diffusers-workflow 0.13.0__tar.gz → 0.15.0__tar.gz

{diffusers_workflow-0.13.0/diffusers_workflow.egg-info → diffusers_workflow-0.15.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: diffusers-workflow
-Version: 0.13.0
+Version: 0.15.0
 Summary: Declarative workflow engine and web UI for Hugging Face Diffusers
 Author: dkackman
 License: Apache-2.0
@@ -37,6 +37,7 @@ Requires-Dist: controlnet-aux>=0.0.10
 Requires-Dist: opencv-python-headless>=5.0.0.93
 Requires-Dist: kornia>=0.8.3
 Requires-Dist: kernels>=0.17.2
+Requires-Dist: comfy-kitchen<0.3,>=0.2.38
 Requires-Dist: spandrel>=0.4.2
 Requires-Dist: facexlib>=0.3.0
 Requires-Dist: numpy>=2.5.4
@@ -168,7 +169,7 @@ once; without it the run fails partway through with a 401/403 from the Hub.
 
 ## Drive it from an agent
 
-Then just ask. The agent has 62 tools covering the whole surface — the
+Then just ask. The agent has 65 tools covering the whole surface — the
 workflow catalog, the real diffusers pipeline signatures, the job queue, the
 gallery, the model cache:
 

{diffusers_workflow-0.13.0 → diffusers_workflow-0.15.0}/README.md

@@ -83,7 +83,7 @@ once; without it the run fails partway through with a 401/403 from the Hub.
 
 ## Drive it from an agent
 
-Then just ask. The agent has 62 tools covering the whole surface — the
+Then just ask. The agent has 65 tools covering the whole surface — the
 workflow catalog, the real diffusers pipeline signatures, the job queue, the
 gallery, the model cache:
 

{diffusers_workflow-0.13.0 → diffusers_workflow-0.15.0/diffusers_workflow.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: diffusers-workflow
-Version: 0.13.0
+Version: 0.15.0
 Summary: Declarative workflow engine and web UI for Hugging Face Diffusers
 Author: dkackman
 License: Apache-2.0
@@ -37,6 +37,7 @@ Requires-Dist: controlnet-aux>=0.0.10
 Requires-Dist: opencv-python-headless>=5.0.0.93
 Requires-Dist: kornia>=0.8.3
 Requires-Dist: kernels>=0.17.2
+Requires-Dist: comfy-kitchen<0.3,>=0.2.38
 Requires-Dist: spandrel>=0.4.2
 Requires-Dist: facexlib>=0.3.0
 Requires-Dist: numpy>=2.5.4
@@ -168,7 +169,7 @@ once; without it the run fails partway through with a 401/403 from the Hub.
 
 ## Drive it from an agent
 
-Then just ask. The agent has 62 tools covering the whole surface — the
+Then just ask. The agent has 65 tools covering the whole surface — the
 workflow catalog, the real diffusers pipeline signatures, the job queue, the
 gallery, the model cache:
 

{diffusers_workflow-0.13.0 → diffusers_workflow-0.15.0}/diffusers_workflow.egg-info/SOURCES.txt

