slide2vec 7.1.1__tar.gz → 7.1.2__tar.gz

{slide2vec-7.1.1 → slide2vec-7.1.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: slide2vec
-Version: 7.1.1
+Version: 7.1.2
 Summary: Embedding of whole slide images with Foundation Models
 Author-email: Clément Grisi <clement.grisi@radboudumc.nl>
 License-Expression: Apache-2.0
@@ -11,94 +11,51 @@ Classifier: Programming Language :: Python :: 3 :: Only
 Classifier: Programming Language :: Python :: 3.10
 Classifier: Programming Language :: Python :: 3.11
 Classifier: Programming Language :: Python :: 3.12
-Classifier: Programming Language :: Python :: 3.13
 Requires-Python: >=3.10
 Description-Content-Type: text/markdown
 License-File: LICENSE
 Requires-Dist: hs2p[asap,cucim,openslide,sam2,tifffile,vips]>=6.1.0
 Requires-Dist: omegaconf
 Requires-Dist: matplotlib
-Requires-Dist: numpy<2
 Requires-Dist: pandas
 Requires-Dist: pillow
 Requires-Dist: rich
 Requires-Dist: torch
 Requires-Dist: torchvision
-Requires-Dist: transformers
+Requires-Dist: transformers<5,>=4.47
 Requires-Dist: wandb
 Requires-Dist: einops
-Requires-Dist: timm>=1.0.3
+Requires-Dist: einops-exts
+Requires-Dist: timm>=1.0.20
 Requires-Dist: huggingface_hub
-Provides-Extra: hoptimus
-Requires-Dist: torch>=2.0; extra == "hoptimus"
-Requires-Dist: torchvision>=0.15.0; extra == "hoptimus"
-Requires-Dist: xformers>=0.0.18; extra == "hoptimus"
-Provides-Extra: virchow
-Requires-Dist: timm>=1.0.3; extra == "virchow"
-Requires-Dist: torch>=2.0; extra == "virchow"
-Provides-Extra: uni
-Requires-Dist: torch>=2.0; extra == "uni"
-Requires-Dist: timm>=1.0.3; extra == "uni"
-Requires-Dist: xformers>=0.0.18; extra == "uni"
+Requires-Dist: packaging
 Provides-Extra: prism
-Requires-Dist: transformers~=4.53.0; extra == "prism"
-Requires-Dist: torch<2.8,>=2.3; extra == "prism"
-Requires-Dist: einops==0.8.0; extra == "prism"
-Requires-Dist: environs==11.0.0; extra == "prism"
-Requires-Dist: sacremoses==0.1.1; extra == "prism"
-Requires-Dist: xformers==0.0.31; extra == "prism"
+Requires-Dist: torch>=2.3; extra == "prism"
+Requires-Dist: environs; extra == "prism"
+Requires-Dist: sacremoses; extra == "prism"
 Provides-Extra: prism2
 Requires-Dist: torch>=2.3; extra == "prism2"
-Requires-Dist: transformers==4.51.3; extra == "prism2"
-Requires-Dist: safetensors; extra == "prism2"
-Requires-Dist: einops; extra == "prism2"
-Requires-Dist: flash-attn>=2.6.3; extra == "prism2"
-Provides-Extra: hibou
-Requires-Dist: scipy~=1.8.1; extra == "hibou"
-Requires-Dist: scikit-image~=0.19.3; extra == "hibou"
-Provides-Extra: moozy
-Requires-Dist: huggingface_hub<1.0,>=0.30.0; extra == "moozy"
+Requires-Dist: transformers<5,>=4.51; extra == "prism2"
 Provides-Extra: titan
-Requires-Dist: torch==2.0.1; extra == "titan"
-Requires-Dist: timm==1.0.3; extra == "titan"
-Requires-Dist: einops==0.6.1; extra == "titan"
-Requires-Dist: einops-exts==0.0.4; extra == "titan"
-Requires-Dist: transformers==4.46.0; extra == "titan"
+Requires-Dist: einops; extra == "titan"
+Requires-Dist: einops-exts; extra == "titan"
+Requires-Dist: timm>=1.0.3; extra == "titan"
+Requires-Dist: transformers<5,>=4.46; extra == "titan"
 Provides-Extra: waiv
-Requires-Dist: transformers<6,>=5.14; extra == "waiv"
-Requires-Dist: safetensors>=0.8; extra == "waiv"
-Requires-Dist: huggingface_hub==1.25.1; extra == "waiv"
+Requires-Dist: transformers>=4.57; extra == "waiv"
+Provides-Extra: flash-attn
+Requires-Dist: flash-attn>=2.6.3; extra == "flash-attn"
+Provides-Extra: xformers
+Requires-Dist: xformers; extra == "xformers"
 Provides-Extra: fm
-Requires-Dist: omegaconf>=2.3.0; extra == "fm"
-Requires-Dist: matplotlib; extra == "fm"
-Requires-Dist: numpy<2; extra == "fm"
-Requires-Dist: pandas; extra == "fm"
-Requires-Dist: pillow; extra == "fm"
-Requires-Dist: rich; extra == "fm"
-Requires-Dist: hs2p[asap,cucim,openslide,sam2,tifffile,vips]>=6.1.0; extra == "fm"
-Requires-Dist: wandb; extra == "fm"
-Requires-Dist: torch<2.8,>=2.3; extra == "fm"
-Requires-Dist: torchvision>=0.18.0; extra == "fm"
-Requires-Dist: einops>=0.8.0; extra == "fm"
-Requires-Dist: timm>=1.0.20; extra == "fm"
-Requires-Dist: huggingface_hub<1.0,>=0.30.0; extra == "fm"
-Requires-Dist: environs; extra == "fm"
-Requires-Dist: einops-exts>=0.0.4; extra == "fm"
-Requires-Dist: transformers>=4.53; extra == "fm"
-Requires-Dist: sacremoses; extra == "fm"
-Requires-Dist: xformers>=0.0.31; extra == "fm"
-Requires-Dist: scipy>=1.8.1; extra == "fm"
-Requires-Dist: scikit-image>=0.19.3; extra == "fm"
-Requires-Dist: torchmetrics>=0.10.3; extra == "fm"
-Requires-Dist: fvcore; extra == "fm"
-Requires-Dist: iopath; extra == "fm"
-Requires-Dist: webdataset; extra == "fm"
-Requires-Dist: scikit-survival; extra == "fm"
-Requires-Dist: scikit-learn; extra == "fm"
+Requires-Dist: slide2vec[prism,prism2,titan,waiv]; extra == "fm"
 Requires-Dist: fairscale; extra == "fm"
-Requires-Dist: packaging==23.2; extra == "fm"
-Requires-Dist: ninja==1.11.1.1; extra == "fm"
-Requires-Dist: psutil<6; extra == "fm"
+Provides-Extra: cpu
+Requires-Dist: torch; extra == "cpu"
+Requires-Dist: torchvision; extra == "cpu"
+Provides-Extra: cu128
+Requires-Dist: torch; extra == "cu128"
+Requires-Dist: torchvision; extra == "cu128"
 Provides-Extra: docs
 Requires-Dist: sphinx>=8.1; extra == "docs"
 Requires-Dist: furo; extra == "docs"
@@ -131,6 +88,10 @@ pip install slide2vec
 
