docs: update Stable Diffusion 1.5 links

2026-05-26 20:53:47 +02:00
31 changed files with 2622 additions and 1707 deletions
@@ -0,0 +1,23 @@
 name: Tier 1 — Smoke (macOS self-hosted)
 # macOS smoke tests run on the self-hosted Apple Silicon runner instead of
 # GitHub-hosted macOS (10x minute multiplier), which exhausts the included
 # Actions minutes too quickly.
 on:
  push:
    branches: [main]
  pull_request:
 jobs:
  smoke:
    runs-on: [self-hosted, macOS, ARM64, coreml]
    timeout-minutes: 20
    steps:
      - uses: actions/checkout@v4
      # The self-hosted runner provides uv; no setup-uv action needed.
      - name: uv sync
        run: uv sync --no-install-project
      - name: Run Tier 1 (synthetic micro-UNet smoke)
        run: uv run pytest -m smoke tests/ -v
@@ -3,4 +3,3 @@ __pycache__/
 models/
 .venv/
 test_results/
 .claude/
@@ -1 +0,0 @@
 3.12
@@ -1,618 +0,0 @@
 # ComfyUI-CoreMLSuite — Converter Extraction Spec for Claude Code
 > **Companion to `MODERNIZATION_SPEC.md`.** That spec hardens the repo and (Phase 3)
 > splits the *inference* math from the framework. **This** spec splits the *conversion*
 > path (`safetensors → CoreML`) out into a standalone, `comfy`-free, pip-installable
 > package that CoreMLSuite then depends on — and that other projects (incl. on-device
 > iOS tooling) can reuse.
 >
 > **Same discipline as the modernization spec:** safety-net first, behavior-preserving
 > until told otherwise, one phase = one branch = one PR, `STOP — VALIDATE` gate between
 > every phase, golden-latent as the regression anchor. `[M2]` = needs macOS/Apple Silicon;
 > `[M2-ANE]` = needs the Neural Engine. Everything else must run on plain Linux/CI.
 ---
 ## 0. How to work (read first — non-negotiable)
 1. **Behavior-preserving until Phase E6.** Phases E1–E5 must not change image output, node
   names, `INPUT_TYPES` field names, or `NODE_CLASS_MAPPINGS` keys. The node graph is the
   public contract; saved user-workflow JSON breaks if these change.
 2. **The conversion package produces an artifact and stops there.** Its job ends at a written
   `.mlpackage` / `.mlmodelc` on disk. It must NOT import `comfy`, `folder_paths`, or
   `comfy_extras`, and must NOT know ComfyUI's `models/unet` layout. Paths are *inputs*.
 3. **The runtime loader stays in the suite.** The loader is the **local** `coreml_suite.coreml_model.CoreMLModel`
   — a thin wrapper over `coremltools.models.MLModel` (NOT Apple's
   `python_coreml_stable_diffusion.coreml_model.CoreMLModel`, which is no longer used; see #58).
   It *runs* a compiled model in Python — a desktop/Python inference concern, not a conversion
   concern. It is NOT moved into the package. (On iOS the `.mlmodelc` is loaded natively; the
   package's output is the deliverable, not a Python runner.)
 4. **Decouple in-repo before splitting repos.** Phases E1–E4 create the package *inside this
   repo* and prove equivalence. The physical second-repo split is Phase E5, only after the
   golden latent is proven identical. Do not create a second repository before Gate E4 passes.
 5. **Reuse the existing regression anchor.** The golden latent / PSNR anchor from
   `MODERNIZATION_SPEC.md` Phase 2 is the cross-cutting proof for every gate here. If it is not
   yet captured, capture it first (it is a prerequisite for E2 onward).
 6. **No new runtime dependencies** without flagging in the gate report (name, why, license, size).
 7. **A failing gate means stop and report**, not work around into the next phase.
 8. **Tooling is `uv`, not bare `pip`/`venv`.** Every environment/install/lock step uses the
   project's `uv` toolchain: `uv venv`, `uv pip install`, `uv pip install -e .`, `uv lock`,
   `uv run pytest`, `uv export`/`uv pip freeze` for baselines. Where this spec says "fresh venv",
   read "`uv venv` + `uv pip install`". Reserve `uv pip` (not `pip`) inside that venv too.
 9. **The package is the single source of truth for *what is possible*; the node is a thin,
   discovery-driven frontend.** See the "Interface contract" pillar below — this is the
   maintainer's hard requirement and it overrides the earlier (now-rescinded) "freeze the
   dropdown list" instruction.
 ---
 ## Interface contract (the maintainer's hard requirement) — read before any phase
 Two coupled guarantees must hold once the package is split out:
 **(A) Updating the converter must NOT require updating CoreMLSuite.**
 This is satisfied by treating the package's public surface as a versioned contract:
 - `convert(...)` and `compile_model(...)` are **keyword-only with defaults** for everything
  past the genuinely-required positionals (`ckpt_path`, `model_version`, `out_path`). New
  capabilities are added as new keyword args with defaults, so an old Suite's call still
  validates against a newer package. **Never** reorder or rename existing parameters.
 - `compose_out_name` (the `.mlpackage` filename = the cache key) **moves into the package** and
  is versioned with it. The Suite must not carry its own copy; if the package changes the naming
  scheme that is a **major** bump (old cached artifacts stop resolving).
 **(B) CoreMLSuite must be able to list *new* conversion types WITHOUT a Suite code change or
 version bump.** Today the node hardcodes its dropdowns:
 ```python
 "model_version": ([ModelVersion.SD15.name, ModelVersion.SDXL.name],),  # hand-typed, also INCOMPLETE (no LCM / SDXL_REFINER)
 "attention_implementation": (list(ATTENTION_IMPLEMENTATIONS),),         # from coreml_suite.attention
 "quantize_nbits": (list(QUANT_NBITS_VALUES), {"default": "none"}),      # from coreml_suite.core.naming
 ```
 These are replaced by **runtime discovery calls into the package**, evaluated inside
 `INPUT_TYPES` (ComfyUI re-evaluates `INPUT_TYPES` on every plugin load):
 ```python
 import coreml_diffusion
 "model_version":            (coreml_diffusion.list_model_versions(),),
 "attention_implementation": (coreml_diffusion.list_attention_impls(),),
 "quantize_nbits":           (coreml_diffusion.list_quant_modes(), {"default": "none"}),
 ```
 Effect: `uv pip install -U coreml_diffusion` + ComfyUI restart surfaces any newly-added type in the old
 plugin's dropdown — **no Suite edit, no Suite version bump.** This is the requirement.
 **The cost, stated honestly (accept this trade-off explicitly at Gate E0):**
 - The Suite becomes a "dumb" frontend; the package is the sole authority on what conversions
  exist. The Suite can no longer guarantee its saved workflows are valid against *arbitrary*
  future package versions.
 - Therefore the package's discovery identifiers (`ModelVersion` values, attn-impl strings, quant
  modes) are an **ADDITIVE-ONLY contract**: the package may *add* identifiers freely (minor bump,
  no Suite change); **removing or renaming an identifier is a breaking change requiring a MAJOR
  bump and a migration note**, because a saved workflow JSON references these strings verbatim.
  Without this rule, "no version bump" silently becomes "randomly broken workflows."
 - `INPUT_TYPES` must **fail soft** when the package is missing/old: wrap the discovery calls so a
  missing `coreml_diffusion` (or an old one lacking a `list_*` function) yields a sane fallback list and a
  logged warning, instead of the node failing to register and disappearing from the menu.
 **Discovery API the package must expose (stable names):**
 ```python
 coreml_diffusion.list_model_versions() -> list[str]   # VERIFIED ones only, e.g. ["SD15","SDXL"] today (.name — see seam.md)
 coreml_diffusion.list_attention_impls() -> list[str]  # ["SPLIT_EINSUM","SPLIT_EINSUM_V2","ORIGINAL"]
 coreml_diffusion.list_quant_modes() -> list[str]      # ["none","8","6","4"]
 coreml_diffusion.CONTRACT_VERSION: str                # bump rules above; Suite may log/compare it
 ```
 These return the *display strings already used today*, so existing workflows keep validating.
 **Verification status is a PACKAGE property, not a node hardcode (maintainer's intent).**
 The Suite wants to expose *every model the converter can verifiably convert*. Today `lcm` and
 `sdxl_refiner` are absent from the converter node not because the Suite chooses to hide them, but
 because they lack a full golden/PSNR verification. So the gating lives in the package as a status:
 ```python
 from enum import Enum
 class Status(Enum):
    VERIFIED = "verified"          # has a golden anchor + passing [M2-ANE] check
    EXPERIMENTAL = "experimental"  # convertible but not yet anchored/verified
 # internal registry, single source of truth.
 # KEY by ModelVersion enum MEMBER so list_* can emit .name. Keying by the lowercase
 # .value string returns ["sd15",...], which the node reverses via ModelVersion[...] -> KeyError.
 _MODEL_STATUS = {ModelVersion.SD15: Status.VERIFIED, ModelVersion.SDXL: Status.VERIFIED,
                 ModelVersion.SDXL_REFINER: Status.EXPERIMENTAL, ModelVersion.LCM: Status.EXPERIMENTAL}
 def list_model_versions(include_experimental: bool = False) -> list[str]:
    return [v.name for v, s in _MODEL_STATUS.items()   # .name -> "SD15","SDXL"; node reverses with ModelVersion[...]
            if s is Status.VERIFIED or (include_experimental and s is Status.EXPERIMENTAL)]
 ```
 Consequence: **promoting a model to VERIFIED in the package expands the Suite's dropdown with no
 Suite change and no Suite bump** — exactly the requirement. The act of verification (E-LCM
 produces an LCM golden anchor; same later for refiner) is what flips the status. The Suite's
 converter node calls `list_model_versions()` (verified-only); a power-user/CLI path may pass
 `include_experimental=True`. Promotion VERIFIED-from-EXPERIMENTAL is additive (minor bump);
 demotion or removal is breaking (major bump + note).
 ---
 ## Naming & layout (chosen — frozen at Gate E0)
 **Distribution name (PyPI):** `coreml-diffusion`. **Import name (Python):** `coreml_diffusion`.
 (PyPI normalizes `-`/`_`; the distribution uses the hyphen, the importable module the underscore.)
 Availability checked: both `coreml-diffusion` and the near variants were free on PyPI at E0.
 **Why this name (the positioning it encodes):** the project's niche is *diffusion models on Apple
 Neural Engine via CoreML, inside ComfyUI and on-device* — **not** Stable Diffusion specifically.
 `sd*` was rejected because it falsely narrows scope to SD; `coreml-diffusion` keeps `coreml` on the
 front for discoverability while `diffusion` honestly states the scope (SD/SDXL/LCM today, Flux and
 other diffusion architectures later) **without** promising arbitrary non-diffusion torch models,
 whose tracing/shape/sample-input pipeline differs. The name must not be re-narrowed to SD in
 future docs. ANE is the *differentiator* (documented in the README), but `coreml` was chosen over
 `ane` in the name for search discoverability per maintainer decision.
 Target package layout (framework-free — zero `comfy` imports):
 ```
 coreml_diffusion/
  __init__.py        # public API surface (see "Public API" below)
  model_version.py   # ModelVersion enum — the SINGLE source of truth, no comfy
  attention.py       # ATTENTION_IMPLEMENTATIONS tuple (from coreml_suite/attention.py) + apply_attention_implementation
  pipeline.py        # get_pipeline (from_single_file), get_unet (cml UNet from ref unet)
  unet.py            # UNet2DConditionModelLCM (moved from coreml_suite/lcm/unet.py)
  inputs.py          # get_sample_input, lcm_inputs, sdxl_inputs,
                     #   get_encoder_hidden_states_shape, get_coreml_inputs, get_inputs_spec
  controlnet.py      # add_cnet_support (conversion-side residual SHAPE calc only)
  convert.py         # convert_unet, convert (orchestration), convert_to_coreml, load_coreml_model
  compile.py         # compile_coreml_model
  quantize.py        # (Phase E6 / MODERNIZATION Phase 6 lands here) palettization 4/6/8-bit
  cli.py             # console entry point: `coreml-diffusion convert ...`
  pyproject.toml     # standalone packaging (at E5)
 ```
 What stays in `coreml_suite/` (the ComfyUI side, thinned):
 - `nodes.py` — still owns **name-encoding** (`out_name` construction), path resolution via
  `folder_paths`, the node `INPUT_TYPES`/mappings, and wrapping the result in `CoreMLModel`.
 - `models.py`, `latents.py`, `controlnet.py` (inference parts), `lcm/utils.py`, `config.py`
  (inference config build) — untouched by this spec except the import-source of `ModelVersion`.
 ### Public API (the contract `coreml_diffusion` exposes)
 ```python
 from coreml_diffusion import ModelVersion, convert, compile_model, compose_out_name
 from coreml_diffusion import list_model_versions, list_attention_impls, list_quant_modes, CONTRACT_VERSION
 # Mirror the CURRENT converter.py signature, made keyword-only past the required positionals
 # and with paths/device injected (no folder_paths, no comfy.model_management):
 # convert(ckpt_path, model_version, out_path, *,
 #         batch_size=1, sample_size=(64, 64), controlnet_support=False,
 #         lora_weights=None, attn_impl=list_attention_impls()[0], config_path=None,
 #         quantize_nbits="none", device=None) -> None   # side effect: writes out_path
 #   (current convert() returns None and writes via convert_unet → coreml_unet.save; keep that,
 #    or change to `return out_path` as a deliberate, documented improvement — pick one at E0.)
 # compile_model(src_path, out_dir, final_name) -> str   # returns compiled .mlmodelc path
 ```
 Note: `convert` takes an **explicit `out_path`** — no `folder_paths`. `device` is injected
 (defaults to torch's default device). `compose_out_name` lives here (cache-key contract) and the
 node imports it from the package. The `list_*` discovery functions back the node's dropdowns.
 ---
 ## The import chains to cut (root cause inventory) — REVISED against current code
 > **State note (verified):** the code moved on since the original draft. Several chains are
 > already cut. Re-verify each line by `grep` before acting; do not assume the original draft.
 **Already done (verify, then skip):**
 - ✅ `converter.py` already imports `from coreml_suite.model_version import ModelVersion`, and
  `model_version.py` is **clean** (`from enum import Enum` only — zero comfy). The old
  "converter → config → comfy" chain is **already broken**. `config.py` still imports comfy, but
  it is **inference-side** (`get_model_config` via `supported_models_base`/`latent_formats`) —
  *not* on the conversion path. Do **not** treat `config.py` as a converter dependency.
 - ✅ `converter.py` now uses `diffusers.UNet2DConditionModel.from_single_file` and a local
  `CoreMLUNetWrapper` (in `coreml_suite/conversion/unet.py`) — it is **no longer** importing the
  Apple `python_coreml_stable_diffusion.unet.UNet2DConditionModel*` internals on the main path.
  A `coreml_suite/conversion/` subpackage already exists (`attention`, `shapes`, `trace`, `unet`).
 - ✅ Name-encoding already extracted to `coreml_suite/core/naming.py` (`compose_out_name`,
  `lora_names_from_params`, `ATTN_SUFFIX`, `QUANT_NBITS_VALUES`) **with characterization tests**
  (`tests/unit/test_characterization_out_name.py`). The pure-naming split is done.
 - ✅ Quantization is **already implemented** in `converter.py` (`quantize_nbits`, k-means
  `palettize_weights`) and surfaced as an optional node input. Phase E6 is therefore *move*, not
  *build* (see revised E6).
 **Still to cut (the real remaining work):**
 1. `coreml_suite/converter.py::get_out_path` → `from folder_paths import get_folder_paths`.
   Main converter still reaches into ComfyUI's model dir. **Cut: `out_path` is an injected arg;
   `folder_paths` resolution moves up into the node** (the node already computes `out_name`).
 2. `coreml_suite/lcm/converter.py` → still has its **own** `from folder_paths import
   get_folder_paths` (`get_out_path`) and (per original draft) `comfy.model_management`. Verify
   the current LCM file and cut both: inject `out_path` and `device`.
 3. Global mutation of the attention impl: confirm where it now lives. Main path appears to route
   through `coreml_suite/conversion/attention.apply_attention_implementation` (cleaner than the
   old global), but `lcm/converter.py` may still set a module global at import. **Ensure the
   package sets attention per-call, never at import time.**
 4. **Duplication LCM vs main:** `lcm/converter.py` still carries its own copies of
   `convert_to_coreml`, `load_coreml_model`, `get_out_path`, `get_sample_input` (the LCM variant
   takes a `scheduler` arg), and hardcodes `SimianLuo/LCM_Dreamshaper_v7`. **Dedupe into the
   single `coreml_diffusion` implementation;** the HF-hardcode consolidation is the *behavior-changing*
   part → deferred to optional **E-LCM**, not E1–E5.
 5. **`compose_out_name` ownership:** currently in `coreml_suite/core/naming.py` and called by the
   node. Per the Interface-contract pillar it must **move into the package** (it is the cache-key
   contract) and the node must import it from `coreml_diffusion`, not keep a copy.
 ---
 ## Phase E0 — Seam decision & inventory (no code change)
 **Objective:** lock the cut line, the interface contract, and naming so later phases don't drift.
 ### Tasks
 1. Produce `docs/extraction/seam.md`: a table of every symbol in `converter.py`,
   `lcm/converter.py`, `lcm/unet.py`, **plus the already-extracted `conversion/` subpackage
   (`attention`, `shapes`, `trace`, `unet`) and `core/naming.py`**, classified
   **CONVERSION → coreml_diffusion** vs **STAYS (comfy/node)**. Note which are already framework-free.
 2. ~~Confirm the current `python_coreml_stable_diffusion` footprint.~~ **DONE (seam.md §6):
   footprint is ZERO** — no runtime imports anywhere; only a docstring mention in
   `core/__init__.py:4`. Main path uses `diffusers` + local `CoreMLUNetWrapper`; the runtime
   `CoreMLModel` (STAYS in suite) is a local coremltools wrapper, not Apple's. No shape/attn helper
   comes from Apple (local `conversion/shapes.py`, `conversion/attention.py`).
 3. **Decide the interface contract concretely (the maintainer's hard requirement):**
   - Discovery functions `list_model_versions / list_attention_impls / list_quant_modes` live in
     the package and return today's display strings verbatim. Node `INPUT_TYPES` calls them.
   - `ModelVersion` values, attn-impl strings, quant modes are **ADDITIVE-ONLY** across package
     versions; removal/rename = MAJOR bump + migration note. Write this into the package's
     versioning policy doc now.
   - `compose_out_name` moves to the package; node imports it (no copy). Confirm the
     characterization tests in `test_characterization_out_name.py` will be re-pointed, not
     duplicated.
   - **Resolve the `model_version` dropdown question (maintainer decided):** the Suite exposes
     *every model the converter can verifiably convert*. `lcm` and `sdxl_refiner` are absent today
     only because they lack a golden/PSNR verification — **not** because the node hardcodes a
     short list. Encode this as a **status registry in the package** (`VERIFIED` vs
     `EXPERIMENTAL`); `list_model_versions()` returns VERIFIED-only by default. The converter node
     calls it plainly. Promoting LCM/refiner to VERIFIED (after E-LCM / a refiner anchor) expands
     the dropdown with **no Suite change**. Do NOT add permanent per-node filtering — the gate is
     verification status, owned by the package.