@@ -9,6 +9,7 @@ diffusers_workflow.egg-info/entry_points.txt
 diffusers_workflow.egg-info/requires.txt
 diffusers_workflow.egg-info/top_level.txt
 dw/__init__.py
+dw/acceleration_warnings.py
 dw/adapter_compatibility.py
 dw/argument_media.py
 dw/argument_warnings.py
@@ -19,6 +20,7 @@ dw/audio_qc.py
 dw/cache_blocks.json
 dw/cache_blocks.py
 dw/content_types.py
+dw/cost_scaling.py
 dw/devices.py
 dw/disk_caches.py
 dw/dissolve_frame_errors.py
@@ -41,8 +43,10 @@ dw/lora_catalog.py
 dw/lora_catalog_schema.json
 dw/lora_hub.py
 dw/media.py
+dw/media_compare.py
 dw/media_frames.py
 dw/media_types.py
+dw/notes.py
 dw/outbound.py
 dw/output_extraction.py
 dw/path_redaction.py
@@ -59,6 +63,7 @@ dw/reference_limits.py
 dw/reference_names.py
 dw/references.py
 dw/result.py
+dw/roles.py
 dw/run.py
 dw/runs.py
 dw/scalar_result_validation.py
@@ -70,6 +75,7 @@ dw/settings.py
 dw/shot_span_preflight.py
 dw/shots.py
 dw/slice_preflight.py
+dw/sparse_attention.py
 dw/step.py
 dw/step_cache.py
 dw/step_value_checks.py
@@ -96,6 +102,7 @@ dw/workflow.py
 dw/workflow_run.py
 dw/workflow_schema.json
 dw/workspace.py
+dw/workspace_copy.py
 dw/writers.py
 dw/community_pipelines/pipeline_flux_rf_inversion.py
 dw/community_pipelines/pipeline_ltx2_refine.py
@@ -132,17 +139,22 @@ dw/pipeline_processors/pipeline.py
 dw/pipeline_processors/placement.py
 dw/pipeline_processors/progress.py
 dw/pipeline_processors/remote.py
+dw/pipeline_processors/timing.py
 dw/server/__init__.py
 dw/server/admission.py
 dw/server/api_models.py
+dw/server/api_models_base.py
+dw/server/api_models_library.py
 dw/server/app.py
 dw/server/assess.py
 dw/server/catalog.py
 dw/server/catalog_shape.py
 dw/server/deps.py
+dw/server/dispatch.py
 dw/server/enhancers.py
 dw/server/exports.py
 dw/server/guides.py
+dw/server/host_ram_need.py
 dw/server/http_security.py
 dw/server/inline_media.py
 dw/server/job_history.py
@@ -181,44 +193,44 @@ dw/server/ui/assets/csharp-BKxtCVv1.js
 dw/server/ui/assets/csp-bTuwJoIa.js
 dw/server/ui/assets/css-DIMkf-bt.js
 dw/server/ui/assets/css.worker-C6lDtRpF.js
-dw/server/ui/assets/cssMode-QzCkHHkC.js
+dw/server/ui/assets/cssMode-D_BuRO9Y.js
 dw/server/ui/assets/cypher-CVaqCwHa.js
 dw/server/ui/assets/dart-onAF5SnQ.js
 dw/server/ui/assets/dockerfile-DZFCIeNp.js
 dw/server/ui/assets/ecl-D05T4iGw.js
 dw/server/ui/assets/editor-DjpsE8to.css
-dw/server/ui/assets/editor.api-LvRbw1iL.js
+dw/server/ui/assets/editor.api-B6pi6BSE.js
 dw/server/ui/assets/editor.worker-BmdWdrHZ.js
 dw/server/ui/assets/elixir-6RTg0lbw.js
 dw/server/ui/assets/flow9-C5_-GSwl.js
-dw/server/ui/assets/freemarker2-BFEReHqW.js
+dw/server/ui/assets/freemarker2-mPiw3NV_.js
 dw/server/ui/assets/fsharp-C8Ef5oNN.js
 dw/server/ui/assets/go-C-y9NEjX.js
 dw/server/ui/assets/graphql-fmXr3nnJ.js
-dw/server/ui/assets/handlebars--qZ0WvL7.js
+dw/server/ui/assets/handlebars-BfGb596Y.js
 dw/server/ui/assets/hcl-CpzslTdj.js
-dw/server/ui/assets/html-Bs0VSJAA.js
+dw/server/ui/assets/html-xXIqw-JZ.js
 dw/server/ui/assets/html.worker-C5yZDBHz.js
-dw/server/ui/assets/htmlMode-DrsyYX5n.js
-dw/server/ui/assets/index-BgBhcVj3.js
-dw/server/ui/assets/index-DSM5LA24.css
+dw/server/ui/assets/htmlMode-B3FC_H4q.js
+dw/server/ui/assets/index-Civ6zDZP.js
+dw/server/ui/assets/index-PRSQwOeN.css
 dw/server/ui/assets/ini-sBoK_t0W.js