 Many models need additional dependencies available through `pip install "slide2vec[fm]"`. See the [model installation guide](https://clemsgrs.github.io/slide2vec/models.html#model-installation) for model-specific extras, separate environments, and upstream packages.
 
+### Reproducible model environments
+
+The repository ships a `uv.lock` file that pins every package for each extra. From a checkout, `uv sync --extra cu128 --extra fm` installs the exact locked stack (use `--extra cpu` for CPU-only PyTorch). Add `--extra flash-attn` for PRISM2 and `gigapath-slide`. The Docker image is built from the same lock; its `SLIDE2VEC_EXTRAS` build argument selects the extras, for example `--build-arg SLIDE2VEC_EXTRAS="cu128"` for the base install only.
+
 For gated models such as Virchow2, request access on the model's Hugging Face page and authenticate with `hf auth login` or an `HF_TOKEN` environment variable.
 
 ## Embed a slide

{slide2vec-7.1.1 → slide2vec-7.1.2}/README.md

@@ -15,6 +15,10 @@ pip install slide2vec
 
 Many models need additional dependencies available through `pip install "slide2vec[fm]"`. See the [model installation guide](https://clemsgrs.github.io/slide2vec/models.html#model-installation) for model-specific extras, separate environments, and upstream packages.
 
+### Reproducible model environments
+
+The repository ships a `uv.lock` file that pins every package for each extra. From a checkout, `uv sync --extra cu128 --extra fm` installs the exact locked stack (use `--extra cpu` for CPU-only PyTorch). Add `--extra flash-attn` for PRISM2 and `gigapath-slide`. The Docker image is built from the same lock; its `SLIDE2VEC_EXTRAS` build argument selects the extras, for example `--build-arg SLIDE2VEC_EXTRAS="cu128"` for the base install only.
+
 For gated models such as Virchow2, request access on the model's Hugging Face page and authenticate with `hf auth login` or an `HF_TOKEN` environment variable.
 