 4. ~~Confirm the `ml-stable-diffusion` git dep is pinned.~~ **N/A — already removed (#58).** Verified:
   zero `python_coreml_stable_diffusion` imports in the repo; `CoreMLModel` is now a local
   coremltools wrapper; the dep is absent from `pyproject.toml`/`requirements.txt`. No SHA to pin.
 ### STOP — VALIDATE (Gate E0)
 ```
 ## Gate E0 report
 - seam.md committed: <path>; symbol counts (move / stay / already-framework-free)
 - python_coreml_stable_diffusion usage (verified by grep): conversion=<list> runtime=<list>
 - Discovery API signatures frozen: list_model_versions (verified-only) / list_attention_impls / list_quant_modes
 - Status registry decided: sd15+sdxl=VERIFIED, lcm+sdxl_refiner=EXPERIMENTAL (gated, not hidden)
 - Additive-only contract policy doc written (incl. promotion=minor, demotion/removal=major): <path>
 - model_version dropdown: expose all (incl. LCM/REFINER) / filtered per node — DECISION: <...>
 - compose_out_name move-not-copy confirmed; tests re-point plan: <...>
 - LCM consolidation deferred to optional E-LCM: YES/NO
 - ml-stable-diffusion: N/A — already removed (#58), not a dependency (was: pin-or-BLOCKER)
 - Package name in-repo: coreml_diffusion (final PyPI name deferred to E5)
 ```
 ---
 ## Phase E1 — Establish `coreml_diffusion` package + discovery API (mostly verification)
 **Objective:** stand up the package namespace and the discovery surface. Much of the comfy-chain
 cut is **already done** — this phase mostly *verifies* that and adds the discovery functions.
 ### Tasks
 1. **Verify (don't redo):** `coreml_suite/model_version.py` is already clean (`Enum` only). Confirm
   `import coreml_suite.model_version` works with **no comfy** (`uv run python -c "..."` in a
   comfy-free `uv venv`). If true, E1's original "extract ModelVersion" task is already satisfied.
 2. Create the `coreml_diffusion/` package skeleton with `__init__.py` exporting the **discovery API**
   backed by the *existing* sources of truth for now (re-export `ModelVersion`, the
   `ATTENTION_IMPLEMENTATIONS` tuple, and `QUANT_NBITS_VALUES`) so values are byte-identical:
   ```python
   def list_model_versions(): return [v.name for v in ModelVersion]   # .name -> "SD15" (node reverses via ModelVersion[...]; .value KeyErrors)
   def list_attention_impls(): return list(ATTENTION_IMPLEMENTATIONS)
   def list_quant_modes(): return list(QUANT_NBITS_VALUES)
   CONTRACT_VERSION = "1.0"
   ```
   (At this stage `coreml_diffusion` may live inside the repo and import from `coreml_suite.*`; the
   physical move of implementation happens in E2. The point of E1 is to freeze the *contract*.)
 3. **Decided (`.name`):** the node renders `ModelVersion.SD15.name` (`"SD15"`) and reverses the
   dropdown string via `ModelVersion[model_version]` (name lookup, `nodes.py:286`). Discovery API
   therefore returns `.name`; `.value` (`"sd15"`) would `KeyError`. Recorded in `seam.md` §5.
 ### Acceptance criteria
 - `uv run python -c "import coreml_diffusion; print(coreml_diffusion.list_model_versions(), coreml_diffusion.list_quant_modes())"`
  works in a **comfy-free** `uv venv` and prints today's exact strings.
 - Existing characterization tests pass unchanged.
 - No node behavior change yet (node still uses its current hardcoded lists in E1).
 ### STOP — VALIDATE (Gate E1)
 ```
 ## Gate E1 report
 - model_version.py confirmed comfy-free (uv, no comfy): PASS/FAIL
 - coreml_diffusion.list_* returns byte-identical strings to current dropdowns: YES/NO (show values)
 - .name vs .value decision for model_version discovery: <...>
 - CONTRACT_VERSION set; additive-only policy linked: <path>
 - Characterization tests unchanged & green (uv run pytest): YES/NO
 ```
 ---
 ## Phase E2 — Move conversion code into `coreml_diffusion` (in-repo, dedup, behavior-preserving)
 **Objective:** physically relocate the conversion mechanics into the framework-free package,
 collapsing the two duplicate converters into one, with paths/device injected.
 ### Tasks
 1. Move into `coreml_diffusion/`: `pipeline.py` (`get_pipeline`, `get_unet`), `unet.py`
   (`UNet2DConditionModelLCM`), `inputs.py` (sample/lcm/sdxl input builders +
   `get_encoder_hidden_states_shape` + `get_coreml_inputs` + `get_inputs_spec`),
   `controlnet.py` (`add_cnet_support`), `convert.py` (`convert_unet`, `convert`,
   `convert_to_coreml`, `load_coreml_model`), `compile.py` (`compile_coreml_model`).
 2. **Dedupe LCM vs main** (the real remaining duplication): delete `lcm/converter.py`'s copies of
   `convert_to_coreml` / `load_coreml_model` / `get_out_path` / `get_sample_input` (LCM variant
   carries a `scheduler` arg — fold that into the shared `get_sample_input` as an optional param)
   in favor of the single `coreml_diffusion` implementation. The main path's helpers
   (`get_unet`/`get_encoder_hidden_states_shape`/`get_coreml_inputs`/`convert_unet`/`convert`) and
   the `conversion/` subpackage (`attention`, `shapes`, `trace`, `unet`) move as-is.
 3. **Inject paths**: replace `get_out_path`'s `folder_paths` reach-in with an injected `out_path`
   argument on `convert(...)`; `folder_paths` resolution moves up into the node (which already
   computes `out_name`). No `folder_paths` import anywhere in `coreml_diffusion`.
 4. **Inject device** where the LCM path used `comfy.model_management` (verify it still does):
   `convert(..., device=None)`, default to torch's default device.
 5. **Attention per-call, never at import:** main path already routes through
   `conversion/attention.apply_attention_implementation` — keep that. If `lcm/converter.py` still
   sets any module global at import, remove it; the package sets attention from the `attn_impl`
   arg inside `convert`.
 6. **Move `compose_out_name` into the package** (`coreml_diffusion/naming.py`); re-point
   `test_characterization_out_name.py` imports to `coreml_diffusion.naming` — assertions and values
   unchanged. The node will import it from the package in E3.
 7. Leave **thin shims** in `coreml_suite/converter.py` and `coreml_suite/lcm/converter.py` that
   re-export from `coreml_diffusion`, preserving the old call signatures the nodes use (nodes untouched
   this phase). Shims map comfy `folder_paths`/device into package args.
 ### Acceptance criteria
 - `uv run pytest -m unit` (Tier 0) imports `coreml_diffusion.*` with **no comfy / no MPS** and is green on Linux.
 - The dedup leaves exactly one implementation of each previously-duplicated function.
 - Characterization tests pass unchanged after the `compose_out_name` re-point.
 - `[M2]` A real SD1.5 conversion via the shim still produces a loadable model.
 - `[M2-ANE]` **Golden latent identical / within tolerance** to the MODERNIZATION Phase 2 anchor
  (same seed/prompt) — proves the move + dedup changed nothing.
 ### STOP — VALIDATE (Gate E2 — first regression gate)
 ```
 ## Gate E2 report
 - Tier 0 import of coreml_diffusion without comfy/MPS (uv run): PASS/FAIL
 - LCM/main duplicated funcs collapsed to one (list old→new): <map>
 - compose_out_name moved to package; char-tests re-pointed & green: YES/NO
 - Paths injected (no folder_paths in package): confirmed
 - Device injected (no comfy.model_management in package): confirmed
 - Attention set per-call, not at import (both main & lcm): confirmed
 - [M2-ANE] Golden latent vs Phase-2 anchor: identical / within tol <x> / DIVERGED (STOP)
 - Node INPUT_TYPES / mappings untouched: confirmed (diff)
 ```
 **If the golden latent diverged at all, STOP and report — do not continue.**
 ---
 ## Phase E3 — Thin the nodes onto the package (behavior-preserving)
 **Objective:** remove the shims; have the ComfyUI nodes call `coreml_diffusion` directly, keeping the
 node contract byte-identical.
 ### Tasks
 1. `CoreMLConverter.convert` (in `coreml_suite/nodes.py`): keep the `folder_paths`-based path
   resolution **in the node**; import `compose_out_name` from `coreml_diffusion` (not `coreml_suite.core`);
   call `coreml_diffusion.convert(...)` and `coreml_diffusion.compile_model(...)` directly; wrap the compiled path
   in `CoreMLModel`.
 2. **Wire the dropdowns to discovery (the maintainer's hard requirement).** Replace the hardcoded
   `INPUT_TYPES` lists with fail-soft discovery calls:
   ```python
   def _discover(fn, fallback):
       try:
           import coreml_diffusion
           return getattr(coreml_diffusion, fn)()
       except Exception as e:        # missing/old package, or import error
           logger.warning(f"coreml_diffusion.{fn} unavailable ({e}); using fallback {fallback}")
           return fallback
   ...
   "model_version":            (_discover("list_model_versions", ["SD15", "SDXL"]),),
   "attention_implementation": (_discover("list_attention_impls", ["SPLIT_EINSUM","SPLIT_EINSUM_V2","ORIGINAL"]),),
   "quantize_nbits":           (_discover("list_quant_modes", ["none","8","6","4"]), {"default": "none"}),
   ```
   This is what makes "update the package → new types appear in the old node, no Suite bump" true.
 3. `COREML_CONVERT_LCM` (in `coreml_suite/lcm/nodes.py`): route through `coreml_diffusion` for the shared
   mechanics. **Keep the existing LCM behavior/HF-hardcode for now** — consolidation is optional E-LCM.
 4. Delete the now-dead `coreml_suite/converter.py` / `coreml_suite/lcm/converter.py` shims (or
   reduce to a one-line re-export if anything external imports them — grep first).
 ### Acceptance criteria
 - `NODE_CLASS_MAPPINGS` / `NODE_DISPLAY_NAME_MAPPINGS` keys: **unchanged** (diff `__init__.py`).
 - Every `INPUT_TYPES` **field name** unchanged. Dropdown **values**: the discovery calls must
  return **a superset of** today's values, with every previously-present value still present and
  spelled identically (additive-only). *(This deliberately replaces the original spec's
  "values must be byte-identical/frozen" criterion — the maintainer requires the list be
  extensible at runtime. Frozen-field-names + additive-only-values is the new contract.)*
 - With `coreml_diffusion` **absent**, the node still registers and shows the fallback lists (fail-soft).
 - `[M2-ANE]` Golden latent still identical to the Phase-2 anchor.
 - `[M2-ANE]` The committed e2e workflow `tests/integration/...` still passes (PSNR > 25).
 ### STOP — VALIDATE (Gate E3)
 ```
 ## Gate E3 report
 - Node mappings diff: empty (confirmed)
 - INPUT_TYPES field-names diff: empty (confirmed)
 - Dropdown values: superset of prior, all prior values still present & identical: YES/NO (show)
 - Fail-soft with coreml_diffusion absent (node still registers): PASS/FAIL
 - compose_out_name now imported from coreml_diffusion (no node-side copy): confirmed
 - [M2-ANE] Golden latent vs anchor: identical / within tol / DIVERGED (STOP)
 - [M2-ANE] e2e workflow PSNR: <value> (> 25?)
 - Dead converter shims removed / reduced: <list>
 ```
 ---
 ## Phase E4 — Standalone packaging & CLI (still in-repo)
 **Objective:** make `coreml_diffusion` independently installable and usable without ComfyUI, with a CLI
 suitable for the planned article and for on-device/iOS conversion workflows.
 ### Tasks
 1. Add `coreml_diffusion/pyproject.toml`: name (working `coreml_diffusion`), `requires-python`, dependencies
   = `coremltools` (pinned to the MODERNIZATION-validated version), `diffusers`, `transformers`,
   `peft`, `omegaconf`, `numpy`, `torch`. **No `ml-stable-diffusion`** (already removed in #58, see
   §0.3) and **no comfy**. Suite pins `transformers>=4.44`/`peft>=0.13`/`omegaconf>=2.3` today;
   grep-confirm each is on the conversion path before listing it. A `[project.scripts]` entry:
   `coreml-diffusion = "coreml_diffusion.cli:main"`.
 2. `coreml_diffusion/cli.py`: `coreml-diffusion convert --ckpt PATH --model-version sd15 --out PATH
   [--height --width --batch-size --attn-impl --controlnet --lora NAME:STRENGTH ... --config PATH]`
   and `coreml-diffusion compile --src PATH --out-dir DIR --name NAME`. Mirrors `convert()`/`compile_model()`.
 3. Tier-0 Linux tests for the CLI **arg→call mapping** (mock the heavy `convert`); the real
   convert remains `[M2]`. Add a `[M2]` smoke test: convert a tiny synthetic UNet end-to-end.
 4. README for the package: install, CLI usage, "produce a `.mlpackage`/`.mlmodelc` for use in a
   Swift/iOS app", and the ANE positioning note (low-power, GPU-free, embeddable; SD1.5/SDXL on
   ANE, **not** a Flux-speed claim).
 ### Acceptance criteria
 - Fresh `python -m venv` + `uv pip install ./coreml-diffusion` (no ComfyUI present) imports and runs
  `coreml-diffusion --help` and the arg-mapping tests on Linux.
 - `[M2]` `coreml-diffusion convert` produces a model file identical (golden) to the node path.
 ### STOP — VALIDATE (Gate E4)
 ```
 ## Gate E4 report
 - uv pip install ./coreml-diffusion in comfy-free venv: PASS/FAIL (log)
 - CLI arg→call tests (Tier 0, Linux): green
 - [M2] CLI-produced model golden vs node-produced model: identical / DIVERGED
 - Package deps list (with pinned SHAs/versions + licenses):
 - New runtime deps vs suite before: <none / list>
 ```
 **This is the gate that proves the package stands alone. Do not split repos before it passes.**
 ---
 ## Phase E5 — Physical split into a second repository
 **Objective:** move `coreml_diffusion/` to its own repo; CoreMLSuite depends on it by pinned version.
 ### Tasks
 1. Create the new repo (maintainer action — agent prepares the tree, not the GitHub repo).
   Choose final distributable name; rename imports if changed (single sweep, recorded).
 2. CoreMLSuite `pyproject.toml` / `requirements.txt`: replace the conversion-only deps with a
   pinned dependency on the new package (`coreml_diffusion==<version>` from PyPI, or `git+...@<tag>`
   until first PyPI release). (There is no `git+...ml-stable-diffusion` line to remove — already
   gone since #58.)
 3. ~~Keep `python_coreml_stable_diffusion` for the loader.~~ **Void.** The loader is the local
   `coreml_suite/coreml_model.py` over `coremltools`; the suite keeps `coremltools` as a direct dep
   for it. No Apple lib involved.
 4. Set up the new repo's CI: Tier 0 on Linux (import + arg-mapping + input-shape math),
   `[M2]`/`[M2-ANE]` on a self-hosted/macOS-ARM runner reusing the golden-latent anchor.
 5. Versioning: SemVer; first release `0.1.0`. Document the compatibility matrix
   (coreml_diffusion ↔ coremltools version ↔ diffusers version). No ml-stable-diffusion axis.
 ### Acceptance criteria
 - CoreMLSuite installs in a fresh venv pulling the new package; e2e workflow still passes `[M2-ANE]`.
 - New repo CI green on Linux (Tier 0) and `[M2-ANE]` golden latent matches the anchor.
 - No conversion code remains in CoreMLSuite (grep: no `ct.convert`, no `from_single_file`,
  no `torch.jit.trace`).
 ### STOP — VALIDATE (Gate E5)
 ```
 ## Gate E5 report
 - New repo tree prepared: <path/branch>; final package name: <name>
 - Suite depends on package by pinned version: <spec>
 - Suite e2e [M2-ANE] PSNR after split: <value> (> 25?)
 - Conversion code fully absent from suite: confirmed (grep output)
 - Compatibility matrix documented: <link>
 - First release tag: 0.1.0
 ```
 ---
 ## Phase E6 — Quantization travels WITH the conversion code (already implemented → move)
 **Objective:** quantization is **already implemented** (k-means `palettize_weights` in
 `converter.py`, `quantize_nbits` node input, `_q<bits>` filename suffix, README tradeoff table).
 There is nothing to *build*. It simply **moves with the conversion code in E2** as part of
 `convert_unet`. This phase is a checkpoint that it survived the extraction intact, plus exposing
 it through the CLI.
 ### Tasks
 1. Confirm the palettization block moved cleanly into `coreml_diffusion` (lives in `convert.py` or a
   `quantize.py` helper called from `convert_unet`). Default `"none"` stays byte-identical.
 2. Expose via CLI flag `--quantize {none,8,6,4}` (E4 already lists this) and via
   `list_quant_modes()` discovery (E1/E3).
 3. The existing README tradeoff table (SD1.5 1×512×512 SPLIT_EINSUM: none/8/6/4 → size/ms/PSNR)
   moves to the package README. Re-confirm one row `[M2-ANE]` so the article can cite a live number.
 ### Acceptance criteria
 - Default (`none`) output byte-identical to pre-extraction (covered by the E2/E3 golden latent).
 - `coreml-diffusion convert --quantize 4` produces a `_q4` artifact matching the node's `_q4` artifact `[M2]`.
 - `list_quant_modes()` drives the node dropdown (no hardcoded copy remains).
 ### STOP — VALIDATE (Gate E6)
 ```
 ## Gate E6 report
 - Palettization relocated into coreml_diffusion, called from convert_unet: confirmed
 - Default none output identical (golden): YES/NO
 - [M2] CLI --quantize {8,6,4} artifacts match node artifacts: YES/NO
 - Tradeoff table in package README with at least one re-confirmed [M2-ANE] row: <link>
 ```
 ---
 ## Phase E-LCM — FIRST task after the split: clean up LCM + verify → promote (behavior-changing, gated)
 > Promoted from "optional, someday" to **the first thing after E5**, per maintainer intent: the
 > Suite should expose every verifiably-convertible model, and LCM is the obvious first cleanup.
 Two coupled goals:
 1. **Consolidate the LCM path.** Make the LCM node use the unified `from_single_file` path in
   `coreml_diffusion.convert(model_version=LCM, ...)` instead of the hardcoded `SimianLuo/LCM_Dreamshaper_v7`
   HF download; drop the duplicated LCM helpers (already deduped in E2). **Behavior change** ⇒
   capture an LCM golden anchor *before* the change, then prove within-tolerance after.
 2. **Verify → promote.** Once the LCM conversion has a passing `[M2-ANE]` golden anchor, flip
   `_MODEL_STATUS["lcm"] = Status.VERIFIED` **in the package** (minor bump). The Suite's dropdown
   gains `lcm` automatically — no Suite change, no Suite bump. This is the end-to-end proof that
   the discovery contract works as designed.