 dw/server/ui/assets/java-BEtHBSE6.js
-dw/server/ui/assets/javascript-C3KR7rul.js
+dw/server/ui/assets/javascript-B-FFcgST.js
 dw/server/ui/assets/json.worker-DdeUAmHa.js
-dw/server/ui/assets/jsonMode-DEXt-mqp.js
+dw/server/ui/assets/jsonMode-BELbxUQD.js
 dw/server/ui/assets/julia-Bri6UV-V.js
 dw/server/ui/assets/kotlin-BOotOW0E.js
 dw/server/ui/assets/less-B9JPFI3C.js
 dw/server/ui/assets/lexon-CfSJPG6W.js
-dw/server/ui/assets/liquid-ByBJuq1e.js
-dw/server/ui/assets/lspLanguageFeatures-6nDoG86-.js
+dw/server/ui/assets/liquid-CxXhb7_e.js
+dw/server/ui/assets/lspLanguageFeatures-sXyBqSzn.js
 dw/server/ui/assets/lua-CsQS60Ue.js
 dw/server/ui/assets/m3-D-oSqn_W.js
 dw/server/ui/assets/markdown-Cimd5fb3.js
-dw/server/ui/assets/mdx-BH2GyjV6.js
+dw/server/ui/assets/mdx-4jPGQ4yb.js
 dw/server/ui/assets/mips-CIPQ_RoX.js
-dw/server/ui/assets/monaco-B6Bmfo5o.js
+dw/server/ui/assets/monaco-ByFdQnrs.js
 dw/server/ui/assets/monaco-n7clI0si.css
 dw/server/ui/assets/msdax-DauUninz.js
 dw/server/ui/assets/mysql-SOo6toE5.js
@@ -234,10 +246,10 @@ dw/server/ui/assets/powerquery-D3hlyOfw.js
 dw/server/ui/assets/powershell-DmHpPYUd.js
 dw/server/ui/assets/protobuf-C531GsRP.js
 dw/server/ui/assets/pug-Z5eAx3Zn.js
-dw/server/ui/assets/python-C1sJMrAm.js
+dw/server/ui/assets/python-D3CL9Y6O.js
 dw/server/ui/assets/qsharp-DkqhCAOL.js
 dw/server/ui/assets/r-BwWrilGY.js
-dw/server/ui/assets/razor-BdPZ4bvZ.js
+dw/server/ui/assets/razor-F_hjmtqk.js
 dw/server/ui/assets/redis-ClamHrr6.js
 dw/server/ui/assets/redshift-DT7zqm-g.js
 dw/server/ui/assets/restructuredtext-BYgofb2h.js
@@ -257,15 +269,15 @@ dw/server/ui/assets/swift-Bxkupp3x.js
 dw/server/ui/assets/systemverilog-Bz4Y3fRF.js
 dw/server/ui/assets/tcl-DISqw1ZD.js
 dw/server/ui/assets/ts.worker-DGGa2mF8.js
-dw/server/ui/assets/tsMode-DW5i-ukN.js
+dw/server/ui/assets/tsMode-BiyDvHYY.js
 dw/server/ui/assets/twig-De2hgUGE.js
-dw/server/ui/assets/typescript-C_QzKQY3.js
+dw/server/ui/assets/typescript-DBE1Uz0U.js
 dw/server/ui/assets/typespec-B8J7ngcE.js
 dw/server/ui/assets/vb-DV3o63ZY.js
 dw/server/ui/assets/wgsl-DpFanUEy.js
-dw/server/ui/assets/workers-DHqaOvh8.js
-dw/server/ui/assets/xml-D3UEcM_2.js
-dw/server/ui/assets/yaml-C0TnGcnd.js
+dw/server/ui/assets/workers-Ci7sZ35J.js
+dw/server/ui/assets/xml-Eqpmo6a1.js
+dw/server/ui/assets/yaml-BNP_JiwW.js
 dw/tasks/assess.py
 dw/tasks/audio_dynamics.py
 dw/tasks/audio_handlers.py
@@ -322,6 +334,8 @@ dw/tasks/video_utils.py
 dw/tasks/voice_attribution.py
 dw/tasks/windows.py
 dw/tasks/zoe_depth.py
+dw/workflows/flux3_klein_upsample.json
+dw/workflows/flux3_klein_upsample_edit.json
 dw/workflows/h3_context_ir.json
 dw/workflows/test.json
 dw_mcp/__init__.py
@@ -357,17 +371,20 @@ tests/test_assemble_and_score_target_lufs.py
 tests/test_assess.py
 tests/test_assessment_rules.py
 tests/test_assets.py
+tests/test_attention_backend_trace.py
 tests/test_attention_slicing.py
 tests/test_audio_utils.py
 tests/test_best_of_n_template.py
 tests/test_borders.py
 tests/test_cache_blocks.py
+tests/test_cache_stats.py
 tests/test_catalog_shape.py
 tests/test_catalog_speed_defaults.py
 tests/test_catalog_structure.py
 tests/test_chain.py
 tests/test_ci_config.py
 tests/test_cli.py
+tests/test_compare_route.py
 tests/test_component_name_errors.py
