfy_flow_shift = 3.1, # ComfyUI ModelSamplingAuraFlow 3.1 (Qwen-Image templates)pipeline_class = "QwenImagePipeline",@@ -525,6 +620,34 @@ _FAMILIES: tuple[DiffusionFamily, ...] = (# qwen-image entry would hand it that family's pipeline, transformer, VAE and exclusion# rules, none of which fit.name = "qwen-image-2.1",+ profile = FamilyProfile(+ # 2.1 is a 32-block SINGLE-stream DiT: no add_* projections, no txt_mlp, so the 20B MMDiT's exclusion list+ # does not apply and this one was measured rather than inherited. ``txt_in`` is here as a QUALITY lever,+ # not the small-M crash guard it is on qwen-image: all four policy arms rendered a deliberately 2-token+ # prompt without tripping ``_int_mm``'s M floor, so nothing crashes either way. Excluding it scored LPIPS+ # 0.0642 against 0.1032 with the generic tokens alone, paired over 16 calibration prompts, -0.0390 CI95+ # [-0.0708, -0.0072], better on 14/16.+ int8_exclude_tokens = ("txt_in",),+ int8_convrot = ConvRotSpec(+ group = 256,+ suffixes = (+ "attn.to_q",+ "attn.to_k",+ "attn.to_v",+ "attn.to_out.0",+ "img_mlp.gate_layer",+ "img_mlp.proj",+ "img_mlp.out",+ ),+ filename = "Qwen-Image-2.1-INT8-ConvRot.safetensors",+ repo = "unsloth/Qwen-Image-2.1-FP8",+ variant_repos = ("unsloth/Qwen-Image-2.1-Turbo-FP8",),+ default_on = True,+ ),+ auto_prefer = _GATED_AUTO_PREFER,+ preview_factors = "qwen_image_21",+ auto_static_skip = (("qwen/qwen-image-2.1", (("default", 3), ("max", 3), ("steps", 25))),),+ ),# ComfyUI QwenImage21: ModelSamplingFlux fixed mu 0.69, no terminal stretch.comfy_flow_shift = flux_mu_shift(0.69),pipeline_class = "QwenImage21Pipeline",@@ -631,6 +754,28 @@ _FAMILIES: tuple[DiffusionFamily, ...] = (),DiffusionFamily(name = "z-image",+ label = "Z-Image",+ profile = FamilyProfile(+ # exactly the Linears the hosted INT8 checkpoint quantizes (all input axes 256-divisible)+ int8_convrot = ConvRotSpec(+ group = 256,+ suffixes = (+ "attention.to_q",+ "attention.to_k",+ "attention.to_v",+ "attention.to_out.0",+ "feed_forward.w1",+ "feed_forward.w2",+ "feed_forward.w3",+ "cap_embedder.1",+ ),+ filename = "Z-Image-Turbo-INT8-ConvRot.safetensors",+ repo = "unsloth/Z-Image-Turbo-FP8",+ default_on = True,+ ),+ auto_prefer = _GATED_AUTO_PREFER,+ preview_factors = "zimage",+ ),filter_reduction_configs_archs = _REDUCTION_RACE_ARCHS,cudnn_benchmark = False,comfy_flow_shift = 3.0, # ComfyUI shift 3 for Turbo and base (Turbo already ships 3.0)@@ -677,6 +822,7 @@ _FAMILIES: tuple[DiffusionFamily, ...] = (),DiffusionFamily(name = "krea-2",+ label = "Krea 2",pipeline_class = "Krea2Pipeline",transformer_class = "Krea2Transformer2DModel",base_repo = "krea/Krea-2-Turbo",@@ -701,6 +847,7 @@ _FAMILIES: tuple[DiffusionFamily, ...] = (# from_pretrained path loads it. NOT aliased to bare "lumina": Lumina-Next checkpoints are a different arch.DiffusionFamily(name = "lumina-2",+ profile = FamilyProfile(auto_bf16_when_resident = _LUMINA2_BF16_WHEN_RESIDENT),pipeline_class = "Lumina2Pipeline",transformer_class = "Lumina2Transformer2DModel",base_repo = "Alpha-VLLM/Lumina-Image-2.0",@@ -721,6 +868,11 @@ _FAMILIES: tuple[DiffusionFamily, ...] = (# covers Full/Dev/Fast.DiffusionFamily(name = "hunyuanimage-2.1",+ profile = FamilyProfile(+ auto_static_skip = (+ ("hunyuanvideo-community/hunyuanimage-2.1-diffusers", (("max", 2), ("steps", 50))),+ ),+ ),# hosted checkpoints, verified bit-identical to on-the-fly quantize (the guider pipeline is not run-to-run# deterministic, so same-seed LPIPS mixes trajectory divergence with harness noise)prequant_repos = (@@ -744,6 +896,10 @@ _FAMILIES: tuple[DiffusionFamily, ...] = (),DiffusionFamily(name = "hidream-i1",+ profile = FamilyProfile(+ auto_bf16_when_resident = _HIDREAM_BF16_WHEN_RESIDENT,+ trusted_repos = ("hidream-ai/hidream-i1-dev", "hidream-ai/hidream-i1-fast"),+ ),# Hosted checkpoints: 28/28 per-case gate pairs per scheme (LPIPS 0.291 int8 / 0.278 fp8); int8 bit-identical# to on-the-fly quantize.prequant_repos = (@@ -767,6 +923,9 @@ _FAMILIES: tuple[DiffusionFamily, ...] = (# encoderDiffusionFamily(name = "ideogram-4",+ profile = FamilyProfile(+ trusted_repos = ("ideogram-ai/ideogram-4-nf4", "ideogram-ai/ideogram-4-nf4-diffusers"),+ ),pipeline_class = "Ideogram4Pipeline",uses_negative_prompt = False,transformer_class = "Ideogram4Transformer2DModel",@@ -779,6 +938,8 @@ _FAMILIES: tuple[DiffusionFamily, ...] = (# pipeline. img2img / inpaint / ControlNet are the standard SDXL pipelines. No GGUF path.DiffusionFamily(name = "sdxl",+ label = "SDXL",+ profile = FamilyProfile(preview_factors = "sdxl"),# ~1.8% slower warm, but compiled-UNet renders then match across servers (ComfyUI's default)cudnn_benchmark = False,pipeline_class = "StableDiffusionXLPipeline",diff --git a/studio/backend/core/inference/diffusion_precision.py b/studio/backend/core/inference/diffusion_precision.pyindex 157616e40bf..7bf92b2333e 100644--- a/studio/backend/core/inference/diffusion_precision.py+++ b/studio/backend/core/inference/diffusion_precision.py@@ -40,6 +40,7 @@ from functools import lru_cachefrom core._torchao_stub import is_stubbed, torch_is_rocmfrom .diffusion_nvfp4_flag import nvfp4_blocked, nvfp4_disabled_message+from .family_profile import profile_table as _profile_tableTE_QUANT_FP8 = "fp8"TE_QUANT_NVFP4 = "nvfp4"@@ -53,15 +54,9 @@ _TE_TORCHAO_MODES = frozenset({TE_QUANT_INT8, TE_QUANT_FP8_DYNAMIC, TE_QUANT_NVF_TEXT_ENCODER_ATTRS = ("text_encoder", "text_encoder_2", "text_encoder_3")# int8 degrades on large text encoders unless the quant-sensitive decoder blocks stay bf16. Per-family (skip_first,-# skip_last) blocks to keep dense, from measured hidden-state fidelity; absent families have no schedule clearing the-# bar, so int8 falls back to fp8. qwen-image (Qwen2.5-VL-7B): first+last 6 gives ~0.997 cosine; flux.2-dev-# (Mistral-Small-24B): first 3 gives ~0.98 (early-layer seeding).-_TE_INT8_SKIP: dict[str, tuple[int, int]] = {- "qwen-image": (6, 6),- "qwen-image-edit": (6, 6),- "qwen-image-layered": (6, 6),- "flux.2-dev": (3, 0),-}+# skip_last) blocks to keep dense (``FamilyProfile.te_int8_skip``, measured hidden-state fidelity); absent families+# have no schedule clearing the bar, so int8 falls back to fp8.+_TE_INT8_SKIP: dict[str, tuple[int, int]] = _profile_table("te_int8_skip")def normalize_te_quant(value: Optional[str]) -> Optional[str]:diff --git a/studio/backend/core/inference/diffusion_preview_factors.py b/studio/backend/core/inference/diffusion_preview_factors.pyindex 9f79a0ef454..72ddc502e03 100644--- a/studio/backend/core/inference/diffusion_preview_factors.py+++ b/studio/backend/core/inference/diffusion_preview_factors.py@@ -6,24 +6,10 @@Each map was fitted by least squares from the final denoised latent of real renders to the same renders' VAEdecode, box-averaged onto the latent grid. Families without an entry get no preview."""-# family name or alias (lower case) -> map key-FAMILY_FACTORS = {- "flux.1": "flux1",- "flux.1-kontext": "flux1",- "flux.2-klein": "flux2",- "flux.2-dev": "flux2",- "qwen-image": "qwen_image",- "qwen-image-edit": "qwen_image",- "qwen-image-2.1": "qwen_image_21",- "z-image": "zimage",- "sdxl": "sdxl",- "wan2.2-ti2v-5b": "wan22_5b",- "hunyuanvideo-1.5": "hv15",- "hunyuanvideo-1.5-720p": "hv15",- "ltx-2": "ltx2",- # 2.5 keeps the 2.3 conv VAE encoder (byte-identical weights), so the same latent space.- "ltx-2.5": "ltx2",-}+# family name (lower case) -> map key: ``FamilyProfile.preview_factors`` on the registry rows (not generated).+from .family_profile import profile_table as _profile_table # noqa: E402++FAMILY_FACTORS = _profile_table("preview_factors")FACTORS: dict = {"flux1": {diff --git a/studio/backend/core/inference/diffusion_transformer_quant.py b/studio/backend/core/inference/diffusion_transformer_quant.pyindex b6791430dca..65771830a39 100644--- a/studio/backend/core/inference/diffusion_transformer_quant.py+++ b/studio/backend/core/inference/diffusion_transformer_quant.py@@ -26,7 +26,6 @@ import os as _osimport re as _reimport sys as _sysimport threading as _threading-from dataclasses import dataclassfrom typing import Any, Callable, Optional# stdlib-only module (no torch), so this stays inside the "imported lazily" promise above.@@ -39,6 +38,8 @@ from .diffusion_nvfp4_flag import (without_nvfp4,)from .diffusion_torchao_patches import install_torchao_int_mm_patch+from .family_profile import AutoPrefer as _AutoPrefer+from .family_profile import profile_table as _profile_table# Also runs in the spawned smoke-probe child, which imports this module and nothing else of the backend.install_torchao_int_mm_patch()@@ -81,92 +82,24 @@ _INT8_EXCLUDE_NAME_TOKENS = ()-# int8 PER-FAMILY exclusions, on top of _INT8_EXCLUDE_NAME_TOKENS. Qwen-Image MMDiT runs every TEXT-stream Linear at M-# = actual prompt tokens (unpadded, unlike FLUX's 512-token T5), so a short prompt falls under _int_mm's M floor of 16-# and the denoise crashes. bf16 there costs ~nothing.-_QWENIMAGE_INT8_EXCLUDES = (- "txt_in",- "add_q_proj",- "add_k_proj",- "add_v_proj",- "to_add_out",- "txt_mlp",+# int8 PER-FAMILY exclusions, on top of _INT8_EXCLUDE_NAME_TOKENS: ``FamilyProfile.int8_exclude_tokens`` on the+# registry rows (small-M text / audio streams, measured quality levers; the reasons sit next to the data). Baked into+# published int8 artifacts and validated at load, so tests/test_diffusion_family_data_freeze.py pins every value.+_INT8_FAMILY_EXCLUDE_NAME_TOKENS: dict[str, tuple[str, ...]] = _profile_table(+ "int8_exclude_tokens", quant = True)-# HunyuanVideo-1.5's attention trim shrinks the text / image streams from their padded lengths to the VALID token-# counts, so every text-stream Linear runs at a tiny M the int8 dynamic path cannot handle. Both failures measured on-# B200. M = 0 (t2v byt5 / image streams trim to zero tokens): torchao returns the input UNPROJECTED (a quantized 1472-# -> 2048 Linear maps [1, 0, 1472] to [1, 0, 1472]), so the 2048-wide cond-type add crashes -- context_embedder_2 /-# image_embedder. M <= 16 (a short prompt, or the empty negative prompt's ~6 tokens): torch._int_mm requires M > 16-# and raises -- the TokenRefiner and every block's context-stream projections. These run at M = text tokens (tens)-# against the video stream's M ~ 32k+, so bf16 here costs nothing measurable and the video-stream linears keep full-# int8 coverage. "context_embedder" also matches "context_embedder_2" (substring check).-_HUNYUAN15_INT8_EXCLUDES = (- "context_embedder",- "image_embedder",- "add_q_proj",- "add_k_proj",- "add_v_proj",- "to_add_out",- "ff_context",-)-# MiniMax-H3 is the same small-M story with one extra trap. Its adaLN projection is named ``adaln_proj``, which no-# token in _INT8_EXCLUDE_NAME_TOKENS matches: "norm" is the closest and it does not appear in the name. On the DENSE-# checkpoint that projection is Linear(2688 -> 96768), so it clears min_features = 512 and gets quantized, then runs-# at M = 1 and raises ``self.size(0) needs to be greater than 16, but got 1`` at the first denoise. Measured: the-# offline builder happily bakes it, and torch.compile then dies on that module. The pruned-modulation form hides the-# bug rather than fixing it: there adaln_proj is Linear(8 -> 96768), which falls under min_features and is skipped for-# free (verified on the fl2va_pruned build: the filter rejects all 51 of them for min_features, in_features = 8). So-# the exclusion is what makes the DENSE path correct, and is a no-op on the pruned one. H3's OTHER small-M linears ---# context_embedder and the two token_refiner blocks, 13 in total -- used to be excluded here for the same reason; they-# are padded instead now, see _INT8_FAMILY_PAD_NAME_TOKENS below.