 Repeat the same recipe for `sdxl_refiner` when it gets an anchor (separate small gate). Do NOT
 bundle E-LCM into E1–E5; it changes behavior and must stand on its own golden.
 ### STOP — VALIDATE (Gate E-LCM)
 ```
 ## Gate E-LCM report
 - LCM golden anchor captured BEFORE change: <path/hash>
 - LCM node now uses unified from_single_file path; HF hardcode removed: confirmed
 - [M2-ANE] LCM golden after change: identical / within tol <x> / DIVERGED (STOP)
 - Status flipped lcm→VERIFIED in package (minor bump <ver>): confirmed
 - Suite dropdown now lists lcm with NO Suite code change / NO Suite bump: confirmed (diff empty)
 - LCM node accepts a checkpoint arg now (documented breaking-ish UI note): <link>
 ```
 ---
 ## Article deliverable (after E4)
 Once the CLI exists and stands alone, the "convert a Comfy/A1111 workflow into an on-device iOS
 app" write-up becomes a clean tutorial: `coreml-diffusion convert` → `.mlmodelc` → load in Swift/CoreML.
 Frame the niche honestly per the README note above (ANE feasibility & power, not raw Flux speed).
 ---
 ## Quick reference: extraction gate discipline
 ```
 E0  Seam decision, interface contract, discovery API frozen → Gate E0  (cut line + additive-only policy?)
 E1  Stand up coreml_diffusion + discovery API (mostly verify)        → Gate E1  (list_* byte-identical, comfy-free?)
 E2  Move conversion code, dedup LCM/main, inject paths/device→ Gate E2  (golden identical? duplicates gone?) ← first regression gate
 E3  Thin nodes onto package + wire discovery dropdowns       → Gate E3  (field-names frozen, values additive, fail-soft, golden identical?)
 E4  Standalone packaging + CLI (uv)                          → Gate E4  (uv pip install w/o comfy? CLI golden?) ← proves it stands alone
 E5  Physical second-repo split                               → Gate E5  (suite depends on pkg? conversion absent?)
 E-LCM  FIRST post-split: clean up LCM, verify → promote      → Gate E-LCM (LCM golden? dropdown gains lcm w/ no Suite bump?)
 E6  Quantization checkpoint (already built → moved in E2)    → Gate E6  (default identical? CLI quant matches?)
 (refiner)  same recipe as E-LCM when an anchor exists        → own small gate (promote sdxl_refiner→VERIFIED)
 ```
 **Interface-contract invariants (the maintainer's hard requirement), restated:**
 - Package API is keyword-only-with-defaults past the required positionals → converter updates
  don't force Suite updates.
 - Node dropdowns are discovery-driven (`coreml_diffusion.list_*`) + fail-soft → new conversion types
  appear in the old plugin with `uv pip install -U coreml_diffusion`, **no Suite code change, no bump**.
 - Discovery identifiers are **additive-only**; removal/rename = MAJOR bump + migration note.
 - `compose_out_name` (cache key) lives in the package, single copy.
 **Golden rule (inherited): never cross a gate with a failing acceptance criterion.
 Stop, report, wait. The golden latent is the single source of truth that the extraction
 changed nothing.**
@@ -1,21 +1,674 @@
-MIT License
+                    GNU GENERAL PUBLIC LICENSE
                       Version 3, 29 June 2007
-Copyright (c) 2023-2026 Adrian Szczepański
+ Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
 Everyone is permitted to copy and distribute verbatim copies
 of this license document, but changing it is not allowed.
-Permission is hereby granted, free of charge, to any person obtaining a copy
+                            Preamble
 of this software and associated documentation files (the "Software"), to deal
 in the Software without restriction, including without limitation the rights
 to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
 copies of the Software, and to permit persons to whom the Software is
 furnished to do so, subject to the following conditions:
-The above copyright notice and this permission notice shall be included in all
+  The GNU General Public License is a free, copyleft license for
-copies or substantial portions of the Software.
+software and other kinds of works.
-THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+  The licenses for most software and other practical works are designed
-IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+to take away your freedom to share and change the works.  By contrast,
-FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+the GNU General Public License is intended to guarantee your freedom to
-AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+share and change all versions of a program--to make sure it remains free
-LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+software for all its users.  We, the Free Software Foundation, use the
-OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+GNU General Public License for most of our software; it applies also to
-SOFTWARE.
+any other work released this way by its authors.  You can apply it to
 your programs, too.
  When we speak of free software, we are referring to freedom, not
 price.  Our General Public Licenses are designed to make sure that you
 have the freedom to distribute copies of free software (and charge for
 them if you wish), that you receive source code or can get it if you
 want it, that you can change the software or use pieces of it in new
 free programs, and that you know you can do these things.
  To protect your rights, we need to prevent others from denying you
 these rights or asking you to surrender the rights.  Therefore, you have
 certain responsibilities if you distribute copies of the software, or if
 you modify it: responsibilities to respect the freedom of others.
  For example, if you distribute copies of such a program, whether
 gratis or for a fee, you must pass on to the recipients the same
 freedoms that you received.  You must make sure that they, too, receive
 or can get the source code.  And you must show them these terms so they
 know their rights.
  Developers that use the GNU GPL protect your rights with two steps:
 (1) assert copyright on the software, and (2) offer you this License
 giving you legal permission to copy, distribute and/or modify it.
  For the developers' and authors' protection, the GPL clearly explains
 that there is no warranty for this free software.  For both users' and
 authors' sake, the GPL requires that modified versions be marked as
 changed, so that their problems will not be attributed erroneously to
 authors of previous versions.
  Some devices are designed to deny users access to install or run
 modified versions of the software inside them, although the manufacturer
 can do so.  This is fundamentally incompatible with the aim of
 protecting users' freedom to change the software.  The systematic
 pattern of such abuse occurs in the area of products for individuals to
 use, which is precisely where it is most unacceptable.  Therefore, we
 have designed this version of the GPL to prohibit the practice for those
 products.  If such problems arise substantially in other domains, we
 stand ready to extend this provision to those domains in future versions
 of the GPL, as needed to protect the freedom of users.
  Finally, every program is threatened constantly by software patents.
 States should not allow patents to restrict development and use of
 software on general-purpose computers, but in those that do, we wish to
 avoid the special danger that patents applied to a free program could
 make it effectively proprietary.  To prevent this, the GPL assures that
 patents cannot be used to render the program non-free.
  The precise terms and conditions for copying, distribution and
 modification follow.
                       TERMS AND CONDITIONS
  0. Definitions.
  "This License" refers to version 3 of the GNU General Public License.
  "Copyright" also means copyright-like laws that apply to other kinds of
 works, such as semiconductor masks.
  "The Program" refers to any copyrightable work licensed under this
 License.  Each licensee is addressed as "you".  "Licensees" and
 "recipients" may be individuals or organizations.
  To "modify" a work means to copy from or adapt all or part of the work
 in a fashion requiring copyright permission, other than the making of an
 exact copy.  The resulting work is called a "modified version" of the
 earlier work or a work "based on" the earlier work.
  A "covered work" means either the unmodified Program or a work based
 on the Program.
  To "propagate" a work means to do anything with it that, without
 permission, would make you directly or secondarily liable for
 infringement under applicable copyright law, except executing it on a
 computer or modifying a private copy.  Propagation includes copying,
 distribution (with or without modification), making available to the
 public, and in some countries other activities as well.
  To "convey" a work means any kind of propagation that enables other
 parties to make or receive copies.  Mere interaction with a user through
 a computer network, with no transfer of a copy, is not conveying.
  An interactive user interface displays "Appropriate Legal Notices"
 to the extent that it includes a convenient and prominently visible
 feature that (1) displays an appropriate copyright notice, and (2)
 tells the user that there is no warranty for the work (except to the
 extent that warranties are provided), that licensees may convey the
 work under this License, and how to view a copy of this License.  If
 the interface presents a list of user commands or options, such as a
 menu, a prominent item in the list meets this criterion.
  1. Source Code.
  The "source code" for a work means the preferred form of the work
 for making modifications to it.  "Object code" means any non-source
 form of a work.
  A "Standard Interface" means an interface that either is an official
 standard defined by a recognized standards body, or, in the case of
 interfaces specified for a particular programming language, one that
 is widely used among developers working in that language.
  The "System Libraries" of an executable work include anything, other
 than the work as a whole, that (a) is included in the normal form of
 packaging a Major Component, but which is not part of that Major
 Component, and (b) serves only to enable use of the work with that
 Major Component, or to implement a Standard Interface for which an
 implementation is available to the public in source code form.  A
 "Major Component", in this context, means a major essential component
 (kernel, window system, and so on) of the specific operating system
 (if any) on which the executable work runs, or a compiler used to
 produce the work, or an object code interpreter used to run it.
  The "Corresponding Source" for a work in object code form means all
 the source code needed to generate, install, and (for an executable
 work) run the object code and to modify the work, including scripts to
 control those activities.  However, it does not include the work's
 System Libraries, or general-purpose tools or generally available free
 programs which are used unmodified in performing those activities but
 which are not part of the work.  For example, Corresponding Source
 includes interface definition files associated with source files for
 the work, and the source code for shared libraries and dynamically
 linked subprograms that the work is specifically designed to require,
 such as by intimate data communication or control flow between those
 subprograms and other parts of the work.
  The Corresponding Source need not include anything that users
 can regenerate automatically from other parts of the Corresponding
 Source.
  The Corresponding Source for a work in source code form is that
 same work.
  2. Basic Permissions.
  All rights granted under this License are granted for the term of
 copyright on the Program, and are irrevocable provided the stated
 conditions are met.  This License explicitly affirms your unlimited
 permission to run the unmodified Program.  The output from running a
 covered work is covered by this License only if the output, given its
 content, constitutes a covered work.  This License acknowledges your
 rights of fair use or other equivalent, as provided by copyright law.
  You may make, run and propagate covered works that you do not
 convey, without conditions so long as your license otherwise remains
 in force.  You may convey covered works to others for the sole purpose
 of having them make modifications exclusively for you, or provide you
 with facilities for running those works, provided that you comply with
 the terms of this License in conveying all material for which you do
 not control copyright.  Those thus making or running the covered works
 for you must do so exclusively on your behalf, under your direction
 and control, on terms that prohibit them from making any copies of
 your copyrighted material outside their relationship with you.
  Conveying under any other circumstances is permitted solely under
 the conditions stated below.  Sublicensing is not allowed; section 10
 makes it unnecessary.
  3. Protecting Users' Legal Rights From Anti-Circumvention Law.
  No covered work shall be deemed part of an effective technological
 measure under any applicable law fulfilling obligations under article
 11 of the WIPO copyright treaty adopted on 20 December 1996, or
 similar laws prohibiting or restricting circumvention of such
 measures.
  When you convey a covered work, you waive any legal power to forbid
 circumvention of technological measures to the extent such circumvention
 is effected by exercising rights under this License with respect to
 the covered work, and you disclaim any intention to limit operation or
 modification of the work as a means of enforcing, against the work's
 users, your or third parties' legal rights to forbid circumvention of
 technological measures.
  4. Conveying Verbatim Copies.
  You may convey verbatim copies of the Program's source code as you
 receive it, in any medium, provided that you conspicuously and
 appropriately publish on each copy an appropriate copyright notice;
 keep intact all notices stating that this License and any
 non-permissive terms added in accord with section 7 apply to the code;
 keep intact all notices of the absence of any warranty; and give all
 recipients a copy of this License along with the Program.
  You may charge any price or no price for each copy that you convey,
 and you may offer support or warranty protection for a fee.
  5. Conveying Modified Source Versions.
  You may convey a work based on the Program, or the modifications to
 produce it from the Program, in the form of source code under the
 terms of section 4, provided that you also meet all of these conditions:
    a) The work must carry prominent notices stating that you modified
    it, and giving a relevant date.
    b) The work must carry prominent notices stating that it is
    released under this License and any conditions added under section
    7.  This requirement modifies the requirement in section 4 to
    "keep intact all notices".
    c) You must license the entire work, as a whole, under this
    License to anyone who comes into possession of a copy.  This
    License will therefore apply, along with any applicable section 7
    additional terms, to the whole of the work, and all its parts,
    regardless of how they are packaged.  This License gives no
    permission to license the work in any other way, but it does not
    invalidate such permission if you have separately received it.
    d) If the work has interactive user interfaces, each must display
    Appropriate Legal Notices; however, if the Program has interactive
    interfaces that do not display Appropriate Legal Notices, your
    work need not make them do so.
  A compilation of a covered work with other separate and independent
 works, which are not by their nature extensions of the covered work,
 and which are not combined with it such as to form a larger program,
 in or on a volume of a storage or distribution medium, is called an
 "aggregate" if the compilation and its resulting copyright are not
 used to limit the access or legal rights of the compilation's users
 beyond what the individual works permit.  Inclusion of a covered work
 in an aggregate does not cause this License to apply to the other
 parts of the aggregate.
  6. Conveying Non-Source Forms.
  You may convey a covered work in object code form under the terms
 of sections 4 and 5, provided that you also convey the
 machine-readable Corresponding Source under the terms of this License,
 in one of these ways:
    a) Convey the object code in, or embodied in, a physical product
    (including a physical distribution medium), accompanied by the
    Corresponding Source fixed on a durable physical medium
    customarily used for software interchange.
    b) Convey the object code in, or embodied in, a physical product
    (including a physical distribution medium), accompanied by a
    written offer, valid for at least three years and valid for as
    long as you offer spare parts or customer support for that product
    model, to give anyone who possesses the object code either (1) a
    copy of the Corresponding Source for all the software in the
    product that is covered by this License, on a durable physical
    medium customarily used for software interchange, for a price no
    more than your reasonable cost of physically performing this
    conveying of source, or (2) access to copy the
    Corresponding Source from a network server at no charge.
    c) Convey individual copies of the object code with a copy of the
    written offer to provide the Corresponding Source.  This
    alternative is allowed only occasionally and noncommercially, and
    only if you received the object code with such an offer, in accord
    with subsection 6b.
    d) Convey the object code by offering access from a designated
    place (gratis or for a charge), and offer equivalent access to the
    Corresponding Source in the same way through the same place at no
    further charge.  You need not require recipients to copy the
    Corresponding Source along with the object code.  If the place to
    copy the object code is a network server, the Corresponding Source
    may be on a different server (operated by you or a third party)
    that supports equivalent copying facilities, provided you maintain
    clear directions next to the object code saying where to find the
    Corresponding Source.  Regardless of what server hosts the
    Corresponding Source, you remain obligated to ensure that it is
    available for as long as needed to satisfy these requirements.
    e) Convey the object code using peer-to-peer transmission, provided
    you inform other peers where the object code and Corresponding
    Source of the work are being offered to the general public at no
    charge under subsection 6d.
  A separable portion of the object code, whose source code is excluded
 from the Corresponding Source as a System Library, need not be
 included in conveying the object code work.
  A "User Product" is either (1) a "consumer product", which means any
 tangible personal property which is normally used for personal, family,
 or household purposes, or (2) anything designed or sold for incorporation
 into a dwelling.  In determining whether a product is a consumer product,
 doubtful cases shall be resolved in favor of coverage.  For a particular
 product received by a particular user, "normally used" refers to a
 typical or common use of that class of product, regardless of the status
 of the particular user or of the way in which the particular user
 actually uses, or expects or is expected to use, the product.  A product
 is a consumer product regardless of whether the product has substantial
 commercial, industrial or non-consumer uses, unless such uses represent
 the only significant mode of use of the product.
  "Installation Information" for a User Product means any methods,
 procedures, authorization keys, or other information required to install
 and execute modified versions of a covered work in that User Product from
 a modified version of its Corresponding Source.  The information must
 suffice to ensure that the continued functioning of the modified object
 code is in no case prevented or interfered with solely because
 modification has been made.
  If you convey an object code work under this section in, or with, or
 specifically for use in, a User Product, and the conveying occurs as
 part of a transaction in which the right of possession and use of the
 User Product is transferred to the recipient in perpetuity or for a
 fixed term (regardless of how the transaction is characterized), the
 Corresponding Source conveyed under this section must be accompanied
 by the Installation Information.  But this requirement does not apply
 if neither you nor any third party retains the ability to install
 modified object code on the User Product (for example, the work has
 been installed in ROM).
  The requirement to provide Installation Information does not include a
 requirement to continue to provide support service, warranty, or updates
 for a work that has been modified or installed by the recipient, or for
 the User Product in which it has been modified or installed.  Access to a
 network may be denied when the modification itself materially and
 adversely affects the operation of the network or violates the rules and
 protocols for communication across the network.
  Corresponding Source conveyed, and Installation Information provided,
 in accord with this section must be in a format that is publicly
 documented (and with an implementation available to the public in
 source code form), and must require no special password or key for
 unpacking, reading or copying.
  7. Additional Terms.
  "Additional permissions" are terms that supplement the terms of this
 License by making exceptions from one or more of its conditions.
 Additional permissions that are applicable to the entire Program shall
 be treated as though they were included in this License, to the extent
 that they are valid under applicable law.  If additional permissions
 apply only to part of the Program, that part may be used separately
 under those permissions, but the entire Program remains governed by
 this License without regard to the additional permissions.
  When you convey a copy of a covered work, you may at your option
 remove any additional permissions from that copy, or from any part of
 it.  (Additional permissions may be written to require their own
 removal in certain cases when you modify the work.)  You may place
 additional permissions on material, added by you to a covered work,
 for which you have or can give appropriate copyright permission.
  Notwithstanding any other provision of this License, for material you
 add to a covered work, you may (if authorized by the copyright holders of
 that material) supplement the terms of this License with terms:
    a) Disclaiming warranty or limiting liability differently from the
    terms of sections 15 and 16 of this License; or
    b) Requiring preservation of specified reasonable legal notices or
    author attributions in that material or in the Appropriate Legal
    Notices displayed by works containing it; or
    c) Prohibiting misrepresentation of the origin of that material, or
    requiring that modified versions of such material be marked in
    reasonable ways as different from the original version; or
    d) Limiting the use for publicity purposes of names of licensors or
    authors of the material; or
    e) Declining to grant rights under trademark law for use of some
    trade names, trademarks, or service marks; or
    f) Requiring indemnification of licensors and authors of that
    material by anyone who conveys the material (or modified versions of
    it) with contractual assumptions of liability to the recipient, for
    any liability that these contractual assumptions directly impose on
    those licensors and authors.
  All other non-permissive additional terms are considered "further
 restrictions" within the meaning of section 10.  If the Program as you
 received it, or any part of it, contains a notice stating that it is
 governed by this License along with a term that is a further
 restriction, you may remove that term.  If a license document contains
 a further restriction but permits relicensing or conveying under this
 License, you may add to a covered work material governed by the terms
 of that license document, provided that the further restriction does
 not survive such relicensing or conveying.
  If you add terms to a covered work in accord with this section, you
 must place, in the relevant source files, a statement of the
 additional terms that apply to those files, or a notice indicating
 where to find the applicable terms.
  Additional terms, permissive or non-permissive, may be stated in the
 form of a separately written license, or stated as exceptions;
 the above requirements apply either way.