 tests/test_component_pruning.py
 tests/test_component_type_errors.py
@@ -376,6 +393,7 @@ tests/test_concat_videos.py
 tests/test_config_objects.py
 tests/test_configuration_schema.py
 tests/test_content_types.py
+tests/test_dependencies_declared.py
 tests/test_depth_estimator.py
 tests/test_device.py
 tests/test_device_helpers.py
@@ -397,7 +415,10 @@ tests/test_find_loop_bed.py
 tests/test_finish.py
 tests/test_fit_templates.py
 tests/test_fit_to_model.py
+tests/test_flux3_catalog.py
 tests/test_for_each.py
+tests/test_forget_workspace_history.py
+tests/test_gallery_roles.py
 tests/test_gather.py
 tests/test_grade.py
 tests/test_guide_chain_validation.py
@@ -410,11 +431,13 @@ tests/test_h3_latent_upscale.py
 tests/test_h3_refine.py
 tests/test_h3_rules.py
 tests/test_h3_schedule.py
+tests/test_h3_sparse_defaults.py
 tests/test_h3_upscaler_source.py
 tests/test_h3_vram_ceiling.py
 tests/test_hf_placement.py
 tests/test_host_memory.py
 tests/test_host_memory_projection.py
+tests/test_host_ram_need.py
 tests/test_hub_cache.py
 tests/test_image_ops.py
 tests/test_image_to_text.py
@@ -468,6 +491,7 @@ tests/test_mcp_twins.py
 tests/test_mcp_workspaces.py
 tests/test_measure_sync.py
 tests/test_media_audio.py
+tests/test_media_compare.py
 tests/test_media_decode_threads.py
 tests/test_media_frames.py
 tests/test_media_info.py
@@ -479,6 +503,9 @@ tests/test_modular_pipeline.py
 tests/test_modular_progress.py
 tests/test_mps_fallback_warning.py
 tests/test_mps_rope_precision.py
+tests/test_notes.py
+tests/test_notes_api.py
+tests/test_notes_bookkeeping.py
 tests/test_null_media.py
 tests/test_numeric_coercion.py
 tests/test_observed_cost.py
@@ -512,6 +539,7 @@ tests/test_reference_limits.py
 tests/test_reference_names.py
 tests/test_reference_skip_sets.py
 tests/test_references.py
+tests/test_rerun_lineage.py
 tests/test_rerun_new_seed.py
 tests/test_resize_bucket.py
 tests/test_restore_faces_helper_cache.py
@@ -519,7 +547,9 @@ tests/test_restore_faces_video.py
 tests/test_result.py
 tests/test_result_fps.py
 tests/test_result_output_naming.py
+tests/test_roles.py
 tests/test_rule_parity.py
+tests/test_run_manifest_roles.py
 tests/test_runs.py
 tests/test_runtime_settings.py
 tests/test_scalar_result_validation.py
@@ -555,10 +585,16 @@ tests/test_shaping_logs.py
 tests/test_shot_span_preflight.py
 tests/test_shots.py
 tests/test_slice_preflight.py
+tests/test_sparse_attention_backends.py
+tests/test_sparse_attention_processor.py
+tests/test_sparse_attention_tiles.py
+tests/test_sparse_attention_wiring.py
 tests/test_speech_generation.py
 tests/test_stabilize.py
+tests/test_stars.py
 tests/test_step.py
 tests/test_step_cache.py
+tests/test_step_timing.py
 tests/test_strip_exif_and_watermark.py
 tests/test_sub_workflow_resolver.py
 tests/test_subfolders.py
@@ -578,6 +614,7 @@ tests/test_trim_video.py
 tests/test_type_helpers.py
 tests/test_type_helpers_concurrency.py
 tests/test_ui_twins.py
+tests/test_unbraced_type_warning.py
 tests/test_updater.py
 tests/test_validate_arguments.py
 tests/test_validation.py
@@ -603,4 +640,5 @@ tests/test_workflow_run_keys.py
 tests/test_workflow_snapshot.py
 tests/test_workflow_step_cache.py
 tests/test_workflow_trust.py
-tests/test_workspace.py
+tests/test_workspace.py
+tests/test_workspace_copy.py