-_MINIMAX_H3_INT8_EXCLUDES = ("adaln_proj",)-# LTX-2 is audiovisual: every block carries an audio-side stream (audio_attn1 / audio_attn2 / audio_ff, fed by-# audio_proj_in) plus the a2v / v2a cross attentions. A video-only run feeds a MINIMAL audio stream -- one token for a-# still -- so those linears run at M = 1, under the same _int_mm floor of 16 as the cases above. The audio stream is-# inert on a video-only run (isolated modalities, out of the loss, out of the LoRA targets), so leaving it in bf16-# costs nothing while the video-stream linears keep full int8 coverage. One token covers every audio-side name,-# including video_to_audio_attn, whose keys/values are that same one-token stream.-# The audio token alone is not enough. LTX-2 also carries LTX2AdaLayerNormSingle projections whose names say nothing-# about audio -- av_cross_attn_video_scale_shift / av_cross_attn_video_a2v_gate (the cross-modality modulation,-# computed unconditionally in forward, isolate_modalities or not) and prompt_adaln / audio_prompt_adaln. Each is a-# Linear over a BATCH-sized input, so M = batch = 1 for a default training run: the same _int_mm floor, reached-# whatever the audio stream does.-_LTX2_INT8_EXCLUDES = ("audio", "av_cross_attn", "adaln")-_INT8_FAMILY_EXCLUDE_NAME_TOKENS: dict[str, tuple[str, ...]] = {- "qwen-image": _QWENIMAGE_INT8_EXCLUDES,- "qwen-image-edit": _QWENIMAGE_INT8_EXCLUDES, # same DiT class + unpadded text stream- "qwen-image-layered": _QWENIMAGE_INT8_EXCLUDES, # same DiT class + unpadded text stream- # 2.1 is a 32-block SINGLE-stream DiT: no add_* projections, no txt_mlp, so the 20B MMDiT's- # exclusion list does not apply and this one was measured rather than inherited. ``txt_in`` is- # here as a QUALITY lever, not the small-M crash guard it is on qwen-image: all four policy arms- # rendered a deliberately 2-token prompt without tripping ``_int_mm``'s M floor, so nothing- # crashes either way. Excluding it scored LPIPS 0.0642 against 0.1032 with the generic tokens- # alone, paired over 16 calibration prompts, -0.0390 CI95 [-0.0708, -0.0072], better on 14/16.- "qwen-image-2.1": ("txt_in",),- "hunyuanvideo-1.5": _HUNYUAN15_INT8_EXCLUDES,- "hunyuanvideo-1.5-720p": _HUNYUAN15_INT8_EXCLUDES,- "minimax-h3": _MINIMAX_H3_INT8_EXCLUDES,- "ltx-2": _LTX2_INT8_EXCLUDES,- "ltx-2.3": _LTX2_INT8_EXCLUDES, # same audiovisual DiT- "ltx-2.5": _LTX2_INT8_EXCLUDES, # the 2.3 DiT less the video FFN bias-}+_QWENIMAGE_INT8_EXCLUDES = _INT8_FAMILY_EXCLUDE_NAME_TOKENS["qwen-image"]-# Int8 PER-FAMILY row PADDING, the alternative to excluding a small-M Linear. ``_int_mm``'s floor is a GEMM shape-# constraint, not a numerical one: padding the activation up to 32 rows, running the matmul and slicing the result-# back returns the caller's rows BITWISE unchanged (torchao's int8 dynamic path scales activations per row, so the pad-# rows cannot reach them -- see diffusion_quant_pad for why that is asserted rather than assumed). So a Linear that-# would have been left dense can be quantized after all, and its weights halve. On MiniMax-H3 that is context_embedder-# plus the two token_refiner blocks: 13 Linears, 798 M parameters, 0.80 GB of bf16 weights, 3.9% of the whole int8-# checkpoint. Measured on B200, 640x384 x 124 frames (see h3_quant/README_int8.md for the paired numbers). Scoped to-# minimax-h3: qwen-image, qwen-image-edit and hunyuanvideo-1.5 have the same small-M shape and could adopt this, but-# each has a PUBLISHED int8 prequant checkpoint whose metadata bakes the current exclusion set, and-# _validate_checkpoint compares that set against exclude_tokens_for_scheme, so changing it invalidates those artifacts-# and they move only together with a rebuild and republish of each.-_MINIMAX_H3_INT8_PAD_TOKENS = ("context_embedder", "token_refiner")-_INT8_FAMILY_PAD_NAME_TOKENS: dict[str, tuple[str, ...]] = {- "minimax-h3": _MINIMAX_H3_INT8_PAD_TOKENS,-}+# Int8 PER-FAMILY row PADDING (``FamilyProfile.int8_pad_tokens``), the alternative to excluding a small-M Linear.+# ``_int_mm``'s floor is a GEMM shape constraint, not a numerical one: padding the activation up to 32 rows, running+# the matmul and slicing the result back returns the caller's rows BITWISE unchanged (torchao's int8 dynamic path+# scales activations per row, so the pad rows cannot reach them -- see diffusion_quant_pad for why that is asserted+# rather than assumed). A family with a PUBLISHED int8 checkpoint moves from exclude to pad only with a rebuild and+# republish, since _validate_checkpoint compares the baked exclusion set against exclude_tokens_for_scheme.+_INT8_FAMILY_PAD_NAME_TOKENS: dict[str, tuple[str, ...]] = _profile_table(+ "int8_pad_tokens", quant = True+)def pad_tokens_for_scheme(scheme: str, family: Optional[str] = None) -> tuple[str, ...]:@@ -209,59 +142,31 @@ def apply_small_m_padding(return wrapped-# Per-family int8 ConvRot (group, rotated fqn suffixes); not an exclusion, so plain artifacts still validate.+# Per-family int8 ConvRot (``FamilyProfile.int8_convrot``): (group, rotated fqn suffixes); not an exclusion, so plain+# artifacts still validate.+_INT8_FAMILY_CONVROT_SPECS = _profile_table("int8_convrot", quant = True)+# A spec with group 0 (ltx-2.5) only names a HOSTED rotated build: no runtime rotation, so no entry here._INT8_FAMILY_CONVROT: dict[str, tuple[int, tuple[str, ...]]] = {- "qwen-image-2.1": (- 256,- (- "attn.to_q",- "attn.to_k",- "attn.to_v",- "attn.to_out.0",- "img_mlp.gate_layer",- "img_mlp.proj",- "img_mlp.out",- ),- ),- # exactly the Linears the hosted INT8 checkpoint quantizes (all input axes 256-divisible)- "z-image": (- 256,- (- "attention.to_q",- "attention.to_k",- "attention.to_v",- "attention.to_out.0",- "feed_forward.w1",- "feed_forward.w2",- "feed_forward.w3",- "cap_embedder.1",- ),- ),+ name: (spec.group, spec.suffixes)+ for name, spec in _INT8_FAMILY_CONVROT_SPECS.items()+ if spec.group}--_INT8_FAMILY_CONVROT_FILENAME: dict[str, str] = {- "qwen-image-2.1": "Qwen-Image-2.1-INT8-ConvRot.safetensors",- "z-image": "Z-Image-Turbo-INT8-ConvRot.safetensors",+ name: spec.filename for name, spec in _INT8_FAMILY_CONVROT_SPECS.items() if spec.filename}# Only this repo gets the rotated name prepended: a same-named file elsewhere is not this build._INT8_FAMILY_CONVROT_REPO: dict[str, str] = {- "qwen-image-2.1": "unsloth/Qwen-Image-2.1-FP8",- "z-image": "unsloth/Z-Image-Turbo-FP8",+ name: spec.repo for name, spec in _INT8_FAMILY_CONVROT_SPECS.items() if spec.repo}-# Variant repos that also host a rotated build (``<Model>-INT8-ConvRot.safetensors``), listed for the same reason._INT8_FAMILY_CONVROT_VARIANT_REPOS: dict[str, tuple[str, ...]] = {- "qwen-image-2.1": ("unsloth/Qwen-Image-2.1-Turbo-FP8",),- # Rotation baked into the file (every quantized Linear, group 256) and read back from its metadata. LTX-2.5 has no- # runtime ConvRot spec above, so quantizing at load stays plain int8.- "ltx-2.5": ("unsloth/LTX-2.5-FP8",),+ name: spec.variant_repos+ for name, spec in _INT8_FAMILY_CONVROT_SPECS.items()+ if spec.variant_repos}--# Families whose int8 runs ConvRot unless the env turns it off; the rest stay opt-in (``=1``). For ltx-2.5 this only-# picks the hosted rotated file (the env's ``0`` picks the plain one).+# Families whose int8 runs ConvRot unless the env turns it off; the rest stay opt-in (``=1``)._INT8_FAMILY_CONVROT_DEFAULT_ON: frozenset[str] = frozenset(- {"qwen-image-2.1", "z-image", "ltx-2.5"}+ name for name, spec in _INT8_FAMILY_CONVROT_SPECS.items() if spec.default_on)INT8_CONVROT_ENV = "UNSLOTH_DIFFUSION_INT8_CONVROT"@@ -372,12 +277,11 @@ def apply_runtime_convrot(return rotated-# nvfp4 raises on an empty activation; HunyuanVideo-1.5's attention trim hits it every t2v render.-_HUNYUAN15_NVFP4_ZERO_ROW_TOKENS = ("image_embedder", "context_embedder_2")-_NVFP4_FAMILY_ZERO_ROW_NAME_TOKENS: dict[str, tuple[str, ...]] = {- "hunyuanvideo-1.5": _HUNYUAN15_NVFP4_ZERO_ROW_TOKENS,- "hunyuanvideo-1.5-720p": _HUNYUAN15_NVFP4_ZERO_ROW_TOKENS,-}+# nvfp4 raises on an empty activation (``FamilyProfile.nvfp4_zero_row_tokens``); HunyuanVideo-1.5's attention trim+# hits it every t2v render.+_NVFP4_FAMILY_ZERO_ROW_NAME_TOKENS: dict[str, tuple[str, ...]] = _profile_table(+ "nvfp4_zero_row_tokens", quant = True+)def zero_row_tokens_for_scheme(scheme: str, family: Optional[str] = None) -> tuple[str, ...]:@@ -445,79 +349,32 @@ _AUTO_LADDER: tuple[tuple[tuple[int, int], tuple[str, ...]], ...] = (((8, 0), (TQ_INT8,)), # Ampere sm_80 / sm_86)-# Families whose activation ranges break specific schemes at the MODEL level (the smoke probe only proves the GEMM-# runs). Measured with the 28-pair prequant accuracy gate on B200: qwen-image + mxfp8 does semantic damage at 1024px,-# + nvfp4 is unusable (LPIPS 0.51). Neither has a known fix, so both stay denied and auto falls through to int8-# (excellent on Qwen). The deny also applies to an EXPLICIT request.-# fp8 was denied here too, for black frames. That cause is gone: it was torchao's fp8 scale chooser having no eps-# clamp, so qwen's all-zero text rows gave scale 0 and NaN, and ``_make_quant_config`` now floors it with-# ``activation_value_lb``. Re-measured on B200 with the floor, and BOTH paths the deny governs clear the 0.10 LPIPS-# bar: prequant checkpoint, full 28-pair gate -> 28/28 PASS (SSIM 0.87-0.99, LPIPS 0.027-0.228, CLIP delta <= 0.019);-# on-the-fly quantize_, gate's simple_object prompt at its own seeds -> LPIPS 0.048 / 0.033 / 0.027 / 0.063, against-# pre-floor plain fp8 at LPIPS 0.712 with SSIM 0.016 and mean luma 0. Keeping the deny cost real speed: it pinned qwen-# to int8 at 1.10x bf16 when compiled fp8 measures 1.21x.-_FAMILY_SCHEME_DENY: dict[str, frozenset[str]] = {- "qwen-image": frozenset({TQ_MXFP8, TQ_NVFP4}),- "qwen-image-edit": frozenset({TQ_MXFP8, TQ_NVFP4}), # same DiT- "qwen-image-layered": frozenset({TQ_MXFP8, TQ_NVFP4}), # same DiT-}+# Families whose activation ranges break specific schemes at the MODEL level (``FamilyProfile.scheme_deny``; the smoke+# probe only proves the GEMM runs). The deny also applies to an EXPLICIT request; a passing NVFP4 gate record lifts an+# nvfp4 deny per base. Measurements next to the data (qwen-image).+_FAMILY_SCHEME_DENY: dict[str, frozenset[str]] = _profile_table("scheme_deny", quant = True)-@dataclass(frozen = True)-class _AutoPrefer:- """A family's head of the ``auto`` order, ahead of ``_AUTO_LADDER``; off below ``floor`` or on consumer GPUs- (unless ``consumer_ok``), and ``gated`` / ``backend`` need a passing record / backend for THIS device."""-- floor: tuple[int, int]- schemes: tuple[str, ...]