 ## Embed a slide

slide2vec-7.1.2/pyproject.toml

@@ -0,0 +1,264 @@
+[build-system]
+requires = ["setuptools>=61.0", "wheel"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "slide2vec"
+version = "7.1.2"
+description = "Embedding of whole slide images with Foundation Models"
+readme = "README.md"
+requires-python = ">=3.10"
+license = "Apache-2.0"
+authors = [
+    { name = "Clément Grisi", email = "clement.grisi@radboudumc.nl" },
+]
+classifiers = [
+    "Programming Language :: Python :: 3",
+    "Programming Language :: Python :: 3 :: Only",
+    "Programming Language :: Python :: 3.10",
+    "Programming Language :: Python :: 3.11",
+    "Programming Language :: Python :: 3.12",
+]
+dependencies = [
+    "hs2p[asap,cucim,openslide,sam2,tifffile,vips]>=6.1.0",
+    "omegaconf",
+    "matplotlib",
+    "pandas",
+    "pillow",
+    "rich",
+    "torch",
+    "torchvision",
+    # <5: Hibou's remote code imports transformers.onnx and
+    # pytorch_utils.find_pruneable_heads_and_indices, both removed in v5.
+    # >=4.47: PRISM's configuring_prism.py needs sub_configs / base_config_key.
+    "transformers>=4.47,<5",
+    "wandb",
+    "einops",
+    # TITAN remote code (used by the built-in conchv15 preset) imports it.
+    "einops-exts",
+    # first release that registers vit_base_patch16_dinov3
+    "timm>=1.0.20",
+    "huggingface_hub",
+    "packaging",
+]
+
+[project.urls]
+Homepage = "https://github.com/clemsgrs/slide2vec"
+"Bug Tracker" = "https://github.com/clemsgrs/slide2vec/issues"
+
+[project.scripts]
+slide2vec = "slide2vec.cli:entrypoint"
+
+# Rule for extras: lower bounds only. An exact or upper pin is kept only when it
+# is structural (the model's code breaks outside it); the reason sits next to
+# the pin. A model whose stack conflicts with the shared stack gets its own
+# extra and a uv conflict below. numpy<2 comes from hs2p[asap].
+[project.optional-dependencies]
+prism = [
+    # perceiver.py calls torch.nn.attention.sdpa_kernel with a list (absent in 2.2)
+    "torch>=2.3",
+    "environs",
+    "sacremoses",
+]
+prism2 = [
+    "torch>=2.3",
+    # the text (diagnostic) variant calls a Phi3Model helper whose signature
+    # changed in 4.52 and which 4.53 removed; slide2vec.encoders.models.prism2
+    # vendors it, so no upper bound. PRISM2 also imports flash_attn: add the
+    # flash-attn extra (CUDA only, so not pulled in here)
+    "transformers>=4.51,<5",
+]
+titan = [
+    "einops",
+    "einops-exts",
+    "timm>=1.0.3",
+    "transformers>=4.46,<5",
+]
+waiv = [
+    # phaet and mascaret load with zero missing keys at the pinned revisions
+    "transformers>=4.57",
+]
+# gigapath-slide needs flash-attn on GPUs with compute capability above 7
+flash-attn = [
+    "flash-attn>=2.6.3",
+]
+xformers = [
+    # optional: PRISM's PERCEIVER_MEM_EFF_ATTN path; gigapath-slide needs it
+    # on GPUs with compute capability 7 or below. Unpinned: each xformers
+    # build pins its torch.
+    "xformers",
+]
+fm = [
+    "slide2vec[prism,prism2,titan,waiv]",
+    # hard import in MUSK and gigapath-slide
+    "fairscale",
+]
+# torch wheel flavour for uv; when neither is selected torch comes from PyPI
+cpu = [
+    "torch",
+    "torchvision",
+]
+cu128 = [
+    "torch",
+    "torchvision",
+]
+docs = [
+    "sphinx>=8.1",
+    "furo",
+    "myst-parser",
+    "sphinx-copybutton",
+    "sphinx-autodoc-typehints",
+]
+testing = [
+    "pytest>=6.0",
+    "pytest-cov>=2.0",
+    "mypy>=0.910",
+    "flake8>=3.9",
+    "flake8-pyproject>=1.2.3",
+    "tox>=3.24",
+]
+
+[dependency-groups]
+dev = [
+    "pytest>=6.0",
+    "pytest-cov>=2.0",
+    "mypy>=0.910",
+    "flake8>=3.9",
+    "flake8-pyproject>=1.2.3",
+    "tox>=3.24",
+    # tests/test_encoder_plugins.py installs a plugin with `python -m pip`;