{diffusers_workflow-0.13.0 → diffusers_workflow-0.15.0}/diffusers_workflow.egg-info/requires.txt

@@ -18,6 +18,7 @@ controlnet-aux>=0.0.10
 opencv-python-headless>=5.0.0.93
 kornia>=0.8.3
 kernels>=0.17.2
+comfy-kitchen<0.3,>=0.2.38
 spandrel>=0.4.2
 facexlib>=0.3.0
 numpy>=2.5.4

diffusers_workflow-0.15.0/dw/acceleration_warnings.py

@@ -0,0 +1,311 @@
+"""Acceleration settings that validate passes and that cannot pay off (#868).
+
+Each of these cost a render to find out in Don's `h3-quality` matrix:
+
+- `cache_short_schedule`: a threshold cache (`first_block`, `mag`) on a
+  schedule under `CACHE_MIN_USEFUL_STEPS`. Consecutive steps of a short
+  schedule differ too much for the residual threshold, so the cache skips
+  nothing and still perturbs the result - on MiniMax-H3 a `first_block` cache
+  at 0.05 skipped no step at 12 steps and took 22% off at 30. `docs/RECIPES_24GB.md` states the
+  rule and `tests/test_catalog_speed_defaults.py` holds the catalog to it;
+  this is where a hand-written workflow meets it. The interval caches
+  (`taylorseer`, `faster`) skip on a fixed interval and `text_kv` skips no
+  step at all, so none of them is held to it.
+- `taylorseer_state_unestimated`: TaylorSeer keeps per-block state on the
+  device that no VRAM estimate charges - P5 ran out of memory two minutes
+  into a run validate had passed. Always said when TaylorSeer is set. Sizing
+  that state would need the architecture data the engine never holds, and
+  charging it in `vram_estimate` would turn a warning into a refusal and move
+  jobs between cards, so the warning only states the estimate it sits on top
+  of.
+- `cache_unsupported`: a cache on a pipeline whose class declares no
+  transformer (a UNet pipeline such as SD 1.5), or whose transformer class has
+  no `enable_cache()` (Z-Image's, in the diffusers this server runs). Read from
+  the pipeline class's constructor alone, nothing is loaded; the run only says
+  "not applied" and carries on uncached.
+- `compile_fullgraph_with_cache`: `fullgraph` forbids the graph break a
+  cache's skip decision needs (`docs/ACCELERATION.md`).
+- `attention_backend_unavailable`: a CUDA-only attention backend on a device
+  that isn't CUDA passes validate and fails at load.
+- `sparse_attention_fallback`: the `sparse` attention backend on a device that
+  isn't CUDA. It never fails a run: it falls back to dense attention, so this
+  says the speedup is not coming, not that the job will break.
+
+Every one is a warning and never a refusal. A value that cannot be resolved
+before the run (a null variable, an unknown device) leaves its warning
+silent rather than guessed at.
+"""
+
+import inspect
+import typing
+
+from . import get_device_type
+from .introspection import load_pipeline_class
+from .references import iter_steps, render_path
+from . import sparse_attention
+from .sparse_attention import SPARSE_BACKEND
+from .vram_estimate import required_vram_gb
+
+# The fewest denoise steps a threshold cache is expected to skip any of. A
+# general fact about residual-threshold caching, not a per-model rule: H3
+# skipped nothing at 12 steps and 22% at 30 (RECIPES_24GB.md); nothing
+# between has been measured yet (#875 measures 20)
+CACHE_MIN_USEFUL_STEPS = 20
+
+# The cache types that decide each step's skip on a residual threshold
+THRESHOLD_CACHES = ("first_block", "mag")
+
+# The attention backends diffusers registers that need CUDA: the flash,
+# flash-3, sage and xformers families, Hub kernels included. Matched by name,
+# never by a `_hub` suffix; no `native`-family backend is here. A name
+# missing from this set only means no warning - today's behavior
+CUDA_ONLY_ATTENTION_BACKENDS = frozenset(
+    {
+        "flash",
+        "flash_hub",
+        "flash_varlen",
+        "flash_varlen_hub",
+        "flash_4_hub",
+        "_flash_3",
+        "_flash_varlen_3",
+        "_flash_3_hub",
+        "_flash_3_varlen_hub",
+        "aiter_fa2_hub",
+        "sage",
+        "sage_hub",
+        "sage_blackwell_hub",
+        "sage_varlen",
+        "_sage_qk_int8_pv_fp8_cuda",
+        "_sage_qk_int8_pv_fp8_cuda_sm90",
+        "_sage_qk_int8_pv_fp16_cuda",
+        "_sage_qk_int8_pv_fp16_triton",
+        "xformers",
+    }
+)
+
+SHORT_SCHEDULE = "cache_short_schedule"
+TAYLORSEER_STATE = "taylorseer_state_unestimated"
+FULLGRAPH_WITH_CACHE = "compile_fullgraph_with_cache"
+BACKEND_UNAVAILABLE = "attention_backend_unavailable"
+SPARSE_FALLBACK = "sparse_attention_fallback"