- consumer_ok: bool = False- gated: bool = False- backend: Optional[str] = None---_FAMILY_AUTO_PREFER: dict[str, _AutoPrefer] = {- "z-image": _AutoPrefer(- floor = (10, 0), schemes = (TQ_INT8, TQ_FP8, TQ_NVFP4, TQ_MXFP8), gated = True- ),- "flux.1": _AutoPrefer(floor = (10, 0), schemes = (TQ_INT8, TQ_FP8, TQ_NVFP4, TQ_MXFP8), gated = True),- "qwen-image": _AutoPrefer(- floor = (10, 0), schemes = (TQ_INT8, TQ_FP8, TQ_NVFP4, TQ_MXFP8), gated = True- ),- "qwen-image-2.1": _AutoPrefer(- floor = (10, 0), schemes = (TQ_INT8, TQ_FP8, TQ_NVFP4, TQ_MXFP8), gated = True- ),- "wan2.2-t2v-a14b": _AutoPrefer(- floor = (10, 0),- schemes = (TQ_NVFP4, TQ_INT8, TQ_FP8, TQ_MXFP8),- backend = "flashinfer",- ),-}+# ``FamilyProfile.auto_prefer``: a family's head of the ``auto`` order, ahead of ``_AUTO_LADDER``.+_FAMILY_AUTO_PREFER: dict[str, _AutoPrefer] = _profile_table("auto_prefer", quant = True)-# AUTO keeps bf16 while it fits for these (measured with both sides compiled: INT8 was at most ~5% faster and failed the-# default-on LPIPS bar); when bf16 would offload, INT8 is still picked.-_FAMILY_AUTO_BF16_WHEN_RESIDENT: dict[str, str] = {- "lumina-2": "INT8 is no faster than compiled bf16 on a card that holds bf16 and changes image detail",- "hidream-i1": "INT8 is barely faster than compiled bf16 on a card that holds bf16 and changes image detail",-}+# AUTO keeps bf16 while it fits for these (``FamilyProfile.auto_bf16_when_resident``: measured with both sides compiled,+# INT8 was at most ~5% faster and failed the default-on LPIPS bar); when bf16 would offload, INT8 is still picked.+_FAMILY_AUTO_BF16_WHEN_RESIDENT: dict[str, str] = _profile_table(+ "auto_bf16_when_resident", quant = True+)def auto_bf16_when_resident_reason(family: Optional[str]) -> Optional[str]:return _FAMILY_AUTO_BF16_WHEN_RESIDENT.get(str(family or "").strip().lower())-# Schemes denied for TRAINING on top of the inference table. Training holds a stricter bar because the evidence above-# is rendering evidence: it says a frozen fp8 forward reconstructs the bf16 image, not that a LoRA converges when its-# frozen linears are fp8. Nobody has run that, so qwen fp8 stays out of the Train UI until someone does. Delete the-# entry once a training run is measured.-_FAMILY_TRAIN_SCHEME_DENY: dict[str, frozenset[str]] = {- "qwen-image": frozenset({TQ_FP8}),- "qwen-image-edit": frozenset({TQ_FP8}),- # MiniMax-H3's packed sequence runs three modalities through one set of linears, so its activation range is not- # the per-family range the fp8 filter was measured against. The trainer refuses both outright; declaring it here- # is what keeps /diffusion/info from advertising a start that is guaranteed to 400.- "minimax-h3": frozenset({TQ_FP8, TQ_MXFP8}),-}+# Schemes denied for TRAINING on top of the inference table (``FamilyProfile.train_scheme_deny``): the inference+# evidence is rendering evidence, not evidence that a LoRA converges on those frozen linears.+_FAMILY_TRAIN_SCHEME_DENY: dict[str, frozenset[str]] = _profile_table(+ "train_scheme_deny", quant = True+)def _nvfp4_gate_passed(family, base_repo) -> bool:diff --git a/studio/backend/core/inference/family_profile.py b/studio/backend/core/inference/family_profile.pynew file mode 100644index 00000000000..3cf1aacbcc7--- /dev/null+++ b/studio/backend/core/inference/family_profile.py@@ -0,0 +1,160 @@+# SPDX-License-Identifier: AGPL-3.0-only+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0++"""Per-family policy data carried on ``DiffusionFamily.profile`` / ``VideoFamily.profile``.++Stdlib only (the transformer-quant smoke-probe child imports its consumers). Everything a new family needs beyond its+registry row lives here as data; the consumer modules (transformer quant, text-encoder quant, untiled VAE decode,+preview, step skip, trusted repos) derive their lookup tables from the registries through ``profile_table`` and keep+their historical names and lookup semantics.++The quant fields are baked into PUBLISHED prequant artifacts and validated at load (``diffusion_prequant.+_validate_checkpoint`` requires an artifact's exclude tokens to be a superset of ``exclude_tokens_for_scheme``), so a+change to ``int8_exclude_tokens`` / ``int8_convrot`` / ``nvfp4_zero_row_tokens`` on a family with a hosted checkpoint+needs a rebuild and republish; tests/test_diffusion_family_data_freeze.py pins every resolved value.++Scheme names are spelled as plain strings ("int8", "fp8", "mxfp8", "nvfp4") to keep this module import-free; they+equal ``diffusion_transformer_quant.TQ_*``.+"""++from __future__ import annotations++from dataclasses import dataclass, field+from typing import Any, Optional+++@dataclass(frozen = True)+class AutoPrefer:+ """A family's head of the transformer-quant ``auto`` order, ahead of the arch ladder; off below ``floor`` or on+ consumer GPUs (unless ``consumer_ok``), and ``gated`` / ``backend`` need a passing NVFP4 gate record / backend for+ THIS device."""++ floor: tuple[int, int]+ schemes: tuple[str, ...]+ consumer_ok: bool = False+ gated: bool = False+ backend: Optional[str] = None+++@dataclass(frozen = True)+class ConvRotSpec:+ """int8 ConvRot: ``group`` Hadamard size and the rotated Linear fqn ``suffixes``. Not an exclusion, so plain int8+ artifacts still validate. ``filename`` / ``repo`` name the hosted rotated build (only ``repo`` gets the rotated+ name prepended), ``variant_repos`` the variant repos that host one under their own name, ``default_on`` whether+ int8 rotates unless ``UNSLOTH_DIFFUSION_INT8_CONVROT`` turns it off (the rest are opt-in with ``=1``). ``group`` 0+ means no runtime rotation: the entry only names a hosted build whose rotation is baked into the file."""++ group: int+ suffixes: tuple[str, ...]+ filename: Optional[str] = None+ repo: Optional[str] = None+ variant_repos: tuple[str, ...] = ()+ default_on: bool = False+++@dataclass(frozen = True)+class UntiledVae:+ """Untiled video VAE decode peak model (video_vae_untiled). Either per latent pixel with an RGB output term+ (``bytes_per_latent_pixel`` + ``output_scale`` = (spatial, temporal)), for a causal decoder that decodes one latent+ frame at a time, or per output frame x latent pixel (``bytes_per_frame_latent_pixel`` + ``temporal_scale``) for a+ decoder holding every frame. Bytes are fp16-decoder bytes; the caller scales by element size."""++ bytes_per_latent_pixel: Optional[float] = None+ output_scale: Optional[tuple[int, int]] = None+ bytes_per_frame_latent_pixel: Optional[float] = None+ temporal_scale: Optional[int] = None+++@dataclass(frozen = True)+class FamilyProfile:+ # int8 name tokens left dense on top of the generic M = 1 tokens (``_int_mm`` needs M > 16, or a measured quality+ # lever). Baked into published int8 artifacts.+ int8_exclude_tokens: tuple[str, ...] = ()+ # int8 name tokens whose small-M Linears get activation rows PADDED to the ``_int_mm`` floor instead of excluded.+ int8_pad_tokens: tuple[str, ...] = ()+ int8_convrot: Optional[ConvRotSpec] = None+ # nvfp4 Linears reachable with an EMPTY activation (zero-row guard after quantization).+ nvfp4_zero_row_tokens: tuple[str, ...] = ()+ # Transformer-quant schemes ruled out at the model level (explicit requests too); a passing NVFP4 gate record+ # lifts an nvfp4 deny per base.+ scheme_deny: frozenset[str] = frozenset()+ # Extra schemes denied for TRAINING on top of ``scheme_deny`` (nvfp4 is denied for training everywhere).+ train_scheme_deny: frozenset[str] = frozenset()+ auto_prefer: Optional[AutoPrefer] = None+ # Reason auto keeps bf16 while bf16 fits (INT8 still picked once bf16 would offload); None = not this family.+ auto_bf16_when_resident: Optional[str] = None+ # Text-encoder int8 keep-bf16 schedule (skip_first, skip_last) decoder blocks; None = int8 falls back to fp8.+ te_int8_skip: Optional[tuple[int, int]] = None+ untiled_vae: Optional[UntiledVae] = None+ # diffusion_preview_factors.FACTORS key for the live latent preview; None = no preview.+ preview_factors: Optional[str] = None+ # Measured auto static step skip, keyed by the lowercased UPSTREAM repo id so unmeasured siblings never match:+ # (repo, ((tier or "steps", value), ...)). See diffusion_cache.AUTO_STATIC_SKIP.+ auto_static_skip: tuple[tuple[str, tuple[tuple[str, int], ...]], ...] = ()+ # Lowercased vendor repos trusted for a full (non-GGUF) load on top of the ones the registry row already names+ # (base_repo, named_variant_bases, train_base_repos, prequant_variant_repos bases). Exact match, safetensors-only,+ # no remote code.+ trusted_repos: tuple[str, ...] = ()+ # Extra exact family strings the transformer-quant tables (exclude / pad / ConvRot / zero-row / denies / auto+ # preferences) answer to, besides ``name``: a builder spelling a family by a version name ("ltx-2.3"). Not general+ # alias resolution, which would hand a family's artifact policy to every alias substring.+ quant_names: tuple[str, ...] = ()+ # Compile the denoiser with static shapes (token counts depend only on (w, h, frames): one recompile per shape,+ # ~1.3x faster); read by the diffusion_autoopt ``ltx_static_compile`` family hook.+ compile_static: bool = False+ # Install the LTX-2 STG compile adapter after a successful compile (``ltx_stg_compile_adapter`` family hook).+ stg_compile_adapter: bool = False+++DEFAULT_PROFILE = FamilyProfile()+++def _families() -> tuple[Any, ...]:+ from .diffusion_families import _FAMILIES as image+ from .video_families import _FAMILIES as video++ return (*image, *video)+++def profile_table(attr: str, *, quant: bool = False) -> dict[str, Any]:+ """``{family name: profile.<attr>}`` over both registries, for families that set it (not None / empty).+ ``quant`` also keys each value under the profile's ``quant_names``."""+ table: dict[str, Any] = {}+ for fam in _families():+ profile = getattr(fam, "profile", DEFAULT_PROFILE)+ value = getattr(profile, attr)+ if value is None or value == () or value == frozenset():+ continue+ table[fam.name] = value+ if quant:+ for name in profile.quant_names:+ table.setdefault(name, value)+ return table+++def profile_for(family: Any) -> FamilyProfile:+ """``family.profile``, else the registry row's profile for its exact ``name`` (a stand-in family object), else the+ empty profile."""+ profile = getattr(family, "profile", None)+ if isinstance(profile, FamilyProfile):+ return profile+ name = getattr(family, "name", None)+ if name:+ for fam in _families():+ if fam.name == name:+ return fam.profile+ return DEFAULT_PROFILE+++def trusted_repo_ids(families: Any) -> frozenset[str]:+ """Lowercased non-``unsloth/`` repo ids ``families`` name as their own bases, plus each profile's+ ``trusted_repos``. ``unsloth/`` ids are trusted by prefix elsewhere, so they are left out here."""+ ids: set[str] = set()+ for fam in families:+ named = [fam.base_repo]+ named += list(getattr(fam, "named_variant_bases", ()) or ())+ named += list(getattr(fam, "train_base_repos", ()) or ())+ named += [row[0] for row in getattr(fam, "prequant_variant_repos", ()) or ()]+ named += list(getattr(fam, "profile", DEFAULT_PROFILE).trusted_repos)+ ids.update(r.strip().lower() for r in named if r)+ return frozenset(r for r in ids if not r.startswith("unsloth/"))diff --git a/studio/backend/core/inference/video.py b/studio/backend/core/inference/video.pyindex 9133221b3e8..01bb603b871 100644--- a/studio/backend/core/inference/video.py+++ b/studio/backend/core/inference/video.py@@ -270,27 +270,12 @@ install_torchao_int_mm_patch()# pipeline (full diffusers repo)_MODEL_KINDS = frozenset({"gguf", "single_file", "pipeline"})-# Vendor base repos allowed to load as full (non-GGUF) artifacts. Exact-match, lowercased, safetensors-only, no remote-# code.