+    # uv environments ship without pip
+    "pip",
+]
+docs = [
+    "sphinx>=8.1",
+    "furo",
+    "myst-parser",
+    "sphinx-copybutton",
+    "sphinx-autodoc-typehints",
+]
+# encoder packages that are not on PyPI (musk, conch, gigapath-slide, sam2)
+models-git = [
+    "musk",
+    "conch",
+    "gigapath",
+    "sam-2",
+]
+
+[tool.uv]
+# ASAP and cucim are Linux-only. numpy<2 (from hs2p[asap]) has no wheels for
+# Python 3.13+, so the lock covers Python 3.10 to 3.12.
+environments = ["sys_platform == 'linux' and python_full_version < '3.13'"]
+# Lock-only: the cu128 index stops at torch 2.11. Hold the CPU and PyPI forks at
+# the same release so the CPU CI image tests the torch the CUDA image ships.
+# This is not part of the published metadata.
+constraint-dependencies = ["torch<2.12"]
+conflicts = [
+    [
+        { extra = "cpu" },
+        { extra = "cu128" },
+    ],
+]
+
+# flash-attn ships only an sdist whose build imports torch; declare its metadata
+# so uv can lock it without building it
+[[tool.uv.dependency-metadata]]
+name = "flash-attn"
+requires-dist = ["torch", "einops"]
+
+# flash-attn's setup.py imports torch, packaging, wheel and psutil. Build it in
+# an isolated environment that holds the same torch as the runtime environment.
+[tool.uv.extra-build-dependencies]
+flash-attn = [
+    { requirement = "torch", match-runtime = true },
+    "packaging",
+    "psutil",
+    "wheel",
+    "ninja",
+]
+
+[tool.uv.sources]
+torch = [
+    { index = "pytorch-cpu", extra = "cpu" },
+    { index = "pytorch-cu128", extra = "cu128" },
+]
+torchvision = [
+    { index = "pytorch-cpu", extra = "cpu" },
+    { index = "pytorch-cu128", extra = "cu128" },
+]
+musk = { git = "https://github.com/lilab-stanford/MUSK.git" }
+conch = { git = "https://github.com/Mahmoodlab/CONCH.git" }
+gigapath = { git = "https://github.com/prov-gigapath/prov-gigapath.git" }
+sam-2 = { git = "https://github.com/facebookresearch/sam2.git" }
+
+[[tool.uv.index]]
+name = "pytorch-cpu"
+url = "https://download.pytorch.org/whl/cpu"
+explicit = true
+
+[[tool.uv.index]]
+name = "pytorch-cu128"
+url = "https://download.pytorch.org/whl/cu128"
+explicit = true
+
+[tool.setuptools]
+include-package-data = true
+zip-safe = false
+
+[tool.setuptools.packages.find]
+include = ["slide2vec*"]
+
+[tool.setuptools.package-data]
+slide2vec = ["py.typed", "cufile.json"]
+"slide2vec.configs" = ["*.yaml"]
+
+[tool.pytest.ini_options]
+addopts = "--cov=slide2vec -m 'not gpu_integration'"
+testpaths = [
+    "tests",
+]
+markers = [
+    "heavy: real-weight foundation-model inference on CPU; minutes per test. Included in the PR suite and available separately through the manual heavy workflow (.github/workflows/nightly-heavy.yaml).",
+    "gpu_integration: real one-GPU versus multi-GPU parity; requires at least two visible CUDA devices. Run explicitly with `CUDA_VISIBLE_DEVICES=0,1 python -m pytest -m gpu_integration --no-cov`.",
+]
+
+[tool.mypy]
+mypy_path = "."
+check_untyped_defs = true
+disallow_any_generics = true
+ignore_missing_imports = true
+no_implicit_optional = true
+show_error_codes = true
+strict_equality = true
+warn_redundant_casts = true
+warn_return_any = true
+warn_unreachable = true
+warn_unused_configs = true
+no_implicit_reexport = true
+
+[tool.flake8]
+max-line-length = 160
+
+[tool.bumpver]
+current_version = "7.1.2"
+version_pattern = "MAJOR.MINOR.PATCH"
+commit = false
+tag = false
+push = false
+files = [
+    "pyproject.toml",
+    "slide2vec/__init__.py",
+    "docs/conf.py",
+]
+
+[tool.bumpver.file_patterns]
+"pyproject.toml" = [
+    '^current_version = "{version}"$',
+    '^version = "{version}"$',
+]
+"slide2vec/__init__.py" = [
+    '^__version__ = "{version}"$',
+]
+"docs/conf.py" = [
+    '^release = "{version}"$',
+]