  8. Termination.
  You may not propagate or modify a covered work except as expressly
 provided under this License.  Any attempt otherwise to propagate or
 modify it is void, and will automatically terminate your rights under
 this License (including any patent licenses granted under the third
 paragraph of section 11).
  However, if you cease all violation of this License, then your
 license from a particular copyright holder is reinstated (a)
 provisionally, unless and until the copyright holder explicitly and
 finally terminates your license, and (b) permanently, if the copyright
 holder fails to notify you of the violation by some reasonable means
 prior to 60 days after the cessation.
  Moreover, your license from a particular copyright holder is
 reinstated permanently if the copyright holder notifies you of the
 violation by some reasonable means, this is the first time you have
 received notice of violation of this License (for any work) from that
 copyright holder, and you cure the violation prior to 30 days after
 your receipt of the notice.
  Termination of your rights under this section does not terminate the
 licenses of parties who have received copies or rights from you under
 this License.  If your rights have been terminated and not permanently
 reinstated, you do not qualify to receive new licenses for the same
 material under section 10.
  9. Acceptance Not Required for Having Copies.
  You are not required to accept this License in order to receive or
 run a copy of the Program.  Ancillary propagation of a covered work
 occurring solely as a consequence of using peer-to-peer transmission
 to receive a copy likewise does not require acceptance.  However,
 nothing other than this License grants you permission to propagate or
 modify any covered work.  These actions infringe copyright if you do
 not accept this License.  Therefore, by modifying or propagating a
 covered work, you indicate your acceptance of this License to do so.
  10. Automatic Licensing of Downstream Recipients.
  Each time you convey a covered work, the recipient automatically
 receives a license from the original licensors, to run, modify and
 propagate that work, subject to this License.  You are not responsible
 for enforcing compliance by third parties with this License.
  An "entity transaction" is a transaction transferring control of an
 organization, or substantially all assets of one, or subdividing an
 organization, or merging organizations.  If propagation of a covered
 work results from an entity transaction, each party to that
 transaction who receives a copy of the work also receives whatever
 licenses to the work the party's predecessor in interest had or could
 give under the previous paragraph, plus a right to possession of the
 Corresponding Source of the work from the predecessor in interest, if
 the predecessor has it or can get it with reasonable efforts.
  You may not impose any further restrictions on the exercise of the
 rights granted or affirmed under this License.  For example, you may
 not impose a license fee, royalty, or other charge for exercise of
 rights granted under this License, and you may not initiate litigation
 (including a cross-claim or counterclaim in a lawsuit) alleging that
 any patent claim is infringed by making, using, selling, offering for
 sale, or importing the Program or any portion of it.
  11. Patents.
  A "contributor" is a copyright holder who authorizes use under this
 License of the Program or a work on which the Program is based.  The
 work thus licensed is called the contributor's "contributor version".
  A contributor's "essential patent claims" are all patent claims
 owned or controlled by the contributor, whether already acquired or
 hereafter acquired, that would be infringed by some manner, permitted
 by this License, of making, using, or selling its contributor version,
 but do not include claims that would be infringed only as a
 consequence of further modification of the contributor version.  For
 purposes of this definition, "control" includes the right to grant
 patent sublicenses in a manner consistent with the requirements of
 this License.
  Each contributor grants you a non-exclusive, worldwide, royalty-free
 patent license under the contributor's essential patent claims, to
 make, use, sell, offer for sale, import and otherwise run, modify and
 propagate the contents of its contributor version.
  In the following three paragraphs, a "patent license" is any express
 agreement or commitment, however denominated, not to enforce a patent
 (such as an express permission to practice a patent or covenant not to
 sue for patent infringement).  To "grant" such a patent license to a
 party means to make such an agreement or commitment not to enforce a
 patent against the party.
  If you convey a covered work, knowingly relying on a patent license,
 and the Corresponding Source of the work is not available for anyone
 to copy, free of charge and under the terms of this License, through a
 publicly available network server or other readily accessible means,
 then you must either (1) cause the Corresponding Source to be so
 available, or (2) arrange to deprive yourself of the benefit of the
 patent license for this particular work, or (3) arrange, in a manner
 consistent with the requirements of this License, to extend the patent
 license to downstream recipients.  "Knowingly relying" means you have
 actual knowledge that, but for the patent license, your conveying the
 covered work in a country, or your recipient's use of the covered work
 in a country, would infringe one or more identifiable patents in that
 country that you have reason to believe are valid.
  If, pursuant to or in connection with a single transaction or
 arrangement, you convey, or propagate by procuring conveyance of, a
 covered work, and grant a patent license to some of the parties
 receiving the covered work authorizing them to use, propagate, modify
 or convey a specific copy of the covered work, then the patent license
 you grant is automatically extended to all recipients of the covered
 work and works based on it.
  A patent license is "discriminatory" if it does not include within
 the scope of its coverage, prohibits the exercise of, or is
 conditioned on the non-exercise of one or more of the rights that are
 specifically granted under this License.  You may not convey a covered
 work if you are a party to an arrangement with a third party that is
 in the business of distributing software, under which you make payment
 to the third party based on the extent of your activity of conveying
 the work, and under which the third party grants, to any of the
 parties who would receive the covered work from you, a discriminatory
 patent license (a) in connection with copies of the covered work
 conveyed by you (or copies made from those copies), or (b) primarily
 for and in connection with specific products or compilations that
 contain the covered work, unless you entered into that arrangement,
 or that patent license was granted, prior to 28 March 2007.
  Nothing in this License shall be construed as excluding or limiting
 any implied license or other defenses to infringement that may
 otherwise be available to you under applicable patent law.
  12. No Surrender of Others' Freedom.
  If conditions are imposed on you (whether by court order, agreement or
 otherwise) that contradict the conditions of this License, they do not
 excuse you from the conditions of this License.  If you cannot convey a
 covered work so as to satisfy simultaneously your obligations under this
 License and any other pertinent obligations, then as a consequence you may
 not convey it at all.  For example, if you agree to terms that obligate you
 to collect a royalty for further conveying from those to whom you convey
 the Program, the only way you could satisfy both those terms and this
 License would be to refrain entirely from conveying the Program.
  13. Use with the GNU Affero General Public License.
  Notwithstanding any other provision of this License, you have
 permission to link or combine any covered work with a work licensed
 under version 3 of the GNU Affero General Public License into a single
 combined work, and to convey the resulting work.  The terms of this
 License will continue to apply to the part which is the covered work,
 but the special requirements of the GNU Affero General Public License,
 section 13, concerning interaction through a network will apply to the
 combination as such.
  14. Revised Versions of this License.
  The Free Software Foundation may publish revised and/or new versions of
 the GNU General Public License from time to time.  Such new versions will
 be similar in spirit to the present version, but may differ in detail to
 address new problems or concerns.
  Each version is given a distinguishing version number.  If the
 Program specifies that a certain numbered version of the GNU General
 Public License "or any later version" applies to it, you have the
 option of following the terms and conditions either of that numbered
 version or of any later version published by the Free Software
 Foundation.  If the Program does not specify a version number of the
 GNU General Public License, you may choose any version ever published
 by the Free Software Foundation.
  If the Program specifies that a proxy can decide which future
 versions of the GNU General Public License can be used, that proxy's
 public statement of acceptance of a version permanently authorizes you
 to choose that version for the Program.
  Later license versions may give you additional or different
 permissions.  However, no additional obligations are imposed on any
 author or copyright holder as a result of your choosing to follow a
 later version.
  15. Disclaimer of Warranty.
  THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
 APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
 HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
 OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
 THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
 PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
 IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
 ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
  16. Limitation of Liability.
  IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
 WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
 THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
 GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
 USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
 DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
 PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
 EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
 SUCH DAMAGES.
  17. Interpretation of Sections 15 and 16.
  If the disclaimer of warranty and limitation of liability provided
 above cannot be given local legal effect according to their terms,
 reviewing courts shall apply local law that most closely approximates
 an absolute waiver of all civil liability in connection with the
 Program, unless a warranty or assumption of liability accompanies a
 copy of the Program in return for a fee.
                     END OF TERMS AND CONDITIONS
            How to Apply These Terms to Your New Programs
  If you develop a new program, and you want it to be of the greatest
 possible use to the public, the best way to achieve this is to make it
 free software which everyone can redistribute and change under these terms.
  To do so, attach the following notices to the program.  It is safest
 to attach them to the start of each source file to most effectively
 state the exclusion of warranty; and each file should have at least
 the "copyright" line and a pointer to where the full notice is found.
    <one line to give the program's name and a brief idea of what it does.>
    Copyright (C) <year>  <name of author>
    This program is free software: you can redistribute it and/or modify
    it under the terms of the GNU General Public License as published by
    the Free Software Foundation, either version 3 of the License, or
    (at your option) any later version.
    This program is distributed in the hope that it will be useful,
    but WITHOUT ANY WARRANTY; without even the implied warranty of
    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
    GNU General Public License for more details.
    You should have received a copy of the GNU General Public License
    along with this program.  If not, see <https://www.gnu.org/licenses/>.
 Also add information on how to contact you by electronic and paper mail.
  If the program does terminal interaction, make it output a short
 notice like this when it starts in an interactive mode:
    <program>  Copyright (C) <year>  <name of author>
    This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
    This is free software, and you are welcome to redistribute it
    under certain conditions; type `show c' for details.
 The hypothetical commands `show w' and `show c' should show the appropriate
 parts of the General Public License.  Of course, your program's commands
 might be different; for a GUI interface, you would use an "about box".
  You should also get your employer (if you work as a programmer) or school,
 if any, to sign a "copyright disclaimer" for the program, if necessary.
 For more information on this, and how to apply and follow the GNU GPL, see
 <https://www.gnu.org/licenses/>.
  The GNU General Public License does not permit incorporating your program
 into proprietary programs.  If your program is a subroutine library, you
 may consider it more useful to permit linking proprietary applications with
 the library.  If this is what you want to do, use the GNU Lesser General
 Public License instead of this License.  But first, please read
 <https://www.gnu.org/licenses/why-not-lgpl.html>.
@@ -11,6 +11,9 @@ from coreml_suite.nodes import (
    CoreMLConverter,
    COREML_LOAD_LORA,
 )
 from coreml_suite.lcm import (
    COREML_CONVERT_LCM,
 )
 NODE_CLASS_MAPPINGS = {
    "CoreMLUNetLoader": CoreMLLoaderUNet,
@@ -19,6 +22,7 @@ NODE_CLASS_MAPPINGS = {
    "CoreMLModelAdapter": CoreMLModelAdapter,
    "Core ML LoRA Loader": COREML_LOAD_LORA,
    "Core ML Converter": CoreMLConverter,
    "Core ML LCM Converter": COREML_CONVERT_LCM,
 }
 NODE_DISPLAY_NAME_MAPPINGS = {
    "CoreMLUNetLoader": "Load Core ML UNet",
@@ -27,4 +31,5 @@ NODE_DISPLAY_NAME_MAPPINGS = {
    "CoreMLModelAdapter": "Core ML Adapter (Experimental)",
    "Core ML LoRA Loader": "Load LoRA to use with Core ML",
    "Core ML Converter": "Convert Checkpoint to Core ML",
    "Core ML LCM Converter": "Convert LCM to Core ML",
 }
@@ -0,0 +1,5 @@
 ATTENTION_IMPLEMENTATIONS = (
    "SPLIT_EINSUM",
    "SPLIT_EINSUM_V2",
    "ORIGINAL",
 )
@@ -4,7 +4,7 @@ from comfy import supported_models_base
 from comfy import latent_formats
 from comfy.model_detection import convert_config
-from coreml_diffusion import ModelVersion
+from coreml_suite.model_version import ModelVersion
 config_map = {
@@ -0,0 +1,9 @@
 """Core ML conversion helpers.
 The conversion approach originates from Apple's ml-stable-diffusion
 (https://github.com/apple/ml-stable-diffusion). This implementation has since
 diverged: it runs natively on diffusers' UNet2DConditionModel with its own
 SPLIT_EINSUM / SPLIT_EINSUM_V2 attention processors and no longer depends on
 that package. The intent is to keep iterating on these methods independently
 while tracking current tooling.
 """
@@ -0,0 +1,239 @@
 import logging
 import torch
 logger = logging.getLogger(__name__)
 CHUNK_SIZE = 512
 def apply_attention_implementation(unet, attention_implementation):
    if attention_implementation == "ORIGINAL":
        return unet
    if attention_implementation == "SPLIT_EINSUM":
        unet.set_attn_processor(SplitEinsumAttnProcessor())
        return unet
    if attention_implementation == "SPLIT_EINSUM_V2":
        unet.set_attn_processor(SplitEinsumV2AttnProcessor())
        return unet
    raise ValueError(f"Unsupported attention implementation: {attention_implementation}")
 class SplitEinsumAttnProcessor:
    def __call__(
        self,
        attn,
        hidden_states,
        encoder_hidden_states=None,
        attention_mask=None,
        temb=None,
        *args,
        **kwargs,
    ):
        return _attention_forward(
            attn,
            hidden_states,
            encoder_hidden_states,
            attention_mask,
            temb,
            split_einsum,
        )
 class SplitEinsumV2AttnProcessor:
    def __call__(
        self,
        attn,
        hidden_states,
        encoder_hidden_states=None,
        attention_mask=None,
        temb=None,
        *args,
        **kwargs,
    ):
        return _attention_forward(
            attn,
            hidden_states,
            encoder_hidden_states,
            attention_mask,
            temb,
            split_einsum_v2,
        )
 def _attention_forward(
    attn,
    hidden_states,
    encoder_hidden_states,
    attention_mask,
    temb,
    attention_fn,
 ):
    residual = hidden_states
    if attn.spatial_norm is not None:
        hidden_states = attn.spatial_norm(hidden_states, temb)
    input_ndim = hidden_states.ndim
    if input_ndim == 4:
        batch_size, channel, height, width = hidden_states.shape
        hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
    else:
        batch_size, _, channel = hidden_states.shape
        height = None
        width = None
    batch_size, key_sequence_length, _ = (
        hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
    )
    if attention_mask is not None:
        attention_mask = attn.prepare_attention_mask(
            attention_mask,
            key_sequence_length,
            batch_size,
        )
        attention_mask = _prepare_split_einsum_mask(
            attention_mask,
            batch_size,
            attn.heads,
            key_sequence_length,
        )
    if attn.group_norm is not None:
        hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
    query = attn.to_q(hidden_states)
    if encoder_hidden_states is None:
        encoder_hidden_states = hidden_states
    elif attn.norm_cross:
        encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
    key = attn.to_k(encoder_hidden_states)
    value = attn.to_v(encoder_hidden_states)
    batch_size = query.shape[0]
    dim_head = attn.inner_kv_dim // attn.heads
    query = _linear_projection_to_bchw(query)
    key = _linear_projection_to_bchw(key)
    value = _linear_projection_to_bchw(value)
    hidden_states = attention_fn(
        query,
        key,
        value,
        attention_mask,
        attn.heads,
        dim_head,
    )
    hidden_states = hidden_states.squeeze(2).transpose(1, 2)
    hidden_states = hidden_states.reshape(batch_size, -1, attn.inner_dim)
    hidden_states = attn.to_out[0](hidden_states)
    hidden_states = attn.to_out[1](hidden_states)
    if input_ndim == 4:
        hidden_states = hidden_states.transpose(-1, -2).reshape(
            batch_size,
            channel,
            height,
            width,
        )
    if attn.residual_connection:
        hidden_states = hidden_states + residual
    hidden_states = hidden_states / attn.rescale_output_factor
    return hidden_states
 def split_einsum(q, k, v, mask, heads, dim_head):
    q_heads = _split_heads(q, heads, dim_head)
    k = k.transpose(1, 3)
    k_heads = [
        k[:, :, :, head_idx * dim_head : (head_idx + 1) * dim_head]
        for head_idx in range(heads)
    ]
    v_heads = _split_heads(v, heads, dim_head)
    weights = [
        torch.einsum("bchq,bkhc->bkhq", query, key) * (dim_head**-0.5)
        for query, key in zip(q_heads, k_heads)
    ]
    if mask is not None:
        weights = [weight + mask for weight in weights]
    weights = [weight.softmax(dim=1) for weight in weights]
    outputs = [
        torch.einsum("bkhq,bchk->bchq", weight, value)
        for weight, value in zip(weights, v_heads)
    ]
    return torch.cat(outputs, dim=1)
 def split_einsum_v2(q, k, v, mask, heads, dim_head):
    query_length = q.size(3)
    num_chunks = query_length // CHUNK_SIZE
    if num_chunks == 0:
        logger.info(
            "SPLIT_EINSUM_V2 query sequence is shorter than %s; using SPLIT_EINSUM.",
            CHUNK_SIZE,
        )
        return split_einsum(q, k, v, mask, heads, dim_head)
    q_heads = _split_heads(q, heads, dim_head)
    q_chunks = [
        [
            head[..., chunk_idx * CHUNK_SIZE : (chunk_idx + 1) * CHUNK_SIZE]
            for chunk_idx in range(num_chunks)
        ]
        for head in q_heads
    ]
    k = k.transpose(1, 3)
    k_heads = [
        k[:, :, :, head_idx * dim_head : (head_idx + 1) * dim_head]
        for head_idx in range(heads)
    ]
    v_heads = _split_heads(v, heads, dim_head)
    head_outputs = []
    for query_chunks, key, value in zip(q_chunks, k_heads, v_heads):
        chunk_outputs = []
        for query_chunk in query_chunks:
            weights = torch.einsum("bchq,bkhc->bkhq", query_chunk, key)
            weights = weights * (dim_head**-0.5)
            if mask is not None:
                weights = weights + mask
            weights = weights.softmax(dim=1)
            chunk_outputs.append(torch.einsum("bkhq,bchk->bchq", weights, value))
        head_outputs.append(torch.cat(chunk_outputs, dim=3))
    return torch.cat(head_outputs, dim=1)
 def _split_heads(x, heads, dim_head):
    return [
        x[:, head_idx * dim_head : (head_idx + 1) * dim_head, :, :]
        for head_idx in range(heads)
    ]
 def _linear_projection_to_bchw(x):
    return x.transpose(1, 2).unsqueeze(2)
 def _prepare_split_einsum_mask(mask, batch_size, heads, key_sequence_length):
    if mask.ndim == 2:
        mask = mask[:, None, :]
    if mask.shape[0] == batch_size * heads:
        mask = mask.reshape(batch_size, heads, -1, key_sequence_length)
        mask = mask[:, 0]
    if mask.ndim == 3:
        mask = mask[:, :, None, None]
    return mask
@@ -0,0 +1,20 @@
 def conv2d_output_shape(height, width, conv):
    """Return the spatial output shape for a torch.nn.Conv2d-like module."""