+CACHE_UNSUPPORTED = "cache_unsupported"
+
+# The component names a cache hook attaches to, as in
+# dw.pipeline_processors.components.get_cache_transformer
+_CACHE_COMPONENTS = ("transformer", "transformer_ref")
+
+
+def _steps(pipeline):
+    """The resolved `num_inference_steps` the pipeline is called with, or
+    None when it is not a number known before the run."""
+    arguments = pipeline.get("arguments")
+    value = (arguments if isinstance(arguments, dict) else {}).get(
+        "num_inference_steps"
+    )
+    if isinstance(value, bool) or not isinstance(value, (int, float)):
+        return None
+    return value
+
+
+def _backend(device_type):
+    """The backend `device_type` belongs to ('cuda' for 'cuda:1'), or None
+    when the device is unknown."""
+    if not isinstance(device_type, str) or not device_type:
+        return None
+    try:
+        return get_device_type(device_type)
+    except (RuntimeError, ValueError):
+        return None
+
+
+def _components(configuration):
+    components = configuration.get("components")
+    if not isinstance(components, dict):
+        return []
+    return [
+        (name, component)
+        for name, component in components.items()
+        if isinstance(component, dict)
+    ]
+
+
+def _estimate_clause(context):
+    """'; this job is estimated at X of Y GB' when the workflow declares a
+    VRAM estimate and the device's capacity is known, else ''."""
+    if context.capacity_gb is None:
+        return ""
+    required = required_vram_gb(
+        context.expanded,
+        context.arguments,
+        device_type=context.device_type,
+        base_dir=context.base_dir,
+        probe=context.probe,
+    )
+    if required is None or not required[1]:
+        # Only a declared vram_estimate is an estimate; a measured `cost`
+        # entry records a card, not what this job needs
+        return ""
+    return (
+        f"; this job is estimated at {required[0]:.1f} of "
+        f"{context.capacity_gb:.1f} GB, and that state comes on top"
+    )
+
+
+def _cache_unsupported(component_type):
+    """Why a cache cannot attach to the pipeline class `component_type`
+    names, or None when it can or that cannot be told from the class alone
+    (an unresolvable class, a **kwargs constructor, a component whose type is
+    not a single class)."""
+    if not isinstance(component_type, str):
+        return None
+    try:
+        cls = load_pipeline_class(component_type)
+        parameters = inspect.signature(cls.__init__).parameters
+    except (ValueError, TypeError):
+        return None
+    if any(p.kind == inspect.Parameter.VAR_KEYWORD for p in parameters.values()):
+        return None
+    names = [name for name in _CACHE_COMPONENTS if name in parameters]
+    if not names:
+        return (
+            f"{cls.__name__} has no transformer component to attach a cache to "
+            f"(a UNet pipeline is not supported)"
+        )
+    try:
+        hints = typing.get_type_hints(cls.__init__)
+    except Exception:
+        return None
+    classes = [hints.get(name) for name in names]
+    if all(isinstance(c, type) and not hasattr(c, "enable_cache") for c in classes):
+        return (
+            f"{cls.__name__}'s transformer ({classes[0].__name__}) does not "
+            f"support diffusers' enable_cache()"
+        )
+    return None
+
+
+def _cache_warnings(cache, pipeline, path, where, context, component_type=None):
+    cache_type = cache.get("type")
+    warnings = []
+    steps = _steps(pipeline)
+    if (
+        cache_type in THRESHOLD_CACHES
+        and steps is not None
+        and steps < CACHE_MIN_USEFUL_STEPS
+    ):
+        warnings.append(
+            f"{path}: {SHORT_SCHEDULE}: a '{cache_type}' cache on {steps:g} "
+            f"steps will likely skip nothing (consecutive steps differ too much "
+            f"for its threshold) and still perturbs the result{where}; drop the "
+            f"cache or run >= {CACHE_MIN_USEFUL_STEPS} steps"
+        )
+    reason = _cache_unsupported(component_type)
+    if reason:
+        warnings.append(
+            f"{path}: {CACHE_UNSUPPORTED}: the '{cache_type}' cache will not be "
+            f"applied: {reason}{where}; the step runs uncached, so drop the cache"
+        )
+    if cache_type == "taylorseer":
+        max_order = cache.get("max_order")
+        if isinstance(max_order, bool) or not isinstance(max_order, int):
+            max_order = 1
+        warnings.append(