-_TRUSTED_NON_GGUF_VIDEO_REPOS = frozenset(- {- "lightricks/ltx-2",- "lightricks/ltx-2.3",- "lightricks/ltx-2.3-fp8",- # LTX-2.5: the Diffusers pack (safetensors + configs, no remote code) and the split single-file repo.- "lightricks/ltx-2.5-diffusers",- "lightricks/ltx-2.5",- # Wan2.2 official diffusers base repos: safetensors-only, no remote code.- "wan-ai/wan2.2-ti2v-5b-diffusers",- "wan-ai/wan2.2-t2v-a14b-diffusers",- "wan-ai/wan2.1-t2v-1.3b-diffusers",- "wan-ai/wan2.1-t2v-14b-diffusers",- # HunyuanVideo-1.5 community Diffusers repacks (tencent's own repo has no model_index.json)- "hunyuanvideo-community/hunyuanvideo-1.5-diffusers-480p_t2v",- "hunyuanvideo-community/hunyuanvideo-1.5-diffusers-720p_t2v",- "minimaxai/minimax-h3",- }-)+# Official non-unsloth bases allowed to load as full (non-GGUF) artifacts: safetensors-only, no remote code, exact+# lowercased match. Derived from the registry rows (base_repo, variant / training bases, FamilyProfile.trusted_repos).+from .video_families import _FAMILIES as _VIDEO_REGISTRY # noqa: E402+from .family_profile import trusted_repo_ids as _trusted_repo_ids # noqa: E402++_TRUSTED_NON_GGUF_VIDEO_REPOS = _trusted_repo_ids(_VIDEO_REGISTRY)def resolve_video_model_kind(gguf_filename: Optional[str], model_kind: Optional[str]) -> str:diff --git a/studio/backend/core/inference/video_families.py b/studio/backend/core/inference/video_families.pyindex 3b7c42e76cb..b9e9f9cad77 100644--- a/studio/backend/core/inference/video_families.py+++ b/studio/backend/core/inference/video_families.py@@ -24,6 +24,7 @@ from typing import Optionalfrom .diffusion_nvfp4_flag import nvfp4_blocked, without_nvfp4from .family_name_match import normalize_family_name, token_in_name, token_length+from .family_profile import AutoPrefer, ConvRotSpec, FamilyProfile, UntiledVae# The request model's ceiling on num_frames, declared HERE so the shape gate and the bound cannot drift: the gate's# refusal names the lattice point above the request, and suggesting one the request model would itself reject is a@@ -283,6 +284,8 @@ class VideoFamily:# One config per reduction instead of a per-process benchmark that can change the render (diffusion_speed.# pin_reduction_configs). Off by default: it can keep a slower config (HunyuanVideo-1.5 ~2% per step).filter_reduction_configs: bool = False+ # Same spelling as DiffusionFamily: pin the filter only on these (major, minor) archs.+ filter_reduction_configs_archs: tuple[tuple[int, int], ...] = field(default_factory = tuple)# Curated GGUF repo for the picker (the DiT as single-file GGUF quants).gguf_repo: Optional[str] = None# Hosted PRE-CAST text-encoder checkpoints as (scheme, component, repo_id); same semantics as@@ -343,6 +346,12 @@ class VideoFamily:supports_references: bool = False# Guidance-distilled families expose neither CFG nor a negative prompt.supports_cfg: bool = True+ # True when Unsloth can TRAIN a LoRA on this family (every such name must resolve through get_trainer).+ trainable: bool = False+ # Display label (Train UI); None = the family name.+ label: Optional[str] = None+ # Per-family policy data; see family_profile.FamilyProfile.+ profile: FamilyProfile = field(default_factory = FamilyProfile)# MiniMax-H3 at ComfyUI's template default (ResolutionSelector 16:9, 0.4 MP, multiple 32 -> 864x480). Both sides are@@ -375,9 +384,77 @@ _HV15_I2V_REFUSAL = ()+# --- Shared FamilyProfile data (see family_profile.FamilyProfile) ---++# int8 small-M exclusions (on top of diffusion_transformer_quant._INT8_EXCLUDE_NAME_TOKENS); baked into the published+# int8 artifacts, so each moves only with a rebuild and republish.+# HunyuanVideo-1.5's attention trim shrinks the text / image streams from their padded lengths to the VALID token+# counts, so every text-stream Linear runs at a tiny M the int8 dynamic path cannot handle. Both failures measured on+# B200. M = 0 (t2v byt5 / image streams trim to zero tokens): torchao returns the input UNPROJECTED (a quantized 1472+# -> 2048 Linear maps [1, 0, 1472] to [1, 0, 1472]), so the 2048-wide cond-type add crashes -- context_embedder_2 /+# image_embedder. M <= 16 (a short prompt, or the empty negative prompt's ~6 tokens): torch._int_mm requires M > 16+# and raises -- the TokenRefiner and every block's context-stream projections. These run at M = text tokens (tens)+# against the video stream's M ~ 32k+, so bf16 here costs nothing measurable and the video-stream linears keep full+# int8 coverage. "context_embedder" also matches "context_embedder_2" (substring check).+_HUNYUAN15_INT8_EXCLUDES = (+ "context_embedder",+ "image_embedder",+ "add_q_proj",+ "add_k_proj",+ "add_v_proj",+ "to_add_out",+ "ff_context",+)+# nvfp4 raises on an empty activation; HunyuanVideo-1.5's attention trim hits it every t2v render.+_HUNYUAN15_NVFP4_ZERO_ROW_TOKENS = ("image_embedder", "context_embedder_2")+# LTX-2 is audiovisual: every block carries an audio-side stream (audio_attn1 / audio_attn2 / audio_ff, fed by+# audio_proj_in) plus the a2v / v2a cross attentions. A video-only run feeds a MINIMAL audio stream -- one token for a+# still -- so those linears run at M = 1, under the same _int_mm floor of 16. The audio stream is inert on a+# video-only run (isolated modalities, out of the loss, out of the LoRA targets), so leaving it in bf16 costs nothing+# while the video-stream linears keep full int8 coverage. One token covers every audio-side name, including+# video_to_audio_attn, whose keys/values are that same one-token stream.+# The audio token alone is not enough. LTX-2 also carries LTX2AdaLayerNormSingle projections whose names say nothing+# about audio -- av_cross_attn_video_scale_shift / av_cross_attn_video_a2v_gate (the cross-modality modulation,+# computed unconditionally in forward, isolate_modalities or not) and prompt_adaln / audio_prompt_adaln. Each is a+# Linear over a BATCH-sized input, so M = batch = 1 for a default training run: the same _int_mm floor, reached+# whatever the audio stream does.+_LTX2_INT8_EXCLUDES = ("audio", "av_cross_attn", "adaln")+# Untiled decode peaks (video_vae_untiled), measured fp16 on a B200. Wan: per latent pixel (h x w), frame count does not+# enter (the causal cache decodes one latent frame at a time), plus the (spatial, temporal) RGB output Wan's decode+# holds about twice (per-frame torch.cat, then clamp); 1021 frames at 1280x704 is 5.2 GiB per fp16 copy. A14B and+# Wan2.1 share the Wan2.1 VAE (8x spatial), measured <= 0.58 MiB per latent pixel at 480p / 720p.+_WAN21_VAE_UNTILED = UntiledVae(bytes_per_latent_pixel = 0.6 * 2**20, output_scale = (8, 4))+# LTX-2 / 2.3 holds every frame: peak per output frame x latent pixel, 0.102-0.105 MiB measured bf16 (B200). LTX-2.5+# has the same conv decoder layout (vae/ of the Diffusers pack); unmeasured, priced as 2.3.+_LTX2_VAE_UNTILED = UntiledVae(bytes_per_frame_latent_pixel = 0.11 * 2**20, temporal_scale = 8)+_FAMILIES: tuple[VideoFamily, ...] = (VideoFamily(name = "minimax-h3",+ trainable = True,+ label = "MiniMax-H3",+ profile = FamilyProfile(+ # Same small-M story as HunyuanVideo-1.5 with one extra trap: the adaLN projection is ``adaln_proj``,+ # which no generic token matches ("norm" does not appear in the name). On the DENSE checkpoint it is+ # Linear(2688 -> 96768), clears min_features = 512, gets quantized, then runs at M = 1 and raises+ # ``self.size(0) needs to be greater than 16, but got 1`` at the first denoise (the offline builder+ # happily bakes it, and torch.compile then dies on that module). On the pruned-modulation form it is+ # Linear(8 -> 96768), under min_features and skipped for free (verified on the fl2va_pruned build: all 51+ # rejected), so the exclusion makes the DENSE path correct and is a no-op on the pruned one.+ int8_exclude_tokens = ("adaln_proj",),+ # H3's OTHER small-M linears (context_embedder + the two token_refiner blocks: 13 Linears, 798 M params,+ # 0.80 GB bf16, 3.9% of the int8 checkpoint) are row-PADDED to the _int_mm floor instead of excluded:+ # padding returns the caller's rows bitwise unchanged (per-row activation scales, asserted in+ # diffusion_quant_pad). Measured on B200, 640x384 x 124 frames (h3_quant/README_int8.md). qwen-image,+ # qwen-image-edit and hunyuanvideo-1.5 have the same small-M shape but each has a PUBLISHED int8+ # checkpoint whose metadata bakes its exclusion set, so they move only with a rebuild and republish.+ int8_pad_tokens = ("context_embedder", "token_refiner"),+ # The packed sequence runs three modalities through one set of linears, so its activation range is not the+ # per-family range the fp8 filter was measured against. The trainer refuses both outright; declaring it+ # here keeps /diffusion/info from advertising a start that is guaranteed to 400.+ train_scheme_deny = frozenset({"fp8", "mxfp8"}),+ auto_static_skip = (("minimaxai/minimax-h3", (("max", 2), ("steps", 30))),),+ ),pipeline_class = "ModularPipeline",transformer_class = "MiniMaxH3Transformer3DModel",base_repo = "MiniMaxAI/MiniMax-H3",@@ -474,6 +551,26 @@ _FAMILIES: tuple[VideoFamily, ...] = (# reference 1280x720 resolves to the 1280x704 latent canvas it actually renders.VideoFamily(name = "ltx-2.5",+ profile = FamilyProfile(+ # the 2.3 DiT less the video FFN bias+ int8_exclude_tokens = _LTX2_INT8_EXCLUDES,+ # Hosted rotated int8 build only: rotation baked into the file (every quantized Linear, group 256) and read+ # back from its metadata. No runtime ConvRot spec (group 0), so quantizing at load stays plain int8;+ # default_on only picks the hosted rotated file (the env's ``0`` picks the plain one).+ int8_convrot = ConvRotSpec(+ group = 0,+ suffixes = (),+ variant_repos = ("unsloth/LTX-2.5-FP8",),+ default_on = True,+ ),+ untiled_vae = _LTX2_VAE_UNTILED,+ # 2.5 keeps the 2.3 conv VAE encoder (byte-identical weights), so the same latent space.+ compile_static = True,+ stg_compile_adapter = True,+ preview_factors = "ltx2",+ # The split single-file repo next to the Diffusers pack (safetensors, no remote code).+ trusted_repos = ("lightricks/ltx-2.5",),+ ),pipeline_class = "LTX2Pipeline",transformer_class = "LTX2VideoTransformer3DModel",base_repo = "Lightricks/LTX-2.5-Diffusers",@@ -510,6 +607,18 @@ _FAMILIES: tuple[VideoFamily, ...] = (# dev config (40 steps, CFG 4).VideoFamily(name = "ltx-2",+ trainable = True,+ label = "LTX-2",+ profile = FamilyProfile(+ int8_exclude_tokens = _LTX2_INT8_EXCLUDES,+ # 2.3 is the same audiovisual DiT, and builders spell it by its version.