    kernel_h, kernel_w = _pair(conv.kernel_size)
    stride_h, stride_w = _pair(conv.stride)
    pad_h, pad_w = _pair(conv.padding)
    dilation_h, dilation_w = _pair(conv.dilation)
    out_h = _conv_output_dim(height, kernel_h, stride_h, pad_h, dilation_h)
    out_w = _conv_output_dim(width, kernel_w, stride_w, pad_w, dilation_w)
    return out_h, out_w
 def _conv_output_dim(size, kernel, stride, padding, dilation):
    return ((size + (2 * padding) - (dilation * (kernel - 1)) - 1) // stride) + 1
 def _pair(value):
    if isinstance(value, tuple):
        return value
    return value, value
@@ -0,0 +1,61 @@
 from types import MethodType
 from diffusers.models.transformers.transformer_2d import Transformer2DModel
 def prepare_unet_for_coreml_trace(unet):
    for module in unet.modules():
        if isinstance(module, Transformer2DModel):
            module._operate_on_continuous_inputs = MethodType(
                _operate_on_continuous_inputs,
                module,
            )
            module._get_output_for_continuous_inputs = MethodType(
                _get_output_for_continuous_inputs,
                module,
            )
    return unet
 def _operate_on_continuous_inputs(self, hidden_states):
    hidden_states = self.norm(hidden_states)
    if not self.use_linear_projection:
        hidden_states = self.proj_in(hidden_states)
        inner_dim = self.inner_dim
        hidden_states = hidden_states.flatten(2).transpose(1, 2)
    else:
        inner_dim = hidden_states.shape[1]
        hidden_states = hidden_states.flatten(2).transpose(1, 2)
        hidden_states = self.proj_in(hidden_states)
    return hidden_states, inner_dim
 def _get_output_for_continuous_inputs(
    self,
    hidden_states,
    residual,
    batch_size,
    height,
    width,
    inner_dim,
 ):
    if not self.use_linear_projection:
        hidden_states = hidden_states.transpose(1, 2).reshape(
            batch_size,
            inner_dim,
            height,
            width,
        )
        hidden_states = self.proj_out(hidden_states)
    else:
        hidden_states = self.proj_out(hidden_states)
        hidden_states = hidden_states.transpose(1, 2).reshape(
            batch_size,
            inner_dim,
            height,
            width,
        )
    return hidden_states + residual
@@ -0,0 +1,54 @@
 import torch
 class CoreMLUNetWrapper(torch.nn.Module):
    """Adapt diffusers UNet inputs to CoreMLSuite's stable Core ML contract."""
    def __init__(self, unet, model_version):
        super().__init__()
        self.unet = unet
        self.model_version = model_version
    def forward(self, sample, timestep, encoder_hidden_states, *extra_inputs):
        input_index = 0
        timestep_cond = None
        if self._is_lcm:
            timestep_cond = extra_inputs[input_index]
            input_index += 1
        added_cond_kwargs = None
        if self._is_sdxl:
            time_ids = extra_inputs[input_index]
            text_embeds = extra_inputs[input_index + 1]
            input_index += 2
            added_cond_kwargs = {
                "time_ids": time_ids,
                "text_embeds": text_embeds,
            }
        additional_residuals = extra_inputs[input_index:]
        down_residuals = None
        mid_residual = None
        if additional_residuals:
            down_residuals = tuple(additional_residuals[:-1])
            mid_residual = additional_residuals[-1]
        outputs = self.unet(
            sample,
            timestep,
            encoder_hidden_states=encoder_hidden_states,
            timestep_cond=timestep_cond,
            added_cond_kwargs=added_cond_kwargs,
            down_block_additional_residuals=down_residuals,
            mid_block_additional_residual=mid_residual,
            return_dict=False,
        )
        return outputs[0]
    @property
    def _is_lcm(self):
        return self.model_version.name == "LCM"
    @property
    def _is_sdxl(self):
        return self.model_version.name in {"SDXL", "SDXL_REFINER"}
@@ -0,0 +1,322 @@
 import gc
 import os
 import time
 import coremltools as ct
 import numpy as np
 import torch
 from diffusers import UNet2DConditionModel
 from coreml_suite.attention import ATTENTION_IMPLEMENTATIONS
 from coreml_suite.conversion.attention import apply_attention_implementation
 from coreml_suite.conversion.shapes import conv2d_output_shape
 from coreml_suite.conversion.trace import prepare_unet_for_coreml_trace
 from coreml_suite.conversion.unet import CoreMLUNetWrapper
 from coreml_suite.logger import logger
 from coreml_suite.model_version import ModelVersion
 DEFAULT_TRACE_TIMESTEP = 999.0
 TEXT_TOKEN_SEQUENCE_LENGTH = 77
 def get_unet(model_version: ModelVersion, ref_unet, attention_implementation):
    ref_unet = prepare_unet_for_coreml_trace(ref_unet)
    unet = apply_attention_implementation(
        ref_unet.eval(),
        attention_implementation,
    )
    return CoreMLUNetWrapper(unet, model_version)
 def get_encoder_hidden_states_shape(ref_unet, batch_size):
    encoder_hidden_states_shape = (
        batch_size,
        TEXT_TOKEN_SEQUENCE_LENGTH,
        ref_unet.config.cross_attention_dim,
    )
    return encoder_hidden_states_shape
 def get_coreml_inputs(sample_inputs):
    coreml_sample_unet_inputs = {
        k: v.numpy().astype(np.float16) for k, v in sample_inputs.items()
    }
    return [
        ct.TensorType(
            name=k,
            shape=v.shape,
            dtype=v.numpy().dtype if isinstance(v, torch.Tensor) else v.dtype,
        )
        for k, v in coreml_sample_unet_inputs.items()
    ]
 def load_coreml_model(out_path):
    logger.info(f"Loading model from {out_path}")
    start = time.time()
    coreml_model = ct.models.MLModel(out_path)
    logger.info(f"Loading {out_path} took {time.time() - start:.1f} seconds")
    return coreml_model
 def convert_to_coreml(
    submodule_name, torchscript_module, sample_inputs, output_names, out_path
 ):
    if os.path.exists(out_path):
        logger.info(f"Skipping export because {out_path} already exists")
        coreml_model = load_coreml_model(out_path)
    else:
        logger.info(f"Converting {submodule_name} to CoreML..")
        coreml_model = ct.convert(
            torchscript_module,
            convert_to="mlprogram",
            minimum_deployment_target=ct.target.macOS13,
            inputs=sample_inputs,
            outputs=[
                ct.TensorType(name=name, dtype=np.float32) for name in output_names
            ],
            skip_model_load=True,
        )
        del torchscript_module
        gc.collect()
    return coreml_model
 def get_out_path(submodule_name, model_name):
    from folder_paths import get_folder_paths
    fname = f"{model_name}_{submodule_name}.mlpackage"
    unet_path = get_folder_paths(submodule_name)[0]
    out_path = os.path.join(unet_path, fname)
    return out_path
 def get_sample_input(batch_size, encoder_hidden_states_shape, sample_shape):
    sample_unet_inputs = dict(
        [
            ("sample", torch.rand(*sample_shape)),
            (
                "timestep",
                torch.tensor([DEFAULT_TRACE_TIMESTEP] * batch_size).to(torch.float32),
            ),
            ("encoder_hidden_states", torch.rand(*encoder_hidden_states_shape)),
        ]
    )
    return sample_unet_inputs
 def lcm_inputs(sample_unet_inputs):
    batch_size = sample_unet_inputs["sample"].shape[0]
    return {"timestep_cond": torch.randn(batch_size, 256).to(torch.float32)}
 def sdxl_inputs(sample_unet_inputs, ref_unet, model_version):
    sample_shape = sample_unet_inputs["sample"].shape
    batch_size = sample_shape[0]
    h = sample_shape[2] * 8
    w = sample_shape[3] * 8
    original_size = (h, w)
    crops_coords_top_left = (0, 0)
    is_refiner = model_version == ModelVersion.SDXL_REFINER
    if is_refiner:
        aesthetic_score = (6.0,)
        time_ids_list = list(original_size + crops_coords_top_left + aesthetic_score)
    else:
        target_size = (h, w)
        time_ids_list = list(original_size + crops_coords_top_left + target_size)
    time_ids = torch.tensor(time_ids_list).repeat(batch_size, 1).to(torch.int64)
    text_embeds_shape = (batch_size, get_sdxl_text_embeds_dim(ref_unet, len(time_ids_list)))
    return {
        "time_ids": time_ids,
        "text_embeds": torch.randn(*text_embeds_shape).to(torch.float32),
    }
 def get_sdxl_text_embeds_dim(ref_unet, time_ids_dim):
    projection_dim = ref_unet.config.projection_class_embeddings_input_dim
    time_embed_dim = ref_unet.config.addition_time_embed_dim
    return projection_dim - (time_ids_dim * time_embed_dim)
 def get_inputs_spec(inputs):
    inputs_spec = {k: (v.shape, v.dtype) for k, v in inputs.items()}
    return inputs_spec
 def add_cnet_support(sample_shape, reference_unet):
    additional_residuals_shapes = []
    batch_size = sample_shape[0]
    h, w = sample_shape[2:]
    # conv_in
    out_h, out_w = conv2d_output_shape(
        h,
        w,
        reference_unet.conv_in,
    )
    additional_residuals_shapes.append(
        (batch_size, reference_unet.conv_in.out_channels, out_h, out_w)
    )
    # down_blocks
    for down_block in reference_unet.down_blocks:
        additional_residuals_shapes += [
            (batch_size, resnet.out_channels, out_h, out_w)
            for resnet in down_block.resnets
        ]
        if hasattr(down_block, "downsamplers") and down_block.downsamplers is not None:
            for downsampler in down_block.downsamplers:
                out_h, out_w = conv2d_output_shape(out_h, out_w, downsampler.conv)
            additional_residuals_shapes.append(
                (
                    batch_size,
                    down_block.downsamplers[-1].conv.out_channels,
                    out_h,
                    out_w,
                )
            )
    # mid_block
    additional_residuals_shapes.append(
        (batch_size, reference_unet.mid_block.resnets[-1].out_channels, out_h, out_w)
    )
    additional_inputs = {}
    for i, shape in enumerate(additional_residuals_shapes):
        sample_residual_input = torch.rand(*shape)
        additional_inputs[f"additional_residual_{i}"] = sample_residual_input
    return additional_inputs
 def convert_unet(
    ref_unet,
    model_version: ModelVersion,
    unet_out_path: str,
    batch_size: int = 1,
    sample_size: tuple[int, int] = (64, 64),
    controlnet_support: bool = False,
    attention_implementation: str = ATTENTION_IMPLEMENTATIONS[0],
    quantize_nbits: str = "none",
 ):
    coreml_unet = get_unet(model_version, ref_unet, attention_implementation)
    sample_shape = (
        batch_size,  # B
        ref_unet.config.in_channels,  # C
        sample_size[0],  # H
        sample_size[1],  # W
    )
    encoder_hidden_states_shape = get_encoder_hidden_states_shape(ref_unet, batch_size)
    sample_inputs = get_sample_input(
        batch_size, encoder_hidden_states_shape, sample_shape
    )
    if model_version == ModelVersion.LCM:
        sample_inputs |= lcm_inputs(sample_inputs)
    if model_version in {ModelVersion.SDXL, ModelVersion.SDXL_REFINER}:
        sample_inputs |= sdxl_inputs(sample_inputs, ref_unet, model_version)
    if controlnet_support:
        sample_inputs |= add_cnet_support(sample_shape, ref_unet)
    sample_inputs_spec = get_inputs_spec(sample_inputs)
    logger.info(f"Sample UNet inputs spec: {sample_inputs_spec}")
    logger.info("JIT tracing..")
    traced_unet = torch.jit.trace(
        coreml_unet, example_inputs=list(sample_inputs.values())
    )
    logger.info("Done.")
    coreml_sample_inputs = get_coreml_inputs(sample_inputs)
    coreml_unet = convert_to_coreml(
        "unet", traced_unet, coreml_sample_inputs, ["noise_pred"], unet_out_path
    )
    del traced_unet
    gc.collect()
    if quantize_nbits != "none":
        # Opt-in k-means weight palettization. The default path
        # (quantize_nbits="none") leaves the traced UNet untouched.
        from coremltools.optimize.coreml import (
            OpPalettizerConfig,
            OptimizationConfig,
            palettize_weights,
        )
        nbits = int(quantize_nbits)
        logger.info(f"Palettizing UNet weights to {nbits}-bit (kmeans)..")
        t0 = time.time()
        cfg = OptimizationConfig(
            global_config=OpPalettizerConfig(mode="kmeans", nbits=nbits)
        )
        coreml_unet = palettize_weights(coreml_unet, config=cfg)
        logger.info(f"Palettization took {time.time() - t0:.1f}s")
    coreml_unet.save(unet_out_path)
    logger.info(f"Saved unet into {unet_out_path}")
 def convert(
    ckpt_path: str,
    model_version: ModelVersion,
    unet_out_path: str,
    batch_size: int = 1,
    sample_size: tuple[int, int] = (64, 64),
    controlnet_support: bool = False,
    lora_weights: list[tuple[str | os.PathLike, float]] = None,
    attn_impl: str = ATTENTION_IMPLEMENTATIONS[0],
    config_path: str = None,
    quantize_nbits: str = "none",
 ):
    if os.path.exists(unet_out_path):
        logger.info(f"Found existing model at {unet_out_path}! Skipping..")
        return
    if attn_impl not in ATTENTION_IMPLEMENTATIONS:
        raise ValueError(
            f"Unsupported attention implementation {attn_impl!r}. "
            f"Expected one of {ATTENTION_IMPLEMENTATIONS}."
        )
    ref_unet = load_unet(ckpt_path, config_path)
    for i, lora_weight in enumerate(lora_weights or []):
        lora_path, strength = lora_weight
        adapter_name = f"lora_{i}"
        ref_unet.load_lora_adapter(lora_path, adapter_name=adapter_name)
        ref_unet.set_adapters([adapter_name], weights=[strength])
        ref_unet.fuse_lora()
    convert_unet(
        ref_unet,
        model_version,
        unet_out_path,
        batch_size,
        sample_size,
        controlnet_support,
        attention_implementation=attn_impl,
        quantize_nbits=quantize_nbits,
    )
 def load_unet(ckpt_path, config_path):
    return UNet2DConditionModel.from_single_file(
        ckpt_path,
        original_config=config_path,
    )
@@ -0,0 +1,68 @@
 """Pure out_name composition for the Core ML UNet artifact.
 Extracted from CoreMLConverter.convert so the filename contract
 can be tested + reused without instantiating the node. The string is the
 cache key: every workflow that references a converted .mlpackage depends
 on it staying byte-for-byte identical.
 """
 from typing import Iterable, Tuple
 ATTN_SUFFIX = {
    "SPLIT_EINSUM": "se",
    "SPLIT_EINSUM_V2": "se2",
    "ORIGINAL": "orig",
 }
 # Palettization bits. "none" = no quantization (default; keeps the
 # unquantized filename intact so existing workflows still resolve their
 # cached .mlpackage). Numeric values append a `_q<bits>` suffix.
 QUANT_NBITS_VALUES = ("none", "8", "6", "4")
 def compose_out_name(
    *,
    ckpt_name: str,
    batch_size: int,
    width: int,
    height: int,
    controlnet_support: bool,
    attention_implementation: str,
    lora_names: Iterable[str] = (),
    quantize_nbits: str = "none",
 ) -> str:
    """Build the .mlpackage stem from convert() parameters.
    Locked behaviour (characterization tests):
      - first '.' in ckpt_name wins (`a.b.c.safetensors` -> `a`)
      - spaces collapse to underscores
      - LoRA names are taken stem-only, sorted, joined with '_' and
        prefixed with '_' when present (caller is expected to pass a
        sorted list; we sort defensively)
      - controlnet adds `_cn`
      - attn suffix is `_se` | `_se2` | `_orig`
    Quantization:
      - quantize_nbits "none" (default) appends nothing — existing
        unquantized .mlpackages keep the old filename
      - "4" / "6" / "8" appends `_q<bits>` after the attn suffix
    """
    if quantize_nbits not in QUANT_NBITS_VALUES:
        raise ValueError(
            f"quantize_nbits={quantize_nbits!r} not in {QUANT_NBITS_VALUES}"
        )
    stem = ckpt_name.split(".")[0]
    sorted_names = sorted(lora_names)
    lora_str = "_" + "_".join(name.split(".")[0] for name in sorted_names) if sorted_names else ""
    cn_suffix = "_cn" if controlnet_support else ""
    attn_suffix = "_" + ATTN_SUFFIX[attention_implementation]
    quant_suffix = f"_q{quantize_nbits}" if quantize_nbits != "none" else ""
    out_name = (
        f"{stem}{lora_str}_{batch_size}x{width}x{height}"
        f"{cn_suffix}{attn_suffix}{quant_suffix}"
    )
    return out_name.replace(" ", "_")
 def lora_names_from_params(lora_params: Iterable[Tuple[str, float]]) -> list[str]:
    """Mirror the sort applied inside CoreMLConverter.convert."""
    return [name for name, _ in sorted(lora_params, key=lambda pair: pair[0])]
@@ -1,8 +1,3 @@
-"""LCM runtime support (sampler-side).
+from .nodes import COREML_CONVERT_LCM
-The dedicated LCM converter node was removed once the standard ``CoreMLConverter``
+__all__ = ["COREML_CONVERT_LCM"]
 gained model-version auto-detection (full-distill LCM is detected from the
 checkpoint). What remains here is runtime sampling support — ``utils`` patches the
 model sampling and supplies the guidance embedding when a converted UNet exposes
 ``timestep_cond``.