{slide2vec-7.1.1 → slide2vec-7.1.2}/slide2vec/__init__.py

@@ -36,7 +36,7 @@ from slide2vec.encoders import (
 )
 
 
-__version__ = "7.1.1"
+__version__ = "7.1.2"
 
 __all__ = [
     "Model",

{slide2vec-7.1.1 → slide2vec-7.1.2}/slide2vec/encoders/models/dinov3.py

@@ -36,9 +36,7 @@ def require_timm_for_dinov3() -> None:
     if Version(timm.__version__) < Version(_MIN_TIMM_VERSION):
         raise ImportError(
             f"dinov3-vitb16 requires timm>={_MIN_TIMM_VERSION} (the DINOv3 "
-            f"architectures were added in that release); found timm=={timm.__version__}. "
-            "Note that the 'titan' extra pins timm==1.0.3 and cannot be installed "
-            "alongside dinov3-vitb16."
+            f"architectures were added in that release); found timm=={timm.__version__}."
         )
 
 

{slide2vec-7.1.1 → slide2vec-7.1.2}/slide2vec/encoders/models/moozy/loading.py

@@ -79,7 +79,7 @@ def load_moozy_inference_components(device: torch.device) -> MoozyInferenceCompo
     except Exception as exc:
         raise RuntimeError(
             "huggingface_hub is required for bundled MOOZY inference support. "
-            "Install it with: pip install 'slide2vec[moozy]'"
+            "Install it with: pip install huggingface_hub"
         ) from exc
 
     checkpoint_path = hf_hub_download(repo_id=_HF_REPO_ID, filename=_HF_CHECKPOINT)