+ quant_names = ("ltx-2.3",),+ untiled_vae = _LTX2_VAE_UNTILED,+ compile_static = True,+ stg_compile_adapter = True,+ preview_factors = "ltx2",+ trusted_repos = ("lightricks/ltx-2.3-fp8",),+ ),pipeline_class = "LTX2Pipeline",transformer_class = "LTX2VideoTransformer3DModel",base_repo = "Lightricks/LTX-2",@@ -544,6 +653,13 @@ _FAMILIES: tuple[VideoFamily, ...] = (# temporal compression 4 gives valid frame counts 4k+1. Defaults 20 steps / CFG 5 (ComfyUI's template).VideoFamily(name = "wan2.2-ti2v-5b",+ profile = FamilyProfile(+ untiled_vae = UntiledVae(bytes_per_latent_pixel = 2.5 * 2**20, output_scale = (16, 4)),+ preview_factors = "wan22_5b",+ auto_static_skip = (+ ("wan-ai/wan2.2-ti2v-5b-diffusers", (("default", 2), ("max", 3), ("steps", 50))),+ ),+ ),comfy_flow_shift = 8.0, # ComfyUI ModelSamplingSD3 8 (TI2V-5B template)pipeline_class = "WanPipeline",transformer_class = "WanTransformer3DModel",@@ -587,6 +703,14 @@ _FAMILIES: tuple[VideoFamily, ...] = (# transformer_2, so cfg2_kwarg is threaded only here.VideoFamily(name = "wan2.2-t2v-a14b",+ profile = FamilyProfile(+ auto_prefer = AutoPrefer(+ floor = (10, 0),+ schemes = ("nvfp4", "int8", "fp8", "mxfp8"),+ backend = "flashinfer",+ ),+ untiled_vae = _WAN21_VAE_UNTILED,+ ),comfy_flow_shift = 5.0, # ComfyUI ModelSamplingSD3 5 (T2V-A14B template)pipeline_class = "WanPipeline",transformer_class = "WanTransformer3DModel",@@ -634,6 +758,7 @@ _FAMILIES: tuple[VideoFamily, ...] = (# Wan2.1-T2V-1.3B: 480P only (upstream calls 720P unstable for 1.3B).VideoFamily(name = "wan2.1-t2v-1.3b",+ profile = FamilyProfile(untiled_vae = _WAN21_VAE_UNTILED),comfy_flow_shift = 8.0, # reference workflow shift; replaces the repo's flow_shift 3.0pipeline_class = "WanPipeline",transformer_class = "WanTransformer3DModel",@@ -654,6 +779,7 @@ _FAMILIES: tuple[VideoFamily, ...] = (# Wan2.1-T2V-14B: same keys as one Wan2.2 A14B expert, run as a single denoiser.VideoFamily(name = "wan2.1-t2v-14b",+ profile = FamilyProfile(untiled_vae = _WAN21_VAE_UNTILED),comfy_flow_shift = 5.0, # reference 14B workflow shift (the 1.3B uses 8); replaces the repo's flow_shift 3.0pipeline_class = "WanPipeline",transformer_class = "WanTransformer3DModel",@@ -676,6 +802,14 @@ _FAMILIES: tuple[VideoFamily, ...] = (# so only the community repacks load.VideoFamily(name = "hunyuanvideo-1.5",+ profile = FamilyProfile(+ int8_exclude_tokens = _HUNYUAN15_INT8_EXCLUDES,+ nvfp4_zero_row_tokens = _HUNYUAN15_NVFP4_ZERO_ROW_TOKENS,+ preview_factors = "hv15",+ auto_static_skip = (+ ("hunyuanvideo-community/hunyuanvideo-1.5-diffusers-480p_t2v", (("max", 2), ("steps", 50))),+ ),+ ),comfy_flow_shift = 7.0, # ComfyUI model default and template shift 7pipeline_class = "HunyuanVideo15Pipeline",transformer_class = "HunyuanVideo15Transformer3DModel",@@ -710,6 +844,11 @@ _FAMILIES: tuple[VideoFamily, ...] = (# own family so a 720p load defaults to 720p sizes; the full-path alias outranks the generic token.VideoFamily(name = "hunyuanvideo-1.5-720p",+ profile = FamilyProfile(+ int8_exclude_tokens = _HUNYUAN15_INT8_EXCLUDES,+ nvfp4_zero_row_tokens = _HUNYUAN15_NVFP4_ZERO_ROW_TOKENS,+ preview_factors = "hv15",+ ),comfy_flow_shift = 7.0, # ComfyUI model default and template shift 7pipeline_class = "HunyuanVideo15Pipeline",transformer_class = "HunyuanVideo15Transformer3DModel",diff --git a/studio/backend/core/inference/video_vae_untiled.py b/studio/backend/core/inference/video_vae_untiled.pyindex 1628aae0c0e..39856a6e615 100644--- a/studio/backend/core/inference/video_vae_untiled.py+++ b/studio/backend/core/inference/video_vae_untiled.py@@ -8,34 +8,27 @@ from __future__ import annotationsimport osfrom typing import Any, Optional+from .family_profile import profile_table as _profile_table+UNTILED_ENV = "UNSLOTH_VIDEO_VAE_UNTILED"-# Untiled fp16 decode peak per latent pixel (h x w), measured on a B200; frame count does not enter (Wan's causal cache-# decodes one latent frame at a time). Scaled by the decoder's element size: fp32 peaked at 21.1 GiB vs fp16 9.6 GiB.-# A14B: Wan2.1 VAE (8x spatial), measured <= 0.58 MiB per latent pixel at 480p / 720p.+# Untiled decode peak models per family (``FamilyProfile.untiled_vae`` on the video registry, measured on a B200; the+# reasoning sits next to the data). Per latent pixel (h x w) + the (spatial, temporal) RGB output for a causal decoder+# (Wan), or per output frame x latent pixel + the temporal scale for a decoder holding every frame (LTX-2 / 2.3 / 2.5).+_UNTILED = _profile_table("untiled_vae")_BYTES_PER_LATENT_PIXEL = {- "wan2.2-ti2v-5b": 2.5 * 2**20,- "wan2.2-t2v-a14b": 0.6 * 2**20,- "wan2.1-t2v-1.3b": 0.6 * 2**20,- "wan2.1-t2v-14b": 0.6 * 2**20,-}-# (spatial, temporal) upsampling to the RGB output, which Wan's decode holds about twice (per-frame torch.cat, then-# clamp) and which grows with frames: 1021 frames at 1280x704 is 5.2 GiB per fp16 copy.-_OUTPUT_SCALE = {- "wan2.2-ti2v-5b": (16, 4),- "wan2.2-t2v-a14b": (8, 4),- "wan2.1-t2v-1.3b": (8, 4),- "wan2.1-t2v-14b": (8, 4),+ name: m.bytes_per_latent_pixel+ for name, m in _UNTILED.items()+ if m.bytes_per_latent_pixel is not None}-# LTX-2 / 2.3 holds every frame: peak per output frame x latent pixel, 0.102-0.105 MiB measured bf16 (B200).+_OUTPUT_SCALE = {name: m.output_scale for name, m in _UNTILED.items() if m.output_scale is not None}_BYTES_PER_FRAME_LATENT_PIXEL = {- "ltx-2": 0.11 * 2**20,- # Same conv decoder layout as 2.3 (vae/ of the Diffusers pack); unmeasured, priced as 2.3.- "ltx-2.5": 0.11 * 2**20,+ name: m.bytes_per_frame_latent_pixel+ for name, m in _UNTILED.items()+ if m.bytes_per_frame_latent_pixel is not None}_TEMPORAL_SCALE = {- "ltx-2": 8,- "ltx-2.5": 8,+ name: m.temporal_scale for name, m in _UNTILED.items() if m.temporal_scale is not None}_MARGIN = 1.25_MARGIN_BYTES = 2 * 2**30diff --git a/studio/backend/core/training/diffusion_train_common.py b/studio/backend/core/training/diffusion_train_common.pyindex 6ebba8c4849..38b25e0d94d 100644--- a/studio/backend/core/training/diffusion_train_common.py+++ b/studio/backend/core/training/diffusion_train_common.py@@ -28,6 +28,7 @@ from core._torchao_stub import (is_stubbed,torch_is_rocm,)+from core.inference.diffusion_families import _FAMILIES as _IMAGE_REGISTRYfrom core.inference.diffusion_families import (detect_family,detect_family_for_pick,@@ -35,6 +36,7 @@ from core.inference.diffusion_families import (trainable_family_names,)from core.inference.diffusion_torchao_patches import install_torchao_int_mm_patch+from core.inference.video_families import _FAMILIES as _VIDEO_REGISTRYfrom core.inference.video_families import detect_video_family, supported_video_family_namesfrom utils.paths.path_utils import drop_appledouble_metadata@@ -72,9 +74,9 @@ _FORCE_BF16_FAMILIES: frozenset[str] = frozenset({"qwen-image", "z-image", "krea-2", "flux.2-klein", "flux.2-dev", "ltx-2", "minimax-h3"})-# The video registry has no trainable flag, so the trainable set lives here and every name in it MUST resolve through-# get_trainer; a video base outside it is refused by name.-TRAINABLE_VIDEO_FAMILIES: frozenset[str] = frozenset({"ltx-2", "minimax-h3"})+# ``VideoFamily.trainable``; every name in it MUST resolve through get_trainer, and a video base outside it is refused+# by name.+TRAINABLE_VIDEO_FAMILIES: frozenset[str] = frozenset(f.name for f in _VIDEO_REGISTRY if f.trainable)# Families whose flow_shift default is "auto" (reproduce the family's INFERENCE sigma distribution) rather than the# identity 1.0. Both schedulers set use_dynamic_shifting, so scheduler.sigmas is the unshifted uniform table and@@ -538,18 +540,10 @@ def train_defaults(family: str) -> dict[str, Any]:return dict(FAMILY_TRAIN_DEFAULTS.get((family or "").strip().lower(), {}))-# Display labels + a short VRAM/access note per trainable family, surfaced by the Train UI so users pick a base with-# realistic expectations.+# Display labels (the registries' ``label``) + a short VRAM/access note per trainable family, surfaced by the Train UI+# so users pick a base with realistic expectations._FAMILY_LABELS = {- "sdxl": "SDXL",- "flux.1": "FLUX.1-dev",- "qwen-image": "Qwen-Image",- "z-image": "Z-Image",- "krea-2": "Krea 2",- "flux.2-klein": "FLUX.2 Klein",- "flux.2-dev": "FLUX.2-dev",- "ltx-2": "LTX-2",- "minimax-h3": "MiniMax-H3",+ fam.name: fam.label for fam in (*_IMAGE_REGISTRY, *_VIDEO_REGISTRY) if fam.label}# params is the transformer size (SDXL is not quoted that way); note is the rest._FAMILY_TRAIN_SPECS: dict[str, dict[str, Any]] = {diff --git a/studio/backend/tests/fixtures/diffusion_family_data_freeze.json b/studio/backend/tests/fixtures/diffusion_family_data_freeze.jsonnew file mode 100644index 00000000000..3831ec8cf19--- /dev/null+++ b/studio/backend/tests/fixtures/diffusion_family_data_freeze.json@@ -0,0 +1,3605 @@+{+ "auto_static_skip_by_base": {+ "flux.1": {+ "black-forest-labs/FLUX.1-schnell": null+ },+ "flux.1-kontext": {+ "black-forest-labs/FLUX.1-Kontext-dev": null+ },+ "flux.2-dev": {+ "black-forest-labs/FLUX.2-dev": null+ },+ "flux.2-klein": {+ "black-forest-labs/FLUX.2-klein-4B": null,+ "black-forest-labs/FLUX.2-klein-9B": null,+ "black-forest-labs/FLUX.2-klein-base-4B": {+ "default": 2,+ "max": 3,+ "steps": 50+ },+ "black-forest-labs/FLUX.2-klein-base-9B": null+ },+ "hidream-i1": {+ "HiDream-ai/HiDream-I1-Full": null+ },+ "hunyuanimage-2.1": {+ "hunyuanvideo-community/HunyuanImage-2.1-Diffusers": {+ "max": 2,+ "steps": 50+ }+ },+ "hunyuanvideo-1.5": {+ "hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v": {+ "max": 2,+ "steps": 50+ }+ },+ "hunyuanvideo-1.5-720p": {+ "hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_t2v": null+ },+ "ideogram-4": {+ "ideogram-ai/ideogram-4-fp8": null+ },+ "krea-2": {+ "krea/Krea-2-Turbo": null+ },+ "ltx-2": {+ "Lightricks/LTX-2": null+ },+ "ltx-2.5": {+ "Lightricks/LTX-2.5-Diffusers": null+ },+ "lumina-2": {+ "Alpha-VLLM/Lumina-Image-2.0": null+ },+ "minimax-h3": {+ "MiniMaxAI/MiniMax-H3": {+ "max": 2,+ "steps": 30+ }+ },+ "qwen-image": {+ "Qwen/Qwen-Image": {+ "default": 2,+ "max": 3,+ "steps": 20+ }+ },+ "qwen-image-2.1": {+ "Qwen/Qwen-Image-2.1": {+ "default": 3,+ "max": 3,+ "steps": 25+ },+ "Qwen/Qwen-Image-2.1-Turbo": null+ },+ "qwen-image-edit": {+ "Qwen/Qwen-Image-Edit-2511": null+ },+ "qwen-image-layered": {+ "Qwen/Qwen-Image-Layered": null+ },+ "sdxl": {+ "stabilityai/stable-diffusion-xl-base-1.0": null+ },+ "wan2.1-t2v-1.3b": {+ "Wan-AI/Wan2.1-T2V-1.3B-Diffusers": null+ },+ "wan2.1-t2v-14b": {+ "Wan-AI/Wan2.1-T2V-14B-Diffusers": null+ },+ "wan2.2-t2v-a14b": {+ "Wan-AI/Wan2.2-T2V-A14B-Diffusers": null+ },+ "wan2.2-ti2v-5b": {+ "Wan-AI/Wan2.2-TI2V-5B-Diffusers": {+ "default": 2,+ "max": 3,+ "steps": 50+ }+ },+ "z-image": {+ "Tongyi-MAI/Z-Image-Turbo": null+ }+ },+ "defaults": {+ "flux.1": [+ 4,+ 0.0+ ],+ "flux.1-kontext": [+ 20,+ 2.5+ ],+ "flux.2-dev": [+ 20,+ 4.0+ ],+ "flux.2-klein": [+ 4,+ 1.0+ ],+ "hidream-i1": [+ 50,+ 5.0+ ],+ "hunyuanimage-2.1": [+ 50,+ 3.25+ ],+ "hunyuanvideo-1.5": {+ "fields": [+ 20,+ 6.0+ ],+ "resolved": [+ 20,+ 6.0+ ]+ },+ "hunyuanvideo-1.5-720p": {+ "fields": [+ 20,+ 6.0+ ],+ "resolved": [+ 20,+ 6.0+ ]+ },+ "ideogram-4": [+ 20,+ 7.0+ ],+ "krea-2": [+ 8,+ 0.0+ ],+ "ltx-2": {+ "fields": [+ 40,+ 4.0+ ],+ "resolved": [+ 40,+ 4.0+ ]+ },+ "ltx-2.5": {+ "fields": [+ 8,+ 1.0+ ],+ "resolved": [+ 8,+ 1.0+ ]+ },+ "lumina-2": [+ 50,+ 4.0+ ],+ "minimax-h3": {+ "fields": [+ 30,+ 1.0+ ],+ "resolved": [+ 30,+ 1.0+ ]+ },+ "qwen-image": [+ 20,+ 4.0+ ],+ "qwen-image-2.1": [+ 25,+ 1.0+ ],+ "qwen-image-edit": [+ 40,+ 4.0+ ],+ "qwen-image-layered": [+ 20,+ 2.5+ ],+ "sdxl": [+ 25,+ 7.0+ ],+ "wan2.1-t2v-1.3b": {+ "fields": [+ 30,+ 6.0+ ],+ "resolved": [+ 30,+ 6.0+ ]+ },+ "wan2.1-t2v-14b": {+ "fields": [+ 20,+ 6.0+ ],+ "resolved": [+ 20,+ 6.0+ ]+ },+ "wan2.2-t2v-a14b": {+ "fields": [+ 20,+ 3.5+ ],+ "resolved": [+ 20,+ 3.5+ ]+ },+ "wan2.2-ti2v-5b": {+ "fields": [+ 20,+ 5.0+ ],+ "resolved": [+ 20,+ 5.0+ ]+ },+ "z-image": [+ 8,+ 0.0+ ]+ },+ "flags": {+ "flux.1": {+ "cudnn_benchmark": false,+ "filter_reduction_configs": false,+ "filter_reduction_configs_archs": [+ [+ 8,+ 0+ ],+ [+ 8,+ 9+ ],+ [+ 12,+ 0+ ]+ ]+ },+ "flux.1-kontext": {+ "cudnn_benchmark": true,+ "filter_reduction_configs": false,+ "filter_reduction_configs_archs": [+ [+ 8,+ 0+ ],+ [+ 8,+ 9+ ],+ [+ 12,+ 0+ ]+ ]+ },+ "flux.2-dev": {+ "cudnn_benchmark": true,+ "filter_reduction_configs": false,+ "filter_reduction_configs_archs": []+ },+ "flux.2-klein": {+ "cudnn_benchmark": true,+ "filter_reduction_configs": false,+ "filter_reduction_configs_archs": []+ },+ "hidream-i1": {+ "cudnn_benchmark": true,+ "filter_reduction_configs": false,+ "filter_reduction_configs_archs": []+ },+ "hunyuanimage-2.1": {+ "cudnn_benchmark": true,+ "filter_reduction_configs": false,+ "filter_reduction_configs_archs": []+ },+ "hunyuanvideo-1.5": {+ "cudnn_benchmark": false,+ "filter_reduction_configs": false,+ "filter_reduction_configs_archs": []+ },+ "hunyuanvideo-1.5-720p": {+ "cudnn_benchmark": false,+ "filter_reduction_configs": false,+ "filter_reduction_configs_archs": []+ },+ "ideogram-4": {+ "cudnn_benchmark": true,+ "filter_reduction_configs": false,+ "filter_reduction_configs_archs": []+ },+ "krea-2": {+ "cudnn_benchmark": true,+ "filter_reduction_configs": false,+ "filter_reduction_configs_archs": []+ },+ "ltx-2": {+ "cudnn_benchmark": false,+ "filter_reduction_configs": true,+ "filter_reduction_configs_archs": []+ },+ "ltx-2.5": {+ "cudnn_benchmark": false,+ "filter_reduction_configs": true,+ "filter_reduction_configs_archs": []+ },+ "lumina-2": {+ "cudnn_benchmark": true,+ "filter_reduction_configs": false,+ "filter_reduction_configs_archs": []+ },+ "minimax-h3": {+ "cudnn_benchmark": true,+ "filter_reduction_configs": false,+ "filter_reduction_configs_archs": []+ },+ "qwen-image": {+ "cudnn_benchmark": false,+ "filter_reduction_configs": false,+ "filter_reduction_configs_archs": [+ [+ 8,+ 0+ ],+ [+ 8,+ 9+ ],+ [+ 12,+ 0+ ]+ ]+ },+ "qwen-image-2.1": {+ "cudnn_benchmark": true,+ "filter_reduction_configs": false,+ "filter_reduction_configs_archs": []+ },+ "qwen-image-edit": {+ "cudnn_benchmark": true,+ "filter_reduction_configs": false,+ "filter_reduction_configs_archs": [+ [+ 8,+ 0+ ],+ [+ 8,+ 9+ ],+ [+ 12,+ 0+ ]+ ]+ },+ "qwen-image-layered": {+ "cudnn_benchmark": true,+ "filter_reduction_configs": false,+ "filter_reduction_configs_archs": [+ [+ 8,+ 0+ ],+ [+ 8,+ 9+ ],+ [+ 12,+ 0+ ]+ ]+ },+ "sdxl": {+ "cudnn_benchmark": false,+ "filter_reduction_configs": false,+ "filter_reduction_configs_archs": []+ },+ "wan2.1-t2v-1.3b": {+ "cudnn_benchmark": false,+ "filter_reduction_configs": false,+ "filter_reduction_configs_archs": []+ },+ "wan2.1-t2v-14b": {+ "cudnn_benchmark": false,+ "filter_reduction_configs": false,+ "filter_reduction_configs_archs": []+ },+ "wan2.2-t2v-a14b": {+ "cudnn_benchmark": false,+ "filter_reduction_configs": false,+ "filter_reduction_configs_archs": []+ },+ "wan2.2-ti2v-5b": {+ "cudnn_benchmark": false,+ "filter_reduction_configs": false,+ "filter_reduction_configs_archs": []+ },+ "z-image": {+ "cudnn_benchmark": false,+ "filter_reduction_configs": false,+ "filter_reduction_configs_archs": [+ [+ 8,+ 0+ ],+ [+ 8,+ 9+ ],+ [+ 12,+ 0+ ]+ ]+ }+ },+ "per_family": {+ "": {+ "auto": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "auto_bf16_when_resident": null,+ "bf16": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "convrot_default_on": false,+ "fp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed"+ ],+ "pad": [],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "page": null,+ "preview_key": null,+ "te_int8_effective": "fp8"+ },+ " qwen-image ": {+ "auto": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "auto_bf16_when_resident": null,+ "bf16": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "convrot_default_on": false,+ "fp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed",+ "txt_in",+ "add_q_proj",+ "add_k_proj",+ "add_v_proj",+ "to_add_out",+ "txt_mlp"+ ],+ "pad": [],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "page": "image",+ "preview_key": "qwen_image",+ "te_int8_effective": "fp8"+ },+ "<None>": {+ "auto": {+ "convrot_spec": [+ 0,+ []+ ],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "bf16": {+ "convrot_spec": [+ 0,+ []+ ],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "fp8": {+ "convrot_spec": [+ 0,+ []+ ],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_spec": [+ 0,+ []+ ],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed"+ ],+ "pad": [],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_spec": [+ 0,+ []+ ],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_spec": [+ 0,+ []+ ],+ "exclude": [],+ "pad": [],+ "zero_row": []+ }+ },+ "QWEN-IMAGE": {+ "auto": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "auto_bf16_when_resident": null,+ "bf16": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "convrot_default_on": false,+ "fp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed",+ "txt_in",+ "add_q_proj",+ "add_k_proj",+ "add_v_proj",+ "to_add_out",+ "txt_mlp"+ ],+ "pad": [],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "page": "image",+ "preview_key": "qwen_image",+ "te_int8_effective": "int8"+ },+ "flux.1": {+ "auto": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "auto_bf16_when_resident": null,+ "bf16": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "convrot_default_on": false,+ "fp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed"+ ],+ "pad": [],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "page": "image",+ "preview_key": "flux1",+ "te_int8_effective": "fp8"+ },+ "flux.1-kontext": {+ "auto": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "auto_bf16_when_resident": null,+ "bf16": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "convrot_default_on": false,+ "fp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed"+ ],+ "pad": [],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "page": "image",+ "preview_key": "flux1",+ "te_int8_effective": "fp8"+ },+ "flux.2-dev": {+ "auto": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "auto_bf16_when_resident": null,+ "bf16": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "convrot_default_on": false,+ "fp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed"+ ],+ "pad": [],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "page": "image",+ "preview_key": "flux2",+ "te_int8_effective": "int8"+ },+ "flux.2-klein": {+ "auto": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "auto_bf16_when_resident": null,+ "bf16": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "convrot_default_on": false,+ "fp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed"+ ],+ "pad": [],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "page": "image",+ "preview_key": "flux2",+ "te_int8_effective": "fp8"+ },+ "hidream-i1": {+ "auto": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "auto_bf16_when_resident": "INT8 is barely faster than compiled bf16 on a card that holds bf16 and changes image detail",+ "bf16": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "convrot_default_on": false,+ "fp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed"+ ],+ "pad": [],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "page": "image",+ "preview_key": null,+ "te_int8_effective": "fp8"+ },+ "hunyuanimage-2.1": {+ "auto": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "auto_bf16_when_resident": null,+ "bf16": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "convrot_default_on": false,+ "fp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed"+ ],+ "pad": [],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "page": "image",+ "preview_key": null,+ "te_int8_effective": "fp8"+ },+ "hunyuanvideo-1.5": {+ "auto": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "auto_bf16_when_resident": null,+ "bf16": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "convrot_default_on": false,+ "fp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed",+ "context_embedder",+ "image_embedder",+ "add_q_proj",+ "add_k_proj",+ "add_v_proj",+ "to_add_out",+ "ff_context"+ ],+ "pad": [],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": [+ "image_embedder",+ "context_embedder_2"+ ]+ },+ "page": "video",+ "preview_key": "hv15",+ "te_int8_effective": "fp8"+ },+ "hunyuanvideo-1.5-720p": {+ "auto": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "auto_bf16_when_resident": null,+ "bf16": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "convrot_default_on": false,+ "fp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed",+ "context_embedder",+ "image_embedder",+ "add_q_proj",+ "add_k_proj",+ "add_v_proj",+ "to_add_out",+ "ff_context"+ ],+ "pad": [],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": [+ "image_embedder",+ "context_embedder_2"+ ]+ },+ "page": "video",+ "preview_key": "hv15",+ "te_int8_effective": "fp8"+ },+ "ideogram-4": {+ "auto": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "auto_bf16_when_resident": null,+ "bf16": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "convrot_default_on": false,+ "fp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed"+ ],+ "pad": [],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "page": "image",+ "preview_key": null,+ "te_int8_effective": "fp8"+ },+ "krea-2": {+ "auto": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "auto_bf16_when_resident": null,+ "bf16": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "convrot_default_on": false,+ "fp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed"+ ],+ "pad": [],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "page": "image",+ "preview_key": null,+ "te_int8_effective": "fp8"+ },+ "ltx-2": {+ "auto": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "auto_bf16_when_resident": null,+ "bf16": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "convrot_default_on": false,+ "fp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed",+ "audio",+ "av_cross_attn",+ "adaln"+ ],+ "pad": [],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "page": "video",+ "preview_key": "ltx2",+ "te_int8_effective": "fp8"+ },+ "ltx-2.5": {+ "auto": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "auto_bf16_when_resident": null,+ "bf16": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "convrot_default_on": true,+ "fp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [+ "unsloth/LTX-2.5-FP8"+ ],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed",+ "audio",+ "av_cross_attn",+ "adaln"+ ],+ "pad": [],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "page": "video",+ "preview_key": "ltx2",+ "te_int8_effective": "fp8"+ },+ "lumina-2": {+ "auto": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "auto_bf16_when_resident": "INT8 is no faster than compiled bf16 on a card that holds bf16 and changes image detail",+ "bf16": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "convrot_default_on": false,+ "fp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed"+ ],+ "pad": [],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "page": "image",+ "preview_key": null,+ "te_int8_effective": "fp8"+ },+ "minimax-h3": {+ "auto": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "auto_bf16_when_resident": null,+ "bf16": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "convrot_default_on": false,+ "fp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed",+ "adaln_proj"+ ],+ "pad": [+ "context_embedder",+ "token_refiner"+ ],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "page": "video",+ "preview_key": null,+ "te_int8_effective": "fp8"+ },+ "qwen-image": {+ "auto": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "auto_bf16_when_resident": null,+ "bf16": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "convrot_default_on": false,+ "fp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed",+ "txt_in",+ "add_q_proj",+ "add_k_proj",+ "add_v_proj",+ "to_add_out",+ "txt_mlp"+ ],+ "pad": [],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "page": "image",+ "preview_key": "qwen_image",+ "te_int8_effective": "int8"+ },+ "qwen-image-2.1": {+ "auto": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "auto_bf16_when_resident": null,+ "bf16": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "convrot_default_on": true,+ "fp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_filename": "Qwen-Image-2.1-INT8-ConvRot.safetensors",+ "convrot_repo": "unsloth/Qwen-Image-2.1-FP8",+ "convrot_spec": [+ 256,+ [+ "attn.to_q",+ "attn.to_k",+ "attn.to_v",+ "attn.to_out.0",+ "img_mlp.gate_layer",+ "img_mlp.proj",+ "img_mlp.out"+ ]+ ],+ "convrot_variant_repos": [+ "unsloth/Qwen-Image-2.1-Turbo-FP8",+ "unsloth/qwen-image-2.1-turbo-fp8"+ ],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed",+ "txt_in"+ ],+ "pad": [],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "page": "image",+ "preview_key": "qwen_image_21",+ "te_int8_effective": "fp8"+ },+ "qwen-image-edit": {+ "auto": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "auto_bf16_when_resident": null,+ "bf16": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "convrot_default_on": false,+ "fp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed",+ "txt_in",+ "add_q_proj",+ "add_k_proj",+ "add_v_proj",+ "to_add_out",+ "txt_mlp"+ ],+ "pad": [],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "page": "image",+ "preview_key": "qwen_image",+ "te_int8_effective": "int8"+ },+ "qwen-image-layered": {+ "auto": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "auto_bf16_when_resident": null,+ "bf16": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "convrot_default_on": false,+ "fp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed",+ "txt_in",+ "add_q_proj",+ "add_k_proj",+ "add_v_proj",+ "to_add_out",+ "txt_mlp"+ ],+ "pad": [],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "page": "image",+ "preview_key": null,+ "te_int8_effective": "int8"+ },+ "sdxl": {+ "auto": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "auto_bf16_when_resident": null,+ "bf16": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "convrot_default_on": false,+ "fp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed"+ ],+ "pad": [],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "page": "image",+ "preview_key": "sdxl",+ "te_int8_effective": "fp8"+ },+ "unknown-family": {+ "auto": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "auto_bf16_when_resident": null,+ "bf16": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "convrot_default_on": false,+ "fp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed"+ ],+ "pad": [],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "page": "image",+ "preview_key": null,+ "te_int8_effective": "fp8"+ },+ "wan2.1-t2v-1.3b": {+ "auto": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "auto_bf16_when_resident": null,+ "bf16": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "convrot_default_on": false,+ "fp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed"+ ],+ "pad": [],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "page": "video",+ "preview_key": null,+ "te_int8_effective": "fp8"+ },+ "wan2.1-t2v-14b": {+ "auto": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "auto_bf16_when_resident": null,+ "bf16": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "convrot_default_on": false,+ "fp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed"+ ],+ "pad": [],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "page": "video",+ "preview_key": null,+ "te_int8_effective": "fp8"+ },+ "wan2.2-t2v-a14b": {+ "auto": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "auto_bf16_when_resident": null,+ "bf16": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "convrot_default_on": false,+ "fp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed"+ ],+ "pad": [],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "page": "video",+ "preview_key": null,+ "te_int8_effective": "fp8"+ },+ "wan2.2-ti2v-5b": {+ "auto": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "auto_bf16_when_resident": null,+ "bf16": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "convrot_default_on": false,+ "fp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed"+ ],+ "pad": [],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "page": "video",+ "preview_key": "wan22_5b",+ "te_int8_effective": "fp8"+ },+ "z-image": {+ "auto": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "auto_bf16_when_resident": null,+ "bf16": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "convrot_default_on": true,+ "fp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "int8": {+ "convrot_filename": "Z-Image-Turbo-INT8-ConvRot.safetensors",+ "convrot_repo": "unsloth/Z-Image-Turbo-FP8",+ "convrot_spec": [+ 256,+ [+ "attention.to_q",+ "attention.to_k",+ "attention.to_v",+ "attention.to_out.0",+ "feed_forward.w1",+ "feed_forward.w2",+ "feed_forward.w3",+ "cap_embedder.1"+ ]+ ],+ "convrot_variant_repos": [],+ "exclude": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed"+ ],+ "pad": [],+ "zero_row": []+ },+ "mxfp8": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "nvfp4": {+ "convrot_filename": null,+ "convrot_repo": null,+ "convrot_spec": [+ 0,+ []+ ],+ "convrot_variant_repos": [],+ "exclude": [],+ "pad": [],+ "zero_row": []+ },+ "page": "image",+ "preview_key": "zimage",+ "te_int8_effective": "fp8"+ }+ },+ "tables": {+ "auto_static_skip": {+ "black-forest-labs/flux.1-dev": {+ "max": 3,+ "steps": 28+ },+ "black-forest-labs/flux.1-krea-dev": {+ "default": 2,+ "max": 3,+ "steps": 28+ },+ "black-forest-labs/flux.2-klein-base-4b": {+ "default": 2,+ "max": 3,+ "steps": 50+ },+ "hunyuanvideo-community/hunyuanimage-2.1-diffusers": {+ "max": 2,+ "steps": 50+ },+ "hunyuanvideo-community/hunyuanvideo-1.5-diffusers-480p_t2v": {+ "max": 2,+ "steps": 50+ },+ "minimaxai/minimax-h3": {+ "max": 2,+ "steps": 30+ },+ "qwen/qwen-image": {+ "default": 2,+ "max": 3,+ "steps": 20+ },+ "qwen/qwen-image-2.1": {+ "default": 3,+ "max": 3,+ "steps": 25+ },+ "wan-ai/wan2.2-ti2v-5b-diffusers": {+ "default": 2,+ "max": 3,+ "steps": 50+ }+ },+ "content_video_families": [+ "hunyuanvideo-1.5",+ "hunyuanvideo-1.5-720p",+ "ltx-2",+ "ltx-2.5",+ "minimax-h3",+ "wan2.1-t2v-1.3b",+ "wan2.1-t2v-14b",+ "wan2.2-t2v-a14b",+ "wan2.2-ti2v-5b"+ ],+ "family_auto_bf16_when_resident": {+ "hidream-i1": "INT8 is barely faster than compiled bf16 on a card that holds bf16 and changes image detail",+ "lumina-2": "INT8 is no faster than compiled bf16 on a card that holds bf16 and changes image detail"+ },+ "family_auto_prefer": {+ "flux.1": {+ "backend": null,+ "consumer_ok": false,+ "floor": [+ 10,+ 0+ ],+ "gated": true,+ "schemes": [+ "int8",+ "fp8",+ "nvfp4",+ "mxfp8"+ ]+ },+ "qwen-image": {+ "backend": null,+ "consumer_ok": false,+ "floor": [+ 10,+ 0+ ],+ "gated": true,+ "schemes": [+ "int8",+ "fp8",+ "nvfp4",+ "mxfp8"+ ]+ },+ "qwen-image-2.1": {+ "backend": null,+ "consumer_ok": false,+ "floor": [+ 10,+ 0+ ],+ "gated": true,+ "schemes": [+ "int8",+ "fp8",+ "nvfp4",+ "mxfp8"+ ]+ },+ "wan2.2-t2v-a14b": {+ "backend": "flashinfer",+ "consumer_ok": false,+ "floor": [+ 10,+ 0+ ],+ "gated": false,+ "schemes": [+ "nvfp4",+ "int8",+ "fp8",+ "mxfp8"+ ]+ },+ "z-image": {+ "backend": null,+ "consumer_ok": false,+ "floor": [+ 10,+ 0+ ],+ "gated": true,+ "schemes": [+ "int8",+ "fp8",+ "nvfp4",+ "mxfp8"+ ]+ }+ },+ "family_scheme_deny": {+ "qwen-image": [+ "mxfp8",+ "nvfp4"+ ],+ "qwen-image-edit": [+ "mxfp8",+ "nvfp4"+ ],+ "qwen-image-layered": [+ "mxfp8",+ "nvfp4"+ ]+ },+ "family_train_scheme_deny": {+ "minimax-h3": [+ "fp8",+ "mxfp8"+ ],+ "qwen-image": [+ "fp8"+ ],+ "qwen-image-edit": [+ "fp8"+ ]+ },+ "int8_exclude_base": [+ "norm",+ "_mod",+ "modulation",+ "timestep_embed",+ "guidance_embed",+ "time_text_embed",+ "pooled",+ "time_embed"+ ],+ "int8_family_convrot": {+ "qwen-image-2.1": [+ 256,+ [+ "attn.to_q",+ "attn.to_k",+ "attn.to_v",+ "attn.to_out.0",+ "img_mlp.gate_layer",+ "img_mlp.proj",+ "img_mlp.out"+ ]+ ],+ "z-image": [+ 256,+ [+ "attention.to_q",+ "attention.to_k",+ "attention.to_v",+ "attention.to_out.0",+ "feed_forward.w1",+ "feed_forward.w2",+ "feed_forward.w3",+ "cap_embedder.1"+ ]+ ]+ },+ "int8_family_convrot_default_on": [+ "ltx-2.5",+ "qwen-image-2.1",+ "z-image"+ ],+ "int8_family_convrot_filename": {+ "qwen-image-2.1": "Qwen-Image-2.1-INT8-ConvRot.safetensors",+ "z-image": "Z-Image-Turbo-INT8-ConvRot.safetensors"+ },+ "int8_family_convrot_repo": {+ "qwen-image-2.1": "unsloth/Qwen-Image-2.1-FP8",+ "z-image": "unsloth/Z-Image-Turbo-FP8"+ },+ "int8_family_convrot_variant_repos": {+ "ltx-2.5": [+ "unsloth/LTX-2.5-FP8"+ ],+ "qwen-image-2.1": [+ "unsloth/Qwen-Image-2.1-Turbo-FP8"+ ]+ },+ "int8_family_exclude": {+ "hunyuanvideo-1.5": [+ "context_embedder",+ "image_embedder",+ "add_q_proj",+ "add_k_proj",+ "add_v_proj",+ "to_add_out",+ "ff_context"+ ],+ "hunyuanvideo-1.5-720p": [+ "context_embedder",+ "image_embedder",+ "add_q_proj",+ "add_k_proj",+ "add_v_proj",+ "to_add_out",+ "ff_context"+ ],+ "ltx-2": [+ "audio",+ "av_cross_attn",+ "adaln"+ ],+ "ltx-2.3": [+ "audio",+ "av_cross_attn",+ "adaln"+ ],+ "ltx-2.5": [+ "audio",+ "av_cross_attn",+ "adaln"+ ],+ "minimax-h3": [+ "adaln_proj"+ ],+ "qwen-image": [+ "txt_in",+ "add_q_proj",+ "add_k_proj",+ "add_v_proj",+ "to_add_out",+ "txt_mlp"+ ],+ "qwen-image-2.1": [+ "txt_in"+ ],+ "qwen-image-edit": [+ "txt_in",+ "add_q_proj",+ "add_k_proj",+ "add_v_proj",+ "to_add_out",+ "txt_mlp"+ ],+ "qwen-image-layered": [+ "txt_in",+ "add_q_proj",+ "add_k_proj",+ "add_v_proj",+ "to_add_out",+ "txt_mlp"+ ]+ },+ "int8_family_pad": {+ "minimax-h3": [+ "context_embedder",+ "token_refiner"+ ]+ },+ "nvfp4_family_zero_row": {+ "hunyuanvideo-1.5": [+ "image_embedder",+ "context_embedder_2"+ ],+ "hunyuanvideo-1.5-720p": [+ "image_embedder",+ "context_embedder_2"+ ]+ },+ "preview_family_factors": {+ "flux.1": "flux1",+ "flux.1-kontext": "flux1",+ "flux.2-dev": "flux2",+ "flux.2-klein": "flux2",+ "hunyuanvideo-1.5": "hv15",+ "hunyuanvideo-1.5-720p": "hv15",+ "ltx-2": "ltx2",+ "ltx-2.5": "ltx2",+ "qwen-image": "qwen_image",+ "qwen-image-2.1": "qwen_image_21",+ "qwen-image-edit": "qwen_image",+ "sdxl": "sdxl",+ "wan2.2-ti2v-5b": "wan22_5b",+ "z-image": "zimage"+ },+ "te_int8_skip": {+ "flux.2-dev": [+ 3,+ 0+ ],+ "qwen-image": [+ 6,+ 6+ ],+ "qwen-image-edit": [+ 