 """
@@ -0,0 +1,259 @@
 import os
 import logging
 import time
 import gc
 import numpy as np
 import torch
 from diffusers import UNet2DConditionModel, LCMScheduler
 from diffusers.loaders import LoraLoaderMixin
 from coreml_suite.conversion.attention import apply_attention_implementation
 from coreml_suite.conversion.shapes import conv2d_output_shape
 from coreml_suite.conversion.unet import CoreMLUNetWrapper
 from coreml_suite.model_version import ModelVersion
 import coremltools as ct
 logging.basicConfig()
 logger = logging.getLogger(__name__)
 logger.setLevel(logging.DEBUG)
 MODEL_VERSION = "SimianLuo/LCM_Dreamshaper_v7"
 MODEL_NAME = MODEL_VERSION.split("/")[-1] + "_4k"
 TEXT_TOKEN_SEQUENCE_LENGTH = 77
 def get_unets():
    ref_unet = UNet2DConditionModel.from_pretrained(
        MODEL_VERSION,
        subfolder="unet",
        device_map=None,
        low_cpu_mem_usage=False,
    )
    cml_unet = CoreMLUNetWrapper(
        apply_attention_implementation(ref_unet.eval(), "SPLIT_EINSUM"),
        ModelVersion.LCM,
    )
    return cml_unet, ref_unet
 def get_encoder_hidden_states_shape(unet_config, batch_size):
    encoder_hidden_states_shape = (
        batch_size,
        TEXT_TOKEN_SEQUENCE_LENGTH,
        unet_config.cross_attention_dim,
    )
    return encoder_hidden_states_shape
 def get_scheduler():
    from comfy.model_management import get_torch_device
    scheduler = LCMScheduler.from_pretrained(MODEL_VERSION, subfolder="scheduler")
    scheduler.set_timesteps(50, get_torch_device(), 50)
    return scheduler
 def get_coreml_inputs(sample_inputs):
    coreml_sample_unet_inputs = {
        k: v.numpy().astype(np.float16) for k, v in sample_inputs.items()
    }
    return [
        ct.TensorType(
            name=k,
            shape=v.shape,
            dtype=v.numpy().dtype if isinstance(v, torch.Tensor) else v.dtype,
        )
        for k, v in coreml_sample_unet_inputs.items()
    ]
 def load_coreml_model(out_path):
    logger.info(f"Loading model from {out_path}")
    start = time.time()
    coreml_model = ct.models.MLModel(out_path)
    logger.info(f"Loading {out_path} took {time.time() - start:.1f} seconds")
    return coreml_model
 def convert_to_coreml(
    submodule_name, torchscript_module, sample_inputs, output_names, out_path
 ):
    if os.path.exists(out_path):
        logger.info(f"Skipping export because {out_path} already exists")
        coreml_model = load_coreml_model(out_path)
    else:
        logger.info(f"Converting {submodule_name} to CoreML..")
        coreml_model = ct.convert(
            torchscript_module,
            convert_to="mlprogram",
            minimum_deployment_target=ct.target.macOS13,
            inputs=sample_inputs,
            outputs=[
                ct.TensorType(name=name, dtype=np.float32) for name in output_names
            ],
            skip_model_load=True,
        )
        del torchscript_module
        gc.collect()
    return coreml_model
 def get_out_path(submodule_name, model_name):
    from folder_paths import get_folder_paths
    fname = f"{model_name}_{submodule_name}.mlpackage"
    unet_path = get_folder_paths(submodule_name)[0]
    out_path = os.path.join(unet_path, fname)
    return out_path
 def get_sample_input(batch_size, encoder_hidden_states_shape, sample_shape, scheduler):
    sample_unet_inputs = dict(
        [
            ("sample", torch.rand(*sample_shape)),
            (
                "timestep",
                torch.tensor([scheduler.timesteps[0].item()] * batch_size).to(
                    torch.float32
                ),
            ),
            ("encoder_hidden_states", torch.rand(*encoder_hidden_states_shape)),
            ("timestep_cond", torch.randn(batch_size, 256).to(torch.float32)),
        ]
    )
    return sample_unet_inputs
 def get_unet_inputs_spec(sample_unet_inputs):
    sample_unet_inputs_spec = {
        k: (v.shape, v.dtype) for k, v in sample_unet_inputs.items()
    }
    return sample_unet_inputs_spec
 def add_cnet_support(sample_shape, reference_unet):
    additional_residuals_shapes = []
    batch_size = sample_shape[0]
    h, w = sample_shape[2:]
    # conv_in
    out_h, out_w = conv2d_output_shape(
        h,
        w,
        reference_unet.conv_in,
    )
    additional_residuals_shapes.append(
        (batch_size, reference_unet.conv_in.out_channels, out_h, out_w)
    )
    # down_blocks
    for down_block in reference_unet.down_blocks:
        additional_residuals_shapes += [
            (batch_size, resnet.out_channels, out_h, out_w)
            for resnet in down_block.resnets
        ]
        if hasattr(down_block, "downsamplers") and down_block.downsamplers is not None:
            for downsampler in down_block.downsamplers:
                out_h, out_w = conv2d_output_shape(out_h, out_w, downsampler.conv)
            additional_residuals_shapes.append(
                (
                    batch_size,
                    down_block.downsamplers[-1].conv.out_channels,
                    out_h,
                    out_w,
                )
            )
    # mid_block
    additional_residuals_shapes.append(
        (batch_size, reference_unet.mid_block.resnets[-1].out_channels, out_h, out_w)
    )
    additional_inputs = {}
    for i, shape in enumerate(additional_residuals_shapes):
        sample_residual_input = torch.rand(*shape)
        additional_inputs[f"additional_residual_{i}"] = sample_residual_input
    return additional_inputs
 def convert(
    out_path: str,
    batch_size: int = 1,
    sample_size: tuple[int, int] = (64, 64),
    controlnet_support: bool = False,
    lora_paths: list[str] = None,
 ):
    lora_paths = lora_paths or []
    coreml_unet, ref_unet = get_unets()
    for lora_path in lora_paths:
        lora_sd, network_alphas = LoraLoaderMixin.lora_state_dict(lora_path)
        LoraLoaderMixin.load_lora_into_unet(lora_sd, network_alphas, ref_unet)
        ref_unet.fuse_lora()
    sample_shape = (
        batch_size,  # B
        ref_unet.config.in_channels,  # C
        sample_size[0],  # H
        sample_size[1],  # W
    )
    encoder_hidden_states_shape = get_encoder_hidden_states_shape(
        ref_unet.config, batch_size
    )
    scheduler = get_scheduler()
    sample_inputs = get_sample_input(
        batch_size, encoder_hidden_states_shape, sample_shape, scheduler
    )
    if controlnet_support:
        sample_inputs |= add_cnet_support(sample_shape, ref_unet)
    sample_inputs_spec = get_unet_inputs_spec(sample_inputs)
    logger.info(f"Sample UNet inputs spec: {sample_inputs_spec}")
    logger.info("JIT tracing..")
    traced_unet = torch.jit.trace(
        coreml_unet, example_inputs=list(sample_inputs.values())
    )
    logger.info("Done.")
    coreml_sample_inputs = get_coreml_inputs(sample_inputs)
    coreml_unet = convert_to_coreml(
        "unet", traced_unet, coreml_sample_inputs, ["noise_pred"], out_path
    )
    del traced_unet
    gc.collect()
    coreml_unet.save(out_path)
    logger.info(f"Saved unet into {out_path}")
 if __name__ == "__main__":
    h = 512
    w = 512
    sample_size = (h // 8, w // 8)
    batch_size = 4
    cn_support_str = "_cn" if True else ""
    out_name = f"{MODEL_NAME}_{batch_size}x{w}x{h}{cn_support_str}"
    out_path = get_out_path("unet", f"{out_name}")
    if not os.path.exists(out_path):
        convert(out_path=out_path, sample_size=sample_size, batch_size=batch_size)
@@ -0,0 +1,70 @@
 import os
 from coremltools import ComputeUnit
 from coreml_suite import COREML_NODE
 from coreml_suite.coreml_model import CoreMLModel
 class COREML_CONVERT_LCM(COREML_NODE):
    """Converts a LCM model to Core ML."""
    @classmethod
    def INPUT_TYPES(cls):
        return {
            "required": {
                "height": ("INT", {"default": 512, "min": 512, "max": 768, "step": 8}),
                "width": ("INT", {"default": 512, "min": 512, "max": 768, "step": 8}),
                "batch_size": ("INT", {"default": 1, "min": 1, "max": 64}),
                "compute_unit": (
                    [
                        ComputeUnit.CPU_AND_NE.name,
                        ComputeUnit.CPU_AND_GPU.name,
                        ComputeUnit.ALL.name,
                        ComputeUnit.CPU_ONLY.name,
                    ],
                ),
                "controlnet_support": ("BOOLEAN", {"default": False}),
            }
        }
    RETURN_TYPES = ("COREML_UNET",)
    RETURN_NAMES = ("coreml_model",)
    FUNCTION = "convert"
    def convert(self, height, width, batch_size, compute_unit, controlnet_support):
        """Converts a LCM model to Core ML.
        Args:
            height (int): Height of the target image.
            width (int): Width of the target image.
            batch_size (int): Batch size.
            compute_unit (str): Compute unit to use when loading the model.
        Returns:
            coreml_model: The converted Core ML model.
        The converted model is also saved to "models/unet" directory and
        can be loaded with the "LCMCoreMLLoaderUNet" node.
        """
        from coreml_suite.lcm import converter as lcm_converter
        h = height
        w = width
        sample_size = (h // 8, w // 8)
        batch_size = batch_size
        cn_support_str = "_cn" if controlnet_support else ""
        out_name = f"{lcm_converter.MODEL_NAME}_{batch_size}x{w}x{h}{cn_support_str}"
        out_path = lcm_converter.get_out_path("unet", f"{out_name}")
        if not os.path.exists(out_path):
            lcm_converter.convert(
                out_path=out_path,
                sample_size=sample_size,
                batch_size=batch_size,
                controlnet_support=controlnet_support,
            )
        return (CoreMLModel(out_path, compute_unit),)
@@ -0,0 +1,98 @@
 from diffusers import UNet2DConditionModel
 from diffusers.models.embeddings import TimestepEmbedding
 class UNet2DConditionModelLCM(UNet2DConditionModel):
    def __init__(
        self,
        time_cond_proj_dim=None,
        **kwargs,
    ):
        super().__init__(**kwargs)
        timestep_input_dim = self.config.block_out_channels[0]
        time_embed_dim = self.config.block_out_channels[0] * 4
        time_embedding = TimestepEmbedding(
            timestep_input_dim, time_embed_dim, cond_proj_dim=time_cond_proj_dim
        )
        self.time_embedding = time_embedding
    def forward(
        self,
        sample,
        timestep,
        encoder_hidden_states,
        timestep_cond,
        *additional_residuals,
    ):
        # 0. Project (or look-up) time embeddings
        t_emb = self.time_proj(timestep)
        emb = self.time_embedding(t_emb, timestep_cond)
        # 1. center input if necessary
        if self.config.center_input_sample:
            sample = 2 * sample - 1.0
        # 2. pre-process
        sample = self.conv_in(sample)
        # 3. down
        down_block_res_samples = (sample,)
        for downsample_block in self.down_blocks:
            if (
                hasattr(downsample_block, "attentions")
                and downsample_block.attentions is not None
            ):
                sample, res_samples = downsample_block(
                    hidden_states=sample,
                    temb=emb,
                    encoder_hidden_states=encoder_hidden_states,
                )
            else:
                sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
            down_block_res_samples += res_samples
        if additional_residuals:
            new_down_block_res_samples = ()
            for i, down_block_res_sample in enumerate(down_block_res_samples):
                down_block_res_sample = down_block_res_sample + additional_residuals[i]
                new_down_block_res_samples += (down_block_res_sample,)
            down_block_res_samples = new_down_block_res_samples
        # 4. mid
        sample = self.mid_block(
            sample, emb, encoder_hidden_states=encoder_hidden_states
        )
        if additional_residuals:
            sample = sample + additional_residuals[-1]
        # 5. up
        for upsample_block in self.up_blocks:
            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
            down_block_res_samples = down_block_res_samples[
                : -len(upsample_block.resnets)
            ]
            if (
                hasattr(upsample_block, "attentions")
                and upsample_block.attentions is not None
            ):
                sample = upsample_block(
                    hidden_states=sample,
                    temb=emb,
                    res_hidden_states_tuple=res_samples,
                    encoder_hidden_states=encoder_hidden_states,
                )
            else:
                sample = upsample_block(
                    hidden_states=sample, temb=emb, res_hidden_states_tuple=res_samples
                )
        # 6. post-process
        sample = self.conv_norm_out(sample)
        sample = self.conv_act(sample)
        sample = self.conv_out(sample)
        return (sample,)
@@ -0,0 +1,8 @@
 from enum import Enum
 class ModelVersion(Enum):
    SD15 = "sd15"
    SDXL = "sdxl"
    SDXL_REFINER = "sdxl_refiner"
    LCM = "lcm"
@@ -4,9 +4,16 @@ from coremltools import ComputeUnit
 import folder_paths
 from coreml_suite import COREML_NODE
 from coreml_suite.attention import ATTENTION_IMPLEMENTATIONS
 from coreml_suite.coreml_model import CoreMLModel
 from coreml_suite.core.naming import (
    QUANT_NBITS_VALUES,
    compose_out_name,
    lora_names_from_params,
 )
 from coreml_suite.lcm.utils import add_lcm_model_options, lcm_patch, is_lcm
 from coreml_suite.logger import logger
 from coreml_suite.model_version import ModelVersion
 from nodes import KSampler, LoraLoader, KSamplerAdvanced
 from coreml_suite.models import (
@@ -17,26 +24,6 @@ from coreml_suite.models import (
 )
 def _discover(fn_name, fallback):
    """Populate a converter dropdown from coreml_diffusion's discovery API.
    Fails soft: if the package is missing, too old to expose ``fn_name``, or
    errors, the node still registers with the fallback list instead of vanishing
    from the menu. Evaluated on every INPUT_TYPES call, so installing a newer
    coreml_diffusion surfaces new conversion types with no Suite change.
    """
    try:
        import coreml_diffusion
        return getattr(coreml_diffusion, fn_name)()
    except Exception as exc:  # missing/old package, import error, etc.
        logger.warning(
            f"coreml_diffusion.{fn_name} unavailable ({exc}); "
            f"using fallback {fallback}"
        )
        return fallback
 class CoreMLSampler(COREML_NODE, KSampler):
    @classmethod
    def INPUT_TYPES(s):
@@ -225,26 +212,24 @@ class CoreMLModelAdapter(COREML_NODE):
 class CoreMLConverter(COREML_NODE):
-    """Converts a Stable Diffusion checkpoint (UNet) to Core ML.
+    """Converts a LCM model to Core ML."""
    The model version (SD15 / SDXL / SDXL refiner / LCM) is auto-detected from
    the checkpoint's architecture, so there is no version dropdown — one node
    converts every supported family, including full-distill LCM.
    """
    @classmethod
    def INPUT_TYPES(cls):
        return {
            "required": {
                "ckpt_name": (folder_paths.get_filename_list("checkpoints"),),
                "model_version": (
                    [
                        ModelVersion.SD15.name,
                        ModelVersion.SDXL.name,
                    ],
                ),
                "height": ("INT", {"default": 512, "min": 8, "step": 8}),
                "width": ("INT", {"default": 512, "min": 8, "step": 8}),
                "batch_size": ("INT", {"default": 1, "min": 1, "max": 64}),
                "attention_implementation": (
-                    _discover(
+                    list(ATTENTION_IMPLEMENTATIONS),
                        "list_attention_impls",
                        ["SPLIT_EINSUM", "SPLIT_EINSUM_V2", "ORIGINAL"],
                    ),
                ),
                "compute_unit": (
                    [
@@ -262,10 +247,7 @@ class CoreMLConverter(COREML_NODE):
                # omits any `required` input. When omitted it defaults to
                # "none", identical to unquantized behavior and filename, so
                # existing cached .mlpackages still resolve.
-                "quantize_nbits": (
+                "quantize_nbits": (list(QUANT_NBITS_VALUES), {"default": "none"}),
                    _discover("list_quant_modes", ["none", "8", "6", "4"]),
                    {"default": "none"},
                ),
                "lora_params": ("LORA_PARAMS",),
            },
        }
@@ -277,6 +259,7 @@ class CoreMLConverter(COREML_NODE):
    def convert(
        self,
        ckpt_name,
        model_version,
        height,
        width,
        batch_size,
@@ -286,11 +269,9 @@ class CoreMLConverter(COREML_NODE):
        quantize_nbits="none",
        lora_params=None,
    ):
-        """Converts a checkpoint's UNet to Core ML.
+        """Converts a LCM model to Core ML.
        Args:
            ckpt_name (str): Checkpoint to convert; its model version is
                auto-detected from the weights.
            height (int): Height of the target image.
            width (int): Width of the target image.
            batch_size (int): Batch size.
@@ -300,8 +281,10 @@ class CoreMLConverter(COREML_NODE):
            coreml_model: The converted Core ML model.
        The converted model is also saved to "models/unet" directory and
-        can be loaded with the "Load Core ML UNet" node.
+        can be loaded with the "LCMCoreMLLoaderUNet" node.
        """
        model_version = ModelVersion[model_version]
        lora_params = lora_params or {}
        lora_params = [(k, v[0]) for k, v in lora_params.items()]
        lora_params = sorted(lora_params, key=lambda lora: lora[0])
@@ -310,16 +293,14 @@ class CoreMLConverter(COREML_NODE):
        h = height
        w = width
        sample_size = (h // 8, w // 8)
-        import coreml_diffusion
+        out_name = compose_out_name(
        out_name = coreml_diffusion.compose_out_name(
            ckpt_name=ckpt_name,
            batch_size=batch_size,
            width=w,
            height=h,
            controlnet_support=controlnet_support,
            attention_implementation=attention_implementation,
-            lora_names=coreml_diffusion.lora_names_from_params(lora_params),
+            lora_names=lora_names_from_params(lora_params),
            quantize_nbits=quantize_nbits,
        )
@@ -330,14 +311,13 @@ class CoreMLConverter(COREML_NODE):
        logger.info(f"Attention implementation: {attention_implementation}")
        if lora_params:
-            logger.info("LoRAs used:")
+            logger.info(f"LoRAs used:")
            for lora_param in lora_params:
                logger.info(f"  {lora_param[0]} - strength: {lora_param[1]}")
-        # Resolve the ComfyUI models/unet path here (a node concern); the package
+        from coreml_suite import converter
-        # takes the output path as an injected argument.
+
-        unet_path = folder_paths.get_folder_paths("unet")[0]
+        unet_out_path = converter.get_out_path("unet", f"{out_name}")
        unet_out_path = os.path.join(unet_path, f"{out_name}_unet.mlpackage")
        ckpt_path = folder_paths.get_full_path("checkpoints", ckpt_name)
        config_filename = ckpt_name.split(".")[0] + ".yaml"
@@ -345,10 +325,10 @@ class CoreMLConverter(COREML_NODE):
        if config_path:
            logger.info(f"Using config file {config_path}")
-        coreml_diffusion.convert(
+        converter.convert(
-            ckpt_path,
+            ckpt_path=ckpt_path,
-            None,  # model_version auto-detected from the checkpoint
+            model_version=model_version,
-            unet_out_path,
+            unet_out_path=unet_out_path,
            sample_size=sample_size,
            batch_size=batch_size,
            controlnet_support=controlnet_support,
@@ -1,34 +1,27 @@
 [build-system]
 requires = ["hatchling"]
 build-backend = "hatchling.build"
 [project]
 name = "comfyui-coremlsuite"
 description = "This extension contains a set of custom nodes for ComfyUI that allow you to use Core ML models in your ComfyUI workflows."