+            f"{path}: {TAYLORSEER_STATE}: TaylorSeer keeps per-block state on "
+            f"the GPU (more with a higher max_order; this one is {max_order}) "
+            f"that the VRAM estimate doesn't include{_estimate_clause(context)}"
+            f"{where}. A run that fits without the cache can run out of memory "
+            f"with it"
+        )
+    return warnings
+
+
+def acceleration_warnings(context):
+    """Every acceleration setting in the expanded definition that cannot pay
+    off, as `"<path>: <code>: <message>"` strings, each said once."""
+    warnings = []
+    backend = _backend(context.device_type)
+    for _, _, pipeline, source, where in iter_steps(
+        context.expanded.get("steps"), context.source_indices, "pipeline"
+    ):
+        configuration = pipeline.get("configuration")
+        if not isinstance(configuration, dict):
+            continue
+        base = ("steps", source, "pipeline", "configuration")
+        cache = configuration.get("cache")
+        if isinstance(cache, dict):
+            warnings += _cache_warnings(
+                cache,
+                pipeline,
+                render_path(base + ("cache",)),
+                where,
+                context,
+                configuration.get("component_type"),
+            )
+
+        backends = [(base + ("attention_backend",), configuration)]
+        for name, component in _components(configuration):
+            component_path = base + ("components", name)
+            compile_config = component.get("compile")
+            if (
+                isinstance(cache, dict)
+                and isinstance(compile_config, dict)
+                and compile_config.get("fullgraph") is True
+            ):
+                warnings.append(
+                    f"{render_path(component_path + ('compile', 'fullgraph'))}: "
+                    f"{FULLGRAPH_WITH_CACHE}: fullgraph forbids the graph break "
+                    f"the '{cache.get('type')}' cache's skip decision needs"
+                    f"{where}; use partial graph mode (fullgraph false) with a "
+                    f"cache"
+                )
+            backends.append((component_path + ("attention_backend",), component))
+
+        if backend is None or backend == "cuda":
+            continue
+        for path, block in backends:
+            name = block.get("attention_backend")
+            if name == SPARSE_BACKEND:
+                warnings.append(
+                    f"{render_path(path)}: {SPARSE_FALLBACK}: sparse attention "
+                    f"needs CUDA; this server is '{context.device_type}', so "
+                    f"the run uses dense attention{where}"
+                )
+                continue
+            if name not in CUDA_ONLY_ATTENTION_BACKENDS:
+                continue
+            warnings.append(
+                f"{render_path(path)}: {BACKEND_UNAVAILABLE}: attention backend "
+                f"'{name}' needs CUDA; this server is '{context.device_type}', "
+                f"so the job will fail at load{where}; leave it unset or use a "
+                f"'native'-family backend"
+            )
+    # A for_each step's members share their written path: say each code once
+    # there, naming the first member it fired for
+    said = {}
+    for warning in warnings:
+        path, _, rest = warning.partition(": ")
+        said.setdefault((path, rest.partition(": ")[0]), warning)
+    return list(said.values())
+
+
+def available_attention_backends(device_type: str) -> list[str]:
+    """The attention backends a server on `device_type` can run, in menu order.
+
+    `native` first, then the other `native`-family names; on CUDA, `sparse`
+    (only where the compiled kernels loaded) and the CUDA-only names. A name
+    diffusers' `AttentionBackendName` lacks is dropped, so the menu never
+    offers one the pipeline would refuse.
+    """
+    from diffusers.models.attention_dispatch import AttentionBackendName
+
+    known = {backend.value for backend in AttentionBackendName}
+    cuda = device_type == "cuda"
+    names = ["native", "_native_flash", "_native_efficient", "_native_math"]
+    if cuda:
+        names.append("_native_cudnn")
+    names.append("flex")
+    if cuda:
+        if sparse_attention._kernels_importable():
+            names.append(SPARSE_BACKEND)
+        names.extend(sorted(CUDA_ONLY_ATTENTION_BACKENDS))
+    return [name for name in names if name == SPARSE_BACKEND or name in known]