6,+ 6+ ],+ "qwen-image-layered": [+ 6,+ 6+ ]+ },+ "train_family_labels": {+ "flux.1": "FLUX.1-dev",+ "flux.2-dev": "FLUX.2-dev",+ "flux.2-klein": "FLUX.2 Klein",+ "krea-2": "Krea 2",+ "ltx-2": "LTX-2",+ "minimax-h3": "MiniMax-H3",+ "qwen-image": "Qwen-Image",+ "sdxl": "SDXL",+ "z-image": "Z-Image"+ },+ "trainable_video_families": [+ "ltx-2",+ "minimax-h3"+ ],+ "trusted_image_repos": [+ "alpha-vllm/lumina-image-2.0",+ "black-forest-labs/flux.1-dev",+ "black-forest-labs/flux.1-kontext-dev",+ "black-forest-labs/flux.1-krea-dev",+ "black-forest-labs/flux.1-schnell",+ "black-forest-labs/flux.2-dev",+ "black-forest-labs/flux.2-klein-4b",+ "black-forest-labs/flux.2-klein-9b",+ "black-forest-labs/flux.2-klein-base-4b",+ "black-forest-labs/flux.2-klein-base-9b",+ "hidream-ai/hidream-i1-dev",+ "hidream-ai/hidream-i1-fast",+ "hidream-ai/hidream-i1-full",+ "hunyuanvideo-community/hunyuanimage-2.1-diffusers",+ "ideogram-ai/ideogram-4-fp8",+ "ideogram-ai/ideogram-4-nf4",+ "ideogram-ai/ideogram-4-nf4-diffusers",+ "krea/krea-2-raw",+ "krea/krea-2-turbo",+ "qwen/qwen-image",+ "qwen/qwen-image-2.1",+ "qwen/qwen-image-2.1-turbo",+ "qwen/qwen-image-2512",+ "qwen/qwen-image-edit-2511",+ "qwen/qwen-image-layered",+ "stabilityai/sdxl-turbo",+ "stabilityai/stable-diffusion-xl-base-1.0",+ "tongyi-mai/z-image",+ "tongyi-mai/z-image-turbo"+ ],+ "trusted_video_repos": [+ "hunyuanvideo-community/hunyuanvideo-1.5-diffusers-480p_t2v",+ "hunyuanvideo-community/hunyuanvideo-1.5-diffusers-720p_t2v",+ "lightricks/ltx-2",+ "lightricks/ltx-2.3",+ "lightricks/ltx-2.3-fp8",+ "lightricks/ltx-2.5",+ "lightricks/ltx-2.5-diffusers",+ "minimaxai/minimax-h3",+ "wan-ai/wan2.1-t2v-1.3b-diffusers",+ "wan-ai/wan2.1-t2v-14b-diffusers",+ "wan-ai/wan2.2-t2v-a14b-diffusers",+ "wan-ai/wan2.2-ti2v-5b-diffusers"+ ],+ "untiled_bytes_per_frame_latent_pixel": {+ "ltx-2": 115343.36,+ "ltx-2.5": 115343.36+ },+ "untiled_bytes_per_latent_pixel": {+ "wan2.1-t2v-1.3b": 629145.6,+ "wan2.1-t2v-14b": 629145.6,+ "wan2.2-t2v-a14b": 629145.6,+ "wan2.2-ti2v-5b": 2621440.0+ },+ "untiled_output_scale": {+ "wan2.1-t2v-1.3b": [+ 8,+ 4+ ],+ "wan2.1-t2v-14b": [+ 8,+ 4+ ],+ "wan2.2-t2v-a14b": [+ 8,+ 4+ ],+ "wan2.2-ti2v-5b": [+ 16,+ 4+ ]+ },+ "untiled_temporal_scale": {+ "ltx-2": 8,+ "ltx-2.5": 8+ }+ },+ "untiled_decode_bytes": {+ "": [+ null,+ null,+ null,+ null,+ null,+ null+ ],+ " qwen-image ": [+ null,+ null,+ null,+ null,+ null,+ null+ ],+ "QWEN-IMAGE": [+ null,+ null,+ null,+ null,+ null,+ null+ ],+ "hunyuanvideo-1.5": [+ null,+ null,+ null,+ null,+ null,+ null+ ],+ "hunyuanvideo-1.5-720p": [+ null,+ null,+ null,+ null,+ null,+ null+ ],+ "ltx-2": [+ 23751504691,+ 47503009382,+ 195696158310,+ 391392316620,+ 12281760972,+ 24563521945+ ],+ "ltx-2.5": [+ 23751504691,+ 47503009382,+ 195696158310,+ 391392316620,+ 12281760972,+ 24563521945+ ],+ "minimax-h3": [+ null,+ null,+ null,+ null,+ null,+ null+ ],+ "unknown-family": [+ null,+ null,+ null,+ null,+ null,+ null+ ],+ "wan2.1-t2v-1.3b": [+ 4007337984,+ 8014675968,+ 5083398144,+ 10166796288,+ 594874368,+ 1189748736+ ],+ "wan2.1-t2v-14b": [+ 4007337984,+ 8014675968,+ 5083398144,+ 10166796288,+ 594874368,+ 1189748736+ ],+ "wan2.2-t2v-a14b": [+ 4007337984,+ 8014675968,+ 5083398144,+ 10166796288,+ 594874368,+ 1189748736+ ],+ "wan2.2-ti2v-5b": [+ 16683663360,+ 33367326720,+ 21071790080,+ 42143580160,+ 2471772160,+ 4943544320+ ]+ }+}diff --git a/studio/backend/tests/test_diffusion_family_data_freeze.py b/studio/backend/tests/test_diffusion_family_data_freeze.pynew file mode 100644index 00000000000..1ffd0e243a5--- /dev/null+++ b/studio/backend/tests/test_diffusion_family_data_freeze.py@@ -0,0 +1,194 @@+# SPDX-License-Identifier: AGPL-3.0-only+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0++"""Freeze every resolved per-family table value, so moving the tables into family data cannot change them.++The int8 exclude / ConvRot / NVFP4 policy sets are baked into published prequant artifacts and checked at load+(``diffusion_prequant._validate_checkpoint`` requires the artifact's exclude tokens to be a superset of+``exclude_tokens_for_scheme``), so a silent change here invalidates hosted checkpoints. The rest (preview factor keys,+untiled VAE coefficients, labels, defaults, trusted repos, ...) is frozen with them so a data move stays a pure move.++A deliberate change regenerates the fixture: ``UNSLOTH_FREEZE_UPDATE=1 pytest tests/test_diffusion_family_data_freeze.py``+and the diff of ``tests/fixtures/diffusion_family_data_freeze.json`` is the review artifact.+"""++from __future__ import annotations++import dataclasses+import json+import os+from pathlib import Path++import pytest++FIXTURE = Path(__file__).parent / "fixtures" / "diffusion_family_data_freeze.json"++_EXTRA_NAMES = ("", "unknown-family", "QWEN-IMAGE", " qwen-image ")+++def _jsonable(value):+ if dataclasses.is_dataclass(value) and not isinstance(value, type):+ return {f.name: _jsonable(getattr(value, f.name)) for f in dataclasses.fields(value)}+ if isinstance(value, dict):+ return {str(k): _jsonable(v) for k, v in sorted(value.items(), key = lambda kv: str(kv[0]))}+ if isinstance(value, (frozenset, set)):+ return sorted(_jsonable(v) for v in value)+ if isinstance(value, (tuple, list)):+ return [_jsonable(v) for v in value]+ return value+++def _snapshot() -> dict:+ from core.inference import diffusion_transformer_quant as tq+ from core.inference import diffusion_precision as prec+ from core.inference import video_vae_untiled as vu+ from core.inference import diffusion_cache as dcache+ from core.inference import diffusion_content as dcontent+ from core.inference.diffusion_families import _FAMILIES as IMAGE, default_generation_params+ from core.inference.video_families import _FAMILIES as VIDEO, default_video_generation_params+ from core.inference.diffusion_preview_factors import FAMILY_FACTORS+ from core.inference import diffusion as dimg+ from core.inference import video as dvid+ from core.training import diffusion_train_common as dtc++ image_names = [f.name for f in IMAGE]+ video_names = [f.name for f in VIDEO]+ names = image_names + video_names + list(_EXTRA_NAMES)+ schemes = list(tq.TQ_SCHEMES) + ["auto", "bf16"]++ # Every repo id the registry names, as candidates for the per-repo ConvRot variant check.+ repos = set()+ for fam in (*IMAGE, *VIDEO):+ repos.add(fam.base_repo)+ for row in getattr(fam, "prequant_repos", ()):+ repos.add(row[-1])+ for row in getattr(fam, "prequant_variant_repos", ()):+ repos.update((row[0], row[-1]))+ repos.update(("unsloth/Qwen-Image-2.1-Turbo-FP8", "unsloth/qwen-image-2.1-turbo-fp8", "", "x/y"))+ repos = sorted(repos)++ per_family = {}+ for name in names:+ row = {}+ for scheme in schemes:+ variant_repos = [r for r in repos if tq.convrot_prequant_variant_repo(scheme, name, r)]+ row[scheme] = {+ "exclude": list(tq.exclude_tokens_for_scheme(scheme, name)),+ "pad": list(tq.pad_tokens_for_scheme(scheme, name)),+ "zero_row": list(tq.zero_row_tokens_for_scheme(scheme, name)),+ "convrot_spec": _jsonable(tq.convrot_spec_for_scheme(scheme, name)),+ "convrot_filename": tq.convrot_prequant_filename(scheme, name),+ "convrot_repo": tq.convrot_prequant_repo(scheme, name),+ "convrot_variant_repos": variant_repos,+ }+ row["convrot_default_on"] = tq.int8_convrot_enabled(name)+ row["auto_bf16_when_resident"] = tq.auto_bf16_when_resident_reason(name)+ row["te_int8_effective"] = prec.effective_te_quant("int8", name)+ row["preview_key"] = FAMILY_FACTORS.get(str(name).strip().lower())+ row["page"] = dcontent.family_page(name)+ per_family[name] = row+ # None as a family (the historical default) on every accessor.+ per_family["<None>"] = {+ scheme: {+ "exclude": list(tq.exclude_tokens_for_scheme(scheme)),+ "pad": list(tq.pad_tokens_for_scheme(scheme)),+ "zero_row": list(tq.zero_row_tokens_for_scheme(scheme)),+ "convrot_spec": _jsonable(tq.convrot_spec_for_scheme(scheme)),+ }+ for scheme in schemes+ }++ untiled = {}+ for name in video_names + list(_EXTRA_NAMES):+ untiled[name] = [+ vu.untiled_decode_bytes(name, shape, itemsize)+ for shape in ((1, 16, 5, 60, 104), (2, 48, 31, 44, 80), (1, 128, 16, 22, 40))+ for itemsize in (2, 4)+ ]++ static_skip = {}+ for fam in (*IMAGE, *VIDEO):+ bases = [fam.base_repo, *getattr(fam, "named_variant_bases", ())]+ static_skip[fam.name] = {b: dcache.auto_static_skip_entry(b) for b in bases}++ flags = {}+ for fam in (*IMAGE, *VIDEO):+ flags[fam.name] = {+ "cudnn_benchmark": getattr(fam, "cudnn_benchmark", True),+ "filter_reduction_configs": bool(getattr(fam, "filter_reduction_configs", False)),+ "filter_reduction_configs_archs": _jsonable(+ tuple(getattr(fam, "filter_reduction_configs_archs", None) or ())+ ),+ }++ defaults = {}+ for fam in IMAGE:+ defaults[fam.name] = list(default_generation_params(fam.base_repo))+ for fam in VIDEO:+ defaults[fam.name] = {+ "fields": [fam.default_steps, fam.default_guidance],+ "resolved": list(+ default_video_generation_params(+ fam.base_repo,+ fallback = (fam.default_steps, fam.default_guidance),+ family = fam.name,+ )+ ),+ }++ return _jsonable(+ {+ "per_family": per_family,+ "tables": {+ "int8_exclude_base": tq._INT8_EXCLUDE_NAME_TOKENS,+ "int8_family_exclude": tq._INT8_FAMILY_EXCLUDE_NAME_TOKENS,+ "int8_family_pad": tq._INT8_FAMILY_PAD_NAME_TOKENS,+ "int8_family_convrot": tq._INT8_FAMILY_CONVROT,+ "int8_family_convrot_filename": tq._INT8_FAMILY_CONVROT_FILENAME,+ "int8_family_convrot_repo": tq._INT8_FAMILY_CONVROT_REPO,+ "int8_family_convrot_variant_repos": tq._INT8_FAMILY_CONVROT_VARIANT_REPOS,+ "int8_family_convrot_default_on": tq._INT8_FAMILY_CONVROT_DEFAULT_ON,+ "nvfp4_family_zero_row": tq._NVFP4_FAMILY_ZERO_ROW_NAME_TOKENS,+ "family_scheme_deny": tq._FAMILY_SCHEME_DENY,+ "family_auto_prefer": tq._FAMILY_AUTO_PREFER,+ "family_auto_bf16_when_resident": tq._FAMILY_AUTO_BF16_WHEN_RESIDENT,+ "family_train_scheme_deny": tq._FAMILY_TRAIN_SCHEME_DENY,+ "te_int8_skip": prec._TE_INT8_SKIP,+ "untiled_bytes_per_latent_pixel": vu._BYTES_PER_LATENT_PIXEL,+ "untiled_output_scale": vu._OUTPUT_SCALE,+ "untiled_bytes_per_frame_latent_pixel": vu._BYTES_PER_FRAME_LATENT_PIXEL,+ "untiled_temporal_scale": vu._TEMPORAL_SCALE,+ "preview_family_factors": FAMILY_FACTORS,+ "auto_static_skip": dcache.AUTO_STATIC_SKIP,+ "train_family_labels": dtc._FAMILY_LABELS,+ "trainable_video_families": dtc.TRAINABLE_VIDEO_FAMILIES,+ "content_video_families": dcontent._VIDEO_FAMILIES,+ "trusted_image_repos": dimg._TRUSTED_NON_GGUF_REPOS,+ "trusted_video_repos": dvid._TRUSTED_NON_GGUF_VIDEO_REPOS,+ },+ "untiled_decode_bytes": untiled,+ "auto_static_skip_by_base": static_skip,+ "flags": flags,+ "defaults": defaults,+ }+ )+++def _clean_env(monkeypatch):+ for var in (+ "UNSLOTH_DIFFUSION_INT8_CONVROT",+ "UNSLOTH_DIFFUSION_NVFP4",+ "UNSLOTH_DIFFUSION_AUTO_STEP_SKIP",+ "UNSLOTH_VIDEO_VAE_UNTILED",+ ):+ monkeypatch.delenv(var, raising = False)+++def test_family_data_resolves_identically(_clean_env):+ snapshot = json.loads(json.dumps(_snapshot()))+ if os.environ.get("UNSLOTH_FREEZE_UPDATE") == "1":+ FIXTURE.write_text(json.dumps(snapshot, indent = 1, sort_keys = True) + "\n")+ frozen = json.loads(FIXTURE.read_text())+ for section in sorted(set(frozen) | set(snapshot)):+ assert snapshot.get(section) == frozen.get(section), section