-version = "2.1.2"
+version = "2.0.2"
 license = "MIT"
-requires-python = ">=3.12"
+requires-python = ">=3.12,<3.13"
 packages = [{ include = "coreml_suite" }]
 dependencies = [
-    # torch is provided by the host (ComfyUI) and intentionally left unpinned
+    "torch>=2.7,<2.8",
    # here: a hard torch cap would downgrade the host's torch and break its
    # torchvision/torchaudio ABI. coreml-diffusion pulls torch>=2.7 transitively.
    # >=0.1.6: model-version auto-detection (convert(model_version=None)) and the
    # dropped <3.13 Python cap (kept in sync with this package's requires-python).
    "coreml-diffusion>=0.1.6,<0.2",
    "coremltools>=9,<10",
    "numpy>=2,<3",
    "diffusers>=0.30",
    "peft>=0.13",
    "omegaconf>=2.3",
    "transformers>=4.44",
 ]
 [project.urls]
 Repository = "https://github.com/aszc-dev/ComfyUI-CoreMLSuite"
 [tool.hatch.build.targets.wheel]
 packages = ["coreml_suite"]
 [tool.comfy]
 PublisherId = "aszc-dev"
 DisplayName = "ComfyUI-CoreMLSuite"
-Icon = "https://raw.githubusercontent.com/aszc-dev/ComfyUI-CoreMLSuite/main/assets/snake.png"
+Icon = ""
 requires-comfyui = ">=0.3.27"
 [dependency-groups]
@@ -1,4 +1,7 @@
-coreml-diffusion>=0.1.4,<0.2
+torch>=2.7,<2.8
 coremltools>=9,<10
 numpy>=2,<3
 diffusers>=0.30
 peft>=0.13
 omegaconf>=2.3
 transformers>=4.44
@@ -1,208 +0,0 @@
 # Conversion Extraction — Seam Inventory (`docs/extraction/seam.md`)
 > **Gate E0 deliverable.** Symbol-by-symbol cut line between the future `coreml_diffusion`
 > package (CONVERSION) and what stays in `coreml_suite` (the ComfyUI side).
 >
 > **Confidence legend:**
 > - ✅ **verified** — read directly from the current source in this repo.
 > - 🔍 **confirm** — inferred / partially seen; Claude Code must `grep`-verify before acting.
 >
 > **Cut rule:** a symbol goes to `coreml_diffusion` iff it participates in producing the `.mlpackage`
 > artifact AND can be made free of `comfy` / `folder_paths` / `comfy_extras`. The runtime
 > *loader* that **runs** a compiled model stays in the suite.
 ---
 ## 1. File-level map
 | File | Side | Status | Note |
 |---|---|---|---|
 | `coreml_suite/model_version.py` | **coreml_diffusion** | ✅ | Already `Enum`-only, zero comfy. Becomes pkg source of truth. |
 | `coreml_suite/attention.py` | **coreml_diffusion** | ✅ | `ATTENTION_IMPLEMENTATIONS` tuple; pure constant. |
 | `coreml_suite/core/naming.py` | **coreml_diffusion** | ✅ | `compose_out_name` = cache-key contract. Move (not copy). |
 | `coreml_suite/converter.py` | **coreml_diffusion** (mostly) | ✅ | Main conversion. One symbol stays-adjacent: `get_out_path` (folder_paths) is replaced by injected `out_path`. |
 | `coreml_suite/conversion/attention.py` | **coreml_diffusion** | ✅ | `apply_attention_implementation`. Imports `logging`,`torch` only — no comfy. |
 | `coreml_suite/conversion/shapes.py` | **coreml_diffusion** | ✅ | `conv2d_output_shape`. Pure math, no imports. |
 | `coreml_suite/conversion/trace.py` | **coreml_diffusion** | ✅ | Imports `types.MethodType`, `diffusers...Transformer2DModel` only — torch/diffusers. |
 | `coreml_suite/conversion/unet.py` | **coreml_diffusion** | ✅ | `CoreMLUNetWrapper`. Imports `torch` only — no comfy. |
 | `coreml_suite/lcm/converter.py` | **coreml_diffusion** (after dedup) | ✅ | Dup helpers deleted; `MODEL_VERSION` HF-hardcode (L22) → E-LCM. `folder_paths` (L111) + `comfy.model_management` (L54) confirmed present → CUT. |
 | `coreml_suite/lcm/unet.py` | **coreml_diffusion** | ✅ | `UNet2DConditionModelLCM(UNet2DConditionModel)`. diffusers-only, no comfy. |
 | `coreml_suite/config.py` | **STAYS** | ✅ | Imports `comfy.supported_models_base`/`latent_formats`/`model_detection`. **Inference-side** (`get_model_config`), NOT conversion. |
 | `coreml_suite/coreml_model.py` | **STAYS** | ✅ | `CoreMLModel` = runtime loader (runs `.mlpackage`). Desktop/Python inference; not used on iOS. |
 | `coreml_suite/nodes.py` | **STAYS** | ✅ | Nodes; will call `coreml_diffusion` + own `folder_paths` path resolution + discovery dropdowns. |
 | `coreml_suite/lcm/nodes.py` | **STAYS** | ✅ | `COREML_CONVERT_LCM` node. |
 | `coreml_suite/models.py` | **STAYS** | ✅ | Inference: `add_sdxl_model_options`, `is_sdxl`, `get_model_patcher`, `get_latent_image`. |
 | `coreml_suite/latents.py` | **STAYS** | ✅ | Inference chunking (MODERNIZATION Phase 3 target, not this spec). |
 | `coreml_suite/controlnet.py` | **STAYS** | ✅ | Inference-side controlnet. Distinct from converter `add_cnet_support`. |
 | `coreml_suite/lcm/utils.py` | **STAYS** | ✅ | `add_lcm_model_options`, `lcm_patch`, `is_lcm`; imports `comfy_extras`. Inference. |
 | `coreml_suite/logger.py` | **both / copy** | ✅ | Trivial. Package gets its own logger; suite keeps its. |
 ---
 ## 2. Symbol-level: `coreml_suite/converter.py` (main conversion)
 | Symbol | Side | Status | Cut action |
 |---|---|---|---|
 | `DEFAULT_TRACE_TIMESTEP`, `TEXT_TOKEN_SEQUENCE_LENGTH` | coreml_diffusion | ✅ | Move as-is (module constants). |
 | `get_unet(model_version, ref_unet, attention_implementation)` | coreml_diffusion | ✅ | Move. Uses `conversion.{trace,attention,unet}`. No comfy. |
 | `get_encoder_hidden_states_shape(ref_unet, batch_size)` | coreml_diffusion | ✅ | Move. Reads `ref_unet.config.cross_attention_dim`. Pure. |
 | `get_coreml_inputs(sample_inputs)` | coreml_diffusion | ✅ | Move. `ct.TensorType` build. |
 | `load_coreml_model(out_path)` | coreml_diffusion | ✅ | Move. `ct.models.MLModel(out_path)`. (Dedup target vs LCM copy.) |
 | `convert_to_coreml(submodule, ts_module, inputs, names, out_path)` | coreml_diffusion | ✅ | Move. `ct.convert(...)`. (Dedup target vs LCM copy.) |
 | `get_sample_input(batch, ehs_shape, sample_shape)` | coreml_diffusion | ✅ | Move. **Merge** with LCM variant (LCM passes extra `scheduler` → optional param). |
 | `lcm_inputs(sample_unet_inputs)` | coreml_diffusion | ✅ | Move. Adds `timestep_cond`. |
 | `sdxl_inputs(sample_unet_inputs, ref_unet, model_version)` | coreml_diffusion | ✅ | Move. `time_ids`/`text_embeds`/`add_embeds`. |
 | `add_cnet_support(sample_shape, ref_unet)` | coreml_diffusion | ✅ | Move. Builds `additional_residual_*` inputs from unet block channels. |
 | `convert_unet(ref_unet, model_version, unet_out_path, ...)` | coreml_diffusion | ✅ | Move. Orchestrates trace→convert→**quant (palettize)**→save. Quant travels here (E6). |
 | `convert(ckpt_path, model_version, unet_out_path, ...)` | coreml_diffusion | ✅ | Move. **Make kw-only past `ckpt_path,model_version,out_path`** (contract). Validates `attn_impl`. |
 | `load_unet(ckpt_path, config_path)` | coreml_diffusion | ✅ | Move. `UNet2DConditionModel.from_single_file`. |
 | `get_out_path(submodule_name, model_name)` | **STAYS (node)** | ✅ | Uses `folder_paths.get_folder_paths`. **Delete from converter; node resolves path and passes `out_path` in.** |
 **Apple `python_coreml_stable_diffusion` footprint on this path:** ✅ **none.** Verified by grep:
 zero imports in `converter.py` / `conversion/*`. Main path uses `diffusers` +
 local `CoreMLUNetWrapper`. (And the runtime `CoreMLModel` is now a local coremltools wrapper too —
 see §6 stale-spec note.)
 ---
 ## 3. Symbol-level: `coreml_suite/lcm/converter.py` (LCM — dedup + defer)
 | Symbol | Side | Status | Cut action |
 |---|---|---|---|
 | `load_coreml_model` (LCM copy) | DELETE | ✅ | Duplicate of main. Remove; use `coreml_diffusion.load_coreml_model`. |
 | `convert_to_coreml` (LCM copy) | DELETE | ✅ | Duplicate of main. Remove. |
 | `get_out_path` (LCM copy, folder_paths) | DELETE | ✅ | Duplicate + comfy. Remove; node injects `out_path`. |
 | `get_sample_input(..., scheduler)` (LCM copy) | MERGE → coreml_diffusion | ✅ | Fold `scheduler` into shared `get_sample_input` as optional param. |
 | `MODEL_NAME` (= LCM_Dreamshaper) | **E-LCM** | ✅ | HF hardcode. Removing it is the behavior change → E-LCM, not E2. |
 | `convert(out_path, sample_size, batch_size, controlnet_support)` (LCM, L190) | coreml_diffusion (via unified) | ✅ | Route through `coreml_diffusion.convert(model_version=LCM, ...)` in E-LCM. |
 | `from comfy.model_management import get_torch_device` (L54, in `get_scheduler`) | **CUT** | ✅ | Confirmed present. Inject `device`. |
 | module-global attention set at import | n/a | ✅ | **No module global.** Attention already per-call: `get_unets` (L36) calls `apply_attention_implementation(ref_unet, "SPLIT_EINSUM")`. No `ATTENTION_IMPLEMENTATION_IN_EFFECT` anywhere in repo. (Note: LCM hardcodes `"SPLIT_EINSUM"` — pass `attn_impl` through in dedup.) |
 ---
 ## 4. Symbol-level: `coreml_suite/core/naming.py` → `coreml_diffusion/naming.py`
 | Symbol | Side | Status | Cut action |
 |---|---|---|---|
 | `compose_out_name(...)` | coreml_diffusion | ✅ | **Move** (cache-key contract). Node imports from pkg. |
 | `lora_names_from_params(...)` | coreml_diffusion | ✅ | Move. |
 | `ATTN_SUFFIX` dict | coreml_diffusion | ✅ | Move. |
 | `QUANT_NBITS_VALUES` | coreml_diffusion | ✅ | Move; backs `list_quant_modes()`. |
 | `tests/unit/test_characterization_out_name.py` | re-point | ✅ | Change import to `coreml_diffusion.naming`. Assertions/values **unchanged**. |
 ---
 ## 5. Discovery API + status registry (new in `coreml_diffusion/__init__.py`)
 ```python
 from enum import Enum
 class Status(Enum):
    VERIFIED = "verified"          # has a golden anchor + passing [M2-ANE] check
    EXPERIMENTAL = "experimental"  # convertible, not yet anchored/verified
 # Single source of truth. Suite gates on this, NOT on a hardcoded node list.
 # KEY by ModelVersion enum MEMBER (not a bare string) so list_model_versions can
 # emit .name — see the .name decision below. Keying by the lowercase .value string
 # (as an earlier draft of this block did) returns ["sd15",...], which the node then
 # reverses via ModelVersion[...] → KeyError. Do NOT key by .value.
 _MODEL_STATUS = {
    ModelVersion.SD15:         Status.VERIFIED,
    ModelVersion.SDXL:         Status.VERIFIED,
    ModelVersion.SDXL_REFINER: Status.EXPERIMENTAL,   # → VERIFIED after a refiner golden anchor
    ModelVersion.LCM:          Status.EXPERIMENTAL,   # → VERIFIED after E-LCM golden anchor
 }
 def list_model_versions(include_experimental: bool = False) -> list[str]:
    return [v.name for v, s in _MODEL_STATUS.items()   # .name → "SD15","SDXL" (see decision)
            if s is Status.VERIFIED or (include_experimental and s is Status.EXPERIMENTAL)]
 def list_attention_impls() -> list[str]:   # from attention.ATTENTION_IMPLEMENTATIONS
    ...
 def list_quant_modes() -> list[str]:       # from naming.QUANT_NBITS_VALUES
    ...
 CONTRACT_VERSION = "1.0"
 # Additive-only: adding an id or promoting EXPERIMENTAL→VERIFIED = minor bump (Suite unaffected).
 # Removing/renaming an id, or demoting VERIFIED→EXPERIMENTAL = MAJOR bump + migration note.
 ```
 **Decision check (`.name` vs `.value`): RESOLVED → `.name`.** ✅
 Verified in current source:
 - Node renders `ModelVersion.SD15.name` / `ModelVersion.SDXL.name` → `"SD15"`, `"SDXL"`
  (`nodes.py:224-225`).
 - Node reverses the dropdown string with `model_version = ModelVersion[model_version]`
  (`nodes.py:286`) — i.e. **lookup by NAME**. Feeding it a `.value` (`"sd15"`) raises `KeyError`.
 - Enum values are lowercase (`model_version.py`: `SD15="sd15"`, `SDXL="sdxl"`,
  `SDXL_REFINER="sdxl_refiner"`, `LCM="lcm"`).
 - `compose_out_name` does NOT consume the model_version string (grep of `core/naming.py` empty) —
  no coupling there, so no constraint from that side.
 **Decision:** `list_model_versions()` returns `.name` (uppercase). Saved workflows store `"SD15"`,
 node already validates them via `ModelVersion[...]`. The `_MODEL_STATUS` block above was corrected
 to key by enum member and emit `.name`. **The earlier `v.value` form was a latent bug.**
 ---
 ## 6. `python_coreml_stable_diffusion` split (Gate E0 line to fill by grep)
 | Use | Side | Status |
 |---|---|---|
 | `coreml_model.CoreMLModel` (runs compiled model) | **STAYS** (suite runtime) | ✅ — **local class**, not Apple's |
 | `unet.UNet2DConditionModel*` internals | **gone** — `converter.py:319` uses `diffusers.UNet2DConditionModel.from_single_file` | ✅ |
 | `AttentionImplementations` enum | gone — local `apply_attention_implementation` + `attention.py` tuple | ✅ |
 | `calculate_conv2d_output_shape` | gone — replaced by `conversion/shapes.conv2d_output_shape` | ✅ |
 > ### ⚠️ SPEC IS STALE: `ml-stable-diffusion` is already fully removed
 > Commit #58 ("replace apple/ml-stable-diffusion with native diffusers conversion") already did
 > the de-Apple work. Verified now:
 > - **Zero** `python_coreml_stable_diffusion` runtime imports anywhere in `coreml_suite` (only a
 >   docstring mention at `core/__init__.py:4`).
 > - `coreml_suite/coreml_model.py:8` `CoreMLModel` is a **local** wrapper over
 >   `coremltools.models.MLModel` (`coreml_model.py:22`) — it does **not** import Apple's class.
 > - `ml-stable-diffusion` / `python_coreml_stable_diffusion` appears in **neither** `pyproject.toml`
 >   **nor** `requirements.txt`. It is not a dependency at all.
 >
 > **Consequences for the spec (correct these in CONVERTER_EXTRACTION_SPEC.md):**
 > - §0.3 premise ("runtime loader = `python_coreml_stable_diffusion.coreml_model.CoreMLModel`,
 >   stays in suite") is **wrong**: the loader is already the local `coreml_model.CoreMLModel`. The
 >   "stays in suite" conclusion still holds; the identity does not.
 > - **Gate E0 item "ml-stable-diffusion pinned SHA — BLOCKER if unpinned" is MOOT** — there is no
 >   such dep to pin. Mark it N/A, not BLOCKER.
 > - **E4/E5 dependency lists must drop `git+...ml-stable-diffusion@<sha>`.** Package runtime deps
 >   are: `coremltools`, `diffusers`, `peft` (LoRA), `omegaconf` (config), `numpy`, `torch`. Confirm
 >   `peft`/`omegaconf` actually used before listing (grep at E4).
 > - The "keep `python_coreml_stable_diffusion` as a suite dep for the loader" instruction in E5 is
 >   **void** — coremltools backs the loader.
 ---
 ## 7. Pre-flight checklist before E1 (run these greps)
 ```
 grep -rn "import comfy"            coreml_suite/conversion coreml_suite/converter.py coreml_suite/lcm/converter.py coreml_suite/lcm/unet.py
 grep -rn "folder_paths"            coreml_suite/converter.py coreml_suite/lcm/converter.py
 grep -rn "model_management"        coreml_suite/lcm
 grep -rn "python_coreml_stable_diffusion" coreml_suite
 grep -rn "ATTENTION_IMPLEMENTATION_IN_EFFECT" coreml_suite
 grep -rn "SimianLuo\|LCM_Dreamshaper" coreml_suite/lcm
 ```
 Every 🔍 above resolves to ✅ or a correction once these run. Do not start moving code (E2)
 with any 🔍 unresolved on the CONVERSION side.
 **STATUS (run 2026-05-26): all 🔍 resolved.** Summary of what the greps found:
 - `conversion/*`, `lcm/unet.py`: comfy-free (torch/diffusers only). ✅
 - `converter.py`: only comfy reach-in is `folder_paths` in `get_out_path` (L91-94) → inject `out_path`.
 - `lcm/converter.py`: `folder_paths` (L111-114) + `comfy.model_management.get_torch_device` (L54)
  → cut both. Dup helpers (`load_coreml_model`,`convert_to_coreml`,`get_out_path`,`get_sample_input`)
  confirmed → dedup E2. `MODEL_VERSION="SimianLuo/LCM_Dreamshaper_v7"` (L22) → E-LCM.
 - No attention module-global anywhere (`ATTENTION_IMPLEMENTATION_IN_EFFECT` absent); already per-call.
  LCM hardcodes `"SPLIT_EINSUM"` in `get_unets` — thread `attn_impl` through during dedup.
 - `.name` vs `.value`: **decided `.name`** (node reverses via `ModelVersion[...]`). §5 corrected.
 - `ml-stable-diffusion`: **already gone** (#58). §6 stale-spec note added — fix the spec's E0/E4/E5
  dep + pinning items.
 Two grep blind-spots to note (the checklist above doesn't cover them, but cheap to add): the
 `folder_paths` grep only scans the two converter files — also grep `coreml_suite/lcm/utils.py`
 (it imports `comfy.model_management` at L3, but it's inference/STAYS, so fine) and confirm no other
 `conversion/` file grew a comfy import since.