slide2vec-7.1.2/slide2vec/encoders/models/prism2.py

@@ -0,0 +1,172 @@
+"""PRISM2 slide encoder implementation."""
+
+import torch
+from transformers import AutoModel, AutoProcessor
+from transformers.models.phi3.modeling_phi3 import Phi3Model
+
+try:
+    from transformers.cache_utils import SlidingWindowCache
+except ImportError:  # pragma: no cover - older transformers
+    SlidingWindowCache = ()  # type: ignore[assignment,misc]
+
+from slide2vec.encoders.base import (
+    SlideEncoder,
+    preferred_default_device,
+    resolve_requested_output_variant,
+)
+from slide2vec.encoders.registry import register_encoder
+
+PRISM2_MODEL_ID = "paige-ai/Prism2"
+PRISM2_REVISION = "450352d0ddc6b42b21ce20794ce0fbefe6b5a47a"
+_PHI3_MASK_HELPER = "_prepare_4d_causal_attention_mask_with_cache_position"
+
+
+def _prepare_4d_causal_attention_mask_with_cache_position(
+    attention_mask: torch.Tensor | None,
+    sequence_length: int,
+    target_length: int,
+    dtype: torch.dtype,
+    cache_position: torch.Tensor,
+    batch_size: int,
+    device: torch.device | None = None,
+    config=None,
+    past_key_values=None,
+    **_: object,
+) -> torch.Tensor | None:
+    """``Phi3Model._prepare_4d_causal_attention_mask_with_cache_position`` as
+    shipped by transformers 4.52.4, accepting the ``device`` keyword of 4.51.
+
+    The pinned PRISM2 remote code (written against 4.51.3) calls it with
+    ``device=`` from the text decoder, which only the ``diagnostic`` output
+    variant runs; the ``base`` variant never reaches it. 4.52 dropped the
+    ``device`` parameter and 4.53 removed the helper, so this one is installed
+    on ``Phi3Model`` for every transformers version. Returns a
+    ``(batch_size, 1, sequence_length, target_length)`` mask in ``dtype``: 0
+    where a query may attend, the dtype minimum elsewhere. A 4D
+    ``attention_mask`` is returned as is.
+    """
+    if attention_mask is not None and attention_mask.dim() == 4:
+        return attention_mask
+    if device is None:
+        device = cache_position.device
+    min_dtype = torch.finfo(dtype).min
+    causal_mask = torch.full(
+        (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
+    )
+    positions = torch.arange(target_length, device=device)
+    diagonal_attend_mask = positions > cache_position.reshape(-1, 1)
+    text_config = config.get_text_config() if config is not None else None
+    sliding_window = getattr(text_config, "sliding_window", None)
+    if getattr(text_config, "use_sliding_window", True) and sliding_window is not None:
+        if not isinstance(past_key_values, SlidingWindowCache) or sequence_length > target_length:
+            sliding_attend_mask = positions <= (cache_position.reshape(-1, 1) - sliding_window)
+            diagonal_attend_mask.bitwise_or_(sliding_attend_mask)
+    causal_mask *= diagonal_attend_mask
+    causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+    if attention_mask is not None:
+        causal_mask = causal_mask.clone()
+        if attention_mask.shape[-1] > target_length:
+            attention_mask = attention_mask[:, :target_length]
+        mask_length = attention_mask.shape[-1]
+        padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(device)
+        padding_mask = padding_mask == 0
+        causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+            padding_mask, min_dtype
+        )
+    return causal_mask
+
+
+def install_phi3_mask_helper() -> None:
+    """Give ``Phi3Model`` the mask helper the PRISM2 remote code calls.
+
+    Installed unconditionally: 4.51 ships a compatible helper, 4.52 ships one
+    that rejects the ``device`` keyword PRISM2 passes, and 4.53 ships none.
+    ``Phi3Model`` itself calls the helper with keywords only, so the vendored
+    one serves every version.
+    """
+    setattr(
+        Phi3Model,
+        _PHI3_MASK_HELPER,
+        staticmethod(_prepare_4d_causal_attention_mask_with_cache_position),
+    )
+
+
+@register_encoder(
+    "prism2",
+    level="slide",
+    tile_encoder="virchow2",
+    tile_encoder_output_variant="cls",
+    output_variants={
+        "base": {"encode_dim": 2560},
+        "diagnostic": {"encode_dim": 3072},
+    },
+    default_output_variant="base",
+    supported_spacing_um=0.5,
+    precision="bf16",
+    source="paige-ai/Prism2",
+)
+class Prism2SlideEncoder(SlideEncoder):
+    def __init__(self, *, output_variant: str | None = None):
+        self._output_variant = resolve_requested_output_variant(
+            output_variant,
+            default="base",
+            allowed=("base", "diagnostic"),
+        )
+        install_phi3_mask_helper()
+        shared_load_kwargs = {
+            "revision": PRISM2_REVISION,
+            "trust_remote_code": True,
+        }
+        self._model = AutoModel.from_pretrained(
+            PRISM2_MODEL_ID,
+            **shared_load_kwargs,
+            torch_dtype="auto",
+        ).eval()
+        self._processor = AutoProcessor.from_pretrained(
+            PRISM2_MODEL_ID,
+            **shared_load_kwargs,
+        )
+        self._device = preferred_default_device()
+
+    @property
+    def encode_dim(self) -> int:
+        return 2560 if self._output_variant == "base" else 3072
+
+    @property
+    def device(self) -> torch.device:
+        return self._device
+
+    def to(self, device: torch.device | str) -> "Prism2SlideEncoder":
+        self._device = torch.device(device)
+        if self._output_variant == "base":
+            # The pinned upstream get_base_embedding path uses image_resampler only.
+            # Keep the excluded Phi-3 decoder on CPU instead of spending GPU memory
+            # on text-generation parameters this base-only preset can never call.
+            self._model.image_resampler.to(self._device)
+        else:
+            diagnostic_modules = (
+                self._model.image_resampler,
+                self._model.img_projection,
+                self._model.text_decoder,
+            )
+            for module in diagnostic_modules:
+                if self._device.type == "cuda":
+                    module.to(self._device, dtype=torch.bfloat16)
+                else:
+                    module.to(self._device)
+        return self
+
+    def encode_slide(
+        self,
+        tile_features: torch.Tensor,
+        coordinates: torch.Tensor | None = None,
+        *,
+        tile_size_lv0: int | None = None,
+    ) -> torch.Tensor:
+        batch = self._processor(tile_embeddings=[tile_features]).to(self._device)
+        embedding_method = (
+            self._model.get_base_embedding
+            if self._output_variant == "base"
+            else self._model.get_diagnostic_embedding
+        )
+        return embedding_method(**batch).squeeze(0)