@@ -107,6 +107,7 @@
  "10": {
    "inputs": {
      "ckpt_name": "dreamshaper_8.safetensors",
      "model_version": "SD15",
      "height": 512,
      "width": 512,
      "batch_size": 1,
@@ -0,0 +1,41 @@
 import platform
 import pytest
 import torch
 from diffusers.models.attention_processor import Attention, AttnProcessor
 from coreml_suite.conversion.attention import (
    SplitEinsumAttnProcessor,
    SplitEinsumV2AttnProcessor,
 )
 pytestmark = pytest.mark.skipif(
    platform.system() != "Darwin" or platform.machine() != "arm64",
    reason="Tier 1 requires macOS on Apple Silicon",
 )
@pytest.mark.parametrize(
    "processor",
    [
        SplitEinsumAttnProcessor(),
        SplitEinsumV2AttnProcessor(),
    ],
 )
 def test_split_einsum_processor_matches_diffusers_attention(processor):
    torch.manual_seed(0)
    reference = Attention(query_dim=32, heads=4, dim_head=8, dropout=0.0)
    reference.set_processor(AttnProcessor())
    candidate = Attention(query_dim=32, heads=4, dim_head=8, dropout=0.0)
    candidate.load_state_dict(reference.state_dict())
    candidate.set_processor(processor)
    hidden_states = torch.randn(2, 17, 32)
    encoder_hidden_states = torch.randn(2, 11, 32)
    expected = reference(hidden_states, encoder_hidden_states=encoder_hidden_states)
    actual = candidate(hidden_states, encoder_hidden_states=encoder_hidden_states)
    assert torch.allclose(actual, expected, atol=1e-5)
@@ -0,0 +1,138 @@
 """Tier 1 smoke: convert a synthetic micro-UNet through coremltools and load
 it back with CoreMLSuite's runtime CoreMLModel wrapper.
 Purpose: catch API breakage in coremltools *without* needing a real SD
 checkpoint, the ANE, or a converted .mlmodelc on disk.
 Runs in minutes on a hosted macOS-ARM runner (no Apple internal stuff).
 What it asserts:
  - coremltools.convert still accepts the call shape we use today
  - the resulting .mlpackage round-trips through CoreMLSuite's CoreMLModel
  - expected_inputs exposes the input names/shapes we declared
  - calling the model returns the named output (`noise_pred`)
 Auto-skips on non-Apple-Silicon hosts so Tier 0 CI on Linux ignores it.
 """
 import platform
 import shutil
 from types import SimpleNamespace
 import numpy as np
 import pytest
 import torch
 import torch.nn as nn
 from coreml_suite.conversion.unet import CoreMLUNetWrapper
 pytestmark = pytest.mark.skipif(
    platform.system() != "Darwin" or platform.machine() != "arm64",
    reason="Tier 1 requires macOS on Apple Silicon",
 )
 # Tiny shapes — large enough to exercise conv2d + linear + addition kernels in
 # coremltools, small enough that conversion finishes in seconds on CPU.
 SAMPLE_SHAPE = (1, 4, 8, 8)
 TIMESTEP_SHAPE = (1,)
 ENCODER_SHAPE = (1, 4, 64)  # native diffusers encoder_hidden_states (batch, tokens, hidden)
 OUT_NAME = "noise_pred"
 class TinyUNet(nn.Module):
    """Minimal UNet-shaped graph: conv -> add(time+context) -> conv.
    Not a real diffusion model. Just enough op variety to exercise the
    PyTorch -> MIL frontend in coremltools and confirm we can still wire
    the inputs/outputs the way CoreMLSuite's runtime expects.
    """
    def __init__(self):
        super().__init__()
        self.conv_in = nn.Conv2d(4, 8, kernel_size=3, padding=1)
        self.conv_out = nn.Conv2d(8, 4, kernel_size=3, padding=1)
        self.time_proj = nn.Linear(1, 8)
        self.text_proj = nn.Linear(64, 8)
    def forward(
        self,
        sample,
        timestep,
        encoder_hidden_states,
        timestep_cond=None,
        added_cond_kwargs=None,
        down_block_additional_residuals=None,
        mid_block_additional_residual=None,
        return_dict=True,
    ):
        h = self.conv_in(sample)
        t_emb = self.time_proj(timestep.unsqueeze(-1)).view(1, 8, 1, 1)
        c_emb = self.text_proj(encoder_hidden_states.mean(1)).view(1, 8, 1, 1)
        h = h + t_emb + c_emb
        return (self.conv_out(h),)
@pytest.fixture(scope="module")
 def tiny_mlpackage(tmp_path_factory):
    """Convert TinyUNet once per test session and reuse the .mlpackage."""
    import coremltools as ct
    torch.manual_seed(0)
    model = CoreMLUNetWrapper(
        TinyUNet().eval(),
        SimpleNamespace(name="SD15"),
    )
    example = (
        torch.randn(*SAMPLE_SHAPE),
        torch.randn(*TIMESTEP_SHAPE),
        torch.randn(*ENCODER_SHAPE),
    )
    traced = torch.jit.trace(model, example)
    mlmodel = ct.convert(
        traced,
        inputs=[
            ct.TensorType(name="sample", shape=SAMPLE_SHAPE, dtype=np.float16),
            ct.TensorType(name="timestep", shape=TIMESTEP_SHAPE, dtype=np.float16),
            ct.TensorType(name="encoder_hidden_states", shape=ENCODER_SHAPE, dtype=np.float16),
        ],
        outputs=[ct.TensorType(name=OUT_NAME, dtype=np.float16)],
        compute_units=ct.ComputeUnit.CPU_ONLY,
        compute_precision=ct.precision.FLOAT16,
        convert_to="mlprogram",
        minimum_deployment_target=ct.target.macOS13,
    )
    out_dir = tmp_path_factory.mktemp("tiny_unet")
    pkg_path = out_dir / "tiny.mlpackage"
    mlmodel.save(str(pkg_path))
    yield pkg_path
    shutil.rmtree(out_dir, ignore_errors=True)
 def test_coremltools_convert_round_trips_via_coreml_model(tiny_mlpackage):
    from coreml_suite.coreml_model import CoreMLModel
    model = CoreMLModel(str(tiny_mlpackage), "CPU_ONLY")
    # expected_inputs is the contract our wrappers depend on. Lock the shape
    # of the dict + a sample entry.
    expected = dict(model.expected_inputs)
    assert set(expected.keys()) == {"sample", "timestep", "encoder_hidden_states"}
    assert tuple(expected["sample"]["shape"]) == SAMPLE_SHAPE
    assert tuple(expected["timestep"]["shape"]) == TIMESTEP_SHAPE
    assert tuple(expected["encoder_hidden_states"]["shape"]) == ENCODER_SHAPE
    # Forward pass: drive the model the way CoreMLModelWrapper does.
    rng = np.random.default_rng(0)
    inputs = {
        "sample": rng.standard_normal(SAMPLE_SHAPE).astype(np.float16),
        "timestep": rng.standard_normal(TIMESTEP_SHAPE).astype(np.float16),
        "encoder_hidden_states": rng.standard_normal(ENCODER_SHAPE).astype(np.float16),
    }
    out = model(**inputs)
    assert isinstance(out, dict), f"unexpected output type: {type(out)}"
    assert OUT_NAME in out, f"missing output {OUT_NAME!r}; got {sorted(out)}"
    assert out[OUT_NAME].shape == SAMPLE_SHAPE, (
        f"output shape drift: got {out[OUT_NAME].shape}, expected {SAMPLE_SHAPE}"
    )
@@ -0,0 +1,197 @@
 """Characterization tests for the .mlpackage filename composition.
 The filename composition is the pure
 coreml_suite.core.naming.compose_out_name function. CoreMLConverter.convert
 calls it; testing the pure function avoids monkey-patching heavy converter
 internals just to capture the string.
 """
 import pytest
 from coreml_suite.core.naming import compose_out_name, lora_names_from_params
 # ---------- attention suffixes ----------------------------------------------
@pytest.mark.parametrize(
    "attn_name,suffix",
    [
        ("SPLIT_EINSUM", "se"),
        ("SPLIT_EINSUM_V2", "se2"),
        ("ORIGINAL", "orig"),
    ],
 )
 def test_attention_suffix(attn_name, suffix):
    out = compose_out_name(
        ckpt_name="dreamshaper_8.safetensors",
        batch_size=1, width=512, height=512,
        controlnet_support=False,
        attention_implementation=attn_name,
    )
    assert out == f"dreamshaper_8_1x512x512_{suffix}"
 # ---------- batch / size ----------------------------------------------------
 def test_includes_batch_and_size():
    out = compose_out_name(
        ckpt_name="dreamshaper_8.safetensors",
        batch_size=4, width=768, height=1024,
        controlnet_support=False,
        attention_implementation="SPLIT_EINSUM",
    )
    assert out == "dreamshaper_8_4x768x1024_se"
 # ---------- ControlNet ------------------------------------------------------
 def test_appends_cn_suffix_when_controlnet_support_true():
    out = compose_out_name(
        ckpt_name="dreamshaper_8.safetensors",
        batch_size=1, width=512, height=512,
        controlnet_support=True,
        attention_implementation="SPLIT_EINSUM",
    )
    assert out == "dreamshaper_8_1x512x512_cn_se"
 # ---------- ckpt name massage -----------------------------------------------
 def test_drops_extension_at_first_period():
    out = compose_out_name(
        ckpt_name="my.checkpoint.v2.safetensors",
        batch_size=1, width=512, height=512,
        controlnet_support=False,
        attention_implementation="SPLIT_EINSUM",
    )
    assert out == "my_1x512x512_se"
 def test_replaces_spaces_with_underscores():
    out = compose_out_name(
        ckpt_name="dream shaper 8.safetensors",
        batch_size=1, width=512, height=512,
        controlnet_support=False,
        attention_implementation="SPLIT_EINSUM",
    )
    assert out == "dream_shaper_8_1x512x512_se"
 # ---------- LoRA suffixes ---------------------------------------------------
 def test_single_lora():
    out = compose_out_name(
        ckpt_name="dreamshaper_8.safetensors",
        batch_size=1, width=512, height=512,
        controlnet_support=False,
        attention_implementation="SPLIT_EINSUM",
        lora_names=["epi_noiseoffset.safetensors"],
    )
    assert out == "dreamshaper_8_epi_noiseoffset_1x512x512_se"
 def test_multiple_loras_sorted():
    out = compose_out_name(
        ckpt_name="dreamshaper_8.safetensors",
        batch_size=1, width=512, height=512,
        controlnet_support=False,
        attention_implementation="SPLIT_EINSUM",
        lora_names=["zoom.safetensors", "alpha.safetensors", "moody.safetensors"],
    )
    assert out == "dreamshaper_8_alpha_moody_zoom_1x512x512_se"
 def test_lora_plus_controlnet():
    out = compose_out_name(
        ckpt_name="dreamshaper_8.safetensors",
        batch_size=1, width=512, height=512,
        controlnet_support=True,
        attention_implementation="SPLIT_EINSUM",
        lora_names=["a.safetensors"],
    )
    assert out == "dreamshaper_8_a_1x512x512_cn_se"
 # ---------- sdxl combinations -----------------------------------------------
 def test_sdxl_1024_original_gpu():
    out = compose_out_name(
        ckpt_name="sd_xl_base_1.0.safetensors",
        batch_size=1, width=1024, height=1024,
        controlnet_support=False,
        attention_implementation="ORIGINAL",
    )
    assert out == "sd_xl_base_1_1x1024x1024_orig"
 # ---------- lora_names_from_params helper ----------------------------------
 def test_lora_names_from_params_sorts_by_name():
    names = lora_names_from_params([
        ("zebra.safetensors", 1.0),
        ("apple.safetensors", 0.5),
        ("mango.safetensors", 0.7),
    ])
    assert names == ["apple.safetensors", "mango.safetensors", "zebra.safetensors"]
 def test_lora_names_from_params_empty_list():
    assert lora_names_from_params([]) == []
 # ---------- quantize_nbits suffix ------------------------------------------
 def test_quantize_nbits_none_appends_nothing():
    """'none' is the default and must keep the unquantized filename so
    existing cached .mlpackages still resolve."""
    out = compose_out_name(
        ckpt_name="dreamshaper_8.safetensors",
        batch_size=1, width=512, height=512,
        controlnet_support=False,
        attention_implementation="SPLIT_EINSUM",
        quantize_nbits="none",
    )
    assert out == "dreamshaper_8_1x512x512_se"
@pytest.mark.parametrize("nbits,suffix", [("4", "_q4"), ("6", "_q6"), ("8", "_q8")])
 def test_quantize_nbits_appends_q_suffix(nbits, suffix):
    out = compose_out_name(
        ckpt_name="dreamshaper_8.safetensors",
        batch_size=1, width=512, height=512,
        controlnet_support=False,
        attention_implementation="SPLIT_EINSUM",
        quantize_nbits=nbits,
    )
    assert out == f"dreamshaper_8_1x512x512_se{suffix}"
 def test_quantize_nbits_with_controlnet_and_lora():
    out = compose_out_name(
        ckpt_name="dreamshaper_8.safetensors",
        batch_size=1, width=512, height=512,
        controlnet_support=True,
        attention_implementation="SPLIT_EINSUM",
        lora_names=["a.safetensors"],
        quantize_nbits="6",
    )
    assert out == "dreamshaper_8_a_1x512x512_cn_se_q6"
 def test_quantize_nbits_invalid_raises():
    import pytest as _pytest
    with _pytest.raises(ValueError, match="quantize_nbits"):
        compose_out_name(
            ckpt_name="x.safetensors",
            batch_size=1, width=512, height=512,
            controlnet_support=False,
            attention_implementation="SPLIT_EINSUM",
            quantize_nbits="16",  # not in {none, 8, 6, 4}
        )
@@ -0,0 +1,183 @@
 from types import SimpleNamespace
 import torch
 from coreml_suite.conversion.attention import (
    SplitEinsumAttnProcessor,
    SplitEinsumV2AttnProcessor,
    apply_attention_implementation,
    split_einsum,
    split_einsum_v2,
 )
 from coreml_suite.conversion.shapes import conv2d_output_shape
 from coreml_suite.conversion.unet import CoreMLUNetWrapper
 class RecordingUNet(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.call = None
    def forward(
        self,
        sample,
        timestep,
        encoder_hidden_states,
        timestep_cond=None,
        added_cond_kwargs=None,
        down_block_additional_residuals=None,
        mid_block_additional_residual=None,
        return_dict=True,
        **kwargs,
    ):
        self.call = {
            "sample": sample,
            "timestep": timestep,
            "encoder_hidden_states": encoder_hidden_states,
            "timestep_cond": timestep_cond,
            "added_cond_kwargs": added_cond_kwargs,
            "down_block_additional_residuals": down_block_additional_residuals,
            "mid_block_additional_residual": mid_block_additional_residual,
            "return_dict": return_dict,
        }
        return (sample + 1,)
 def test_conv2d_output_shape_matches_torch_conv2d_contract():
    conv = torch.nn.Conv2d(
        4,
        8,
        kernel_size=(3, 5),
        stride=(2, 3),
        padding=(1, 2),
        dilation=(1, 2),
    )
    assert conv2d_output_shape(17, 19, conv) == (9, 5)
 def test_unet_wrapper_passes_context_through_for_sd15():
    unet = RecordingUNet()
    wrapper = CoreMLUNetWrapper(unet, SimpleNamespace(name="SD15"))
    sample = torch.randn(2, 4, 8, 8)
    timestep = torch.randn(2)
    context = torch.randn(2, 77, 768)
    out = wrapper(sample, timestep, context)
    assert torch.equal(out, sample + 1)
    assert unet.call["encoder_hidden_states"] is context
    assert unet.call["return_dict"] is False
 def test_unet_wrapper_routes_lcm_sdxl_and_controlnet_inputs():
    unet = RecordingUNet()
    wrapper = CoreMLUNetWrapper(unet, SimpleNamespace(name="LCM"))
    sample = torch.randn(1, 4, 8, 8)
    timestep = torch.randn(1)
    context = torch.randn(1, 77, 768)
    timestep_cond = torch.randn(1, 256)
    down_residual = torch.randn(1, 320, 8, 8)
    mid_residual = torch.randn(1, 1280, 1, 1)
    wrapper(sample, timestep, context, timestep_cond, down_residual, mid_residual)
    assert unet.call["timestep_cond"] is timestep_cond
    assert len(unet.call["down_block_additional_residuals"]) == 1
    assert unet.call["down_block_additional_residuals"][0] is down_residual
    assert unet.call["mid_block_additional_residual"] is mid_residual
 def test_unet_wrapper_routes_sdxl_added_conditioning():
    unet = RecordingUNet()
    wrapper = CoreMLUNetWrapper(unet, SimpleNamespace(name="SDXL"))
    sample = torch.randn(1, 4, 8, 8)
    timestep = torch.randn(1)
    context = torch.randn(1, 77, 2048)
    time_ids = torch.randn(1, 6)
    text_embeds = torch.randn(1, 1280)
    wrapper(sample, timestep, context, time_ids, text_embeds)
    assert unet.call["added_cond_kwargs"]["time_ids"] is time_ids
    assert unet.call["added_cond_kwargs"]["text_embeds"] is text_embeds
 def test_split_einsum_matches_original_attention_math():
    torch.manual_seed(0)
    batch = 2
    heads = 3
    dim_head = 4
    sequence = 16
    channels = heads * dim_head
    q = torch.randn(batch, channels, 1, sequence)
    k = torch.randn(batch, channels, 1, sequence)
    v = torch.randn(batch, channels, 1, sequence)
    expected = _original_attention(q, k, v, None, heads, dim_head)
    # split-einsum reorders the float32 reductions vs the reference, so equality
    # only holds up to rounding; the drift exceeds allclose's default atol on
    # some BLAS backends (e.g. Linux x86 CI).
    assert torch.allclose(split_einsum(q, k, v, None, heads, dim_head), expected, atol=1e-6)
    assert torch.allclose(split_einsum_v2(q, k, v, None, heads, dim_head), expected, atol=1e-6)
 def test_split_einsum_v2_chunked_path_matches_original_attention_math():
    torch.manual_seed(0)
    batch = 1
    heads = 2
    dim_head = 2
    sequence = 512
    channels = heads * dim_head
    q = torch.randn(batch, channels, 1, sequence)
    k = torch.randn(batch, channels, 1, sequence)
    v = torch.randn(batch, channels, 1, sequence)
    expected = _original_attention(q, k, v, None, heads, dim_head)
    assert torch.allclose(
        split_einsum_v2(q, k, v, None, heads, dim_head),
        expected,
        atol=1e-6,
    )
 def test_apply_attention_implementation_sets_split_processors():
    unet = RecordingProcessorUNet()
    assert apply_attention_implementation(unet, "ORIGINAL") is unet
    assert unet.processor is None
    apply_attention_implementation(unet, "SPLIT_EINSUM")
    assert isinstance(unet.processor, SplitEinsumAttnProcessor)
    apply_attention_implementation(unet, "SPLIT_EINSUM_V2")
    assert isinstance(unet.processor, SplitEinsumV2AttnProcessor)
 class RecordingProcessorUNet:
    def __init__(self):
        self.processor = None
    def set_attn_processor(self, processor):
        self.processor = processor
 def _original_attention(q, k, v, mask, heads, dim_head):
    batch = q.size(0)
    mh_q = q.view(batch, heads, dim_head, -1)
    mh_k = k.view(batch, heads, dim_head, -1)
    mh_v = v.view(batch, heads, dim_head, -1)
    weights = torch.einsum("bhcq,bhck->bhqk", mh_q, mh_k)
    weights = weights * (dim_head**-0.5)
    if mask is not None:
        weights = weights + mask
    weights = weights.softmax(dim=3)
    attn = torch.einsum("bhqk,bhck->bhcq", weights, mh_v)
    return attn.contiguous().view(batch, heads * dim_head, 1, -1)