feat(vendor): add LLM2Vec embedding model

pyproject.toml

@@ -36,9 +36,6 @@ possibly-missing-attribute = "ignore"
 missing-argument = "ignore"
 unused-type-ignore-comment = "ignore"
 
-[tool.bandit]
-exclude_dirs = ["tests", "docs"]
-
 
 [tool.coverage.run]
 source = ["src/pruna"]
@@ -70,29 +67,21 @@ name = "pruna_internal"
 url = "https://prunaai.pythonanywhere.com/simple/"
 explicit = true
 
-[[tool.uv.index]]
-name = "intel-pytorch-extension"
-url = "https://pytorch-extension.intel.com/release-whl/stable/cpu/cn/"
-explicit = true
-
 [tool.uv]
 index-strategy = "first-index"
+exclude-newer = "1 week"  # protection against compromised dependencies
+# trusted dev wheels that are missing an upload date
+exclude-newer-package = { gptqmodel = false, "stable-fast-pruna" = false }
 
 conflicts = [
     [{ extra = "awq" }, { extra = "vbench" }],
     [{ extra = "vllm" }, { extra = "vbench" }],
-    [{ extra = "intel" }, { extra = "awq" }],
     [{ extra = "gptq" }, { extra = "awq" }],
-    # intel is incompatible with all stable-fast variants and vllm
-    [{ extra = "intel" }, { extra = "stable-fast" }, { extra = "stable-fast-extraindex" }],
-    [{ extra = "intel" }, { extra = "full" }, { extra = "stable-fast-extraindex" }],
-    [{ extra = "intel" }, { extra = "vllm" }],
     [{ extra = "kvpress" }, { extra = "vbench" }],
 ]
 
 [tool.uv.sources]
 gptqmodel = { index = "pruna_internal", marker = "sys_platform != 'darwin' or platform_machine != 'arm64'" }
-intel-extension-for-pytorch = { index = "intel-pytorch-extension" }
 stable-fast-pruna = { index = "pruna_internal", extra = "stable-fast-extraindex" }
 
 [project]
@@ -171,6 +160,21 @@ vllm = [
     "vllm>=0.16.0",
     "ray",
 ]
+rapidata = [
+    "rapidata>=3.0.0",
+]
+upscale = [
+    "realesrgan",
+]
+evaluation = [
+    "pruna[rapidata]",
+    "pruna[lmharness]",
+    "outlines>1.2.0,<2.0.0",
+    "litellm>=1.0.0",
+]
+oneig-reasoning = [
+    "hf_transfer>=0.1.9",
+]
 stable-fast = [
     "xformers>=0.0.30",
     "stable-fast-pruna>=1.0.8,<1.0.9",
@@ -195,18 +199,12 @@ awq = [
     "llmcompressor>=0.9",
     "torch>=2.9.0"
 ]
-upscale = [
-    "realesrgan",
-]
 full = [
     "pruna[stable-fast]",
 ]
 vbench = [
     "vbench-pruna; sys_platform != 'darwin'",
 ]
-rapidata = [
-    "rapidata>=3.0.0"
-]
 dev = [
     "wget",
     "python-dotenv",
@@ -233,22 +231,13 @@ dev = [
     "types-PyYAML",
     "logbar",
     "pytest-xdist>=3.8.0",
+    "pruna[evaluation]",
 ]
 cpu = []
 lmharness = [
     "lm-eval>=0.4.0"
 ]
-evaluation = [
-    "pruna[rapidata]",
-    "pruna[lmharness]"
-]
 
-# Intel extension is tightly coupled with the torch version
-intel = [
-    "intel-extension-for-pytorch>=2.7.0",
-    "torch>=2.7.0,<2.9.0",
-    "torchvision>=0.22.0,<0.24.0",
-]
 kvpress = [
     "kvpress>=0.5.2",
 ]

src/pruna/evaluation/metrics/metric_torch.py

@@ -50,6 +50,26 @@
 )
 from pruna.logging.logger import pruna_logger
 
+_PRUNA_TASK_ROUTING_KWARGS: tuple[str, ...] = (
+    "vlm_type",
+    "model_name",
+    "structured_output",
+    "vlm_kwargs",
+    "api_key",
+)
+
+
+def _strip_task_routing_kwargs(kwargs: dict[str, Any]) -> None:
+    """
+    Drop kwargs :class:`~pruna.evaluation.task.Task` passes when building mixed metric lists.
+
+    Torchmetrics classes often end with ``**kwargs`` and would otherwise accept bogus keys
+    until a lower layer raises. Stripping here keeps :class:`TorchMetricWrapper` the single
+    choke point between Pruna routing and torchmetrics constructors.
+    """
+    for key in _PRUNA_TASK_ROUTING_KWARGS:
+        kwargs.pop(key, None)
+
 
 def default_update(metric: Metric, *args, **kwargs) -> None:
     """
@@ -124,9 +144,7 @@ def arniqa_update(metric: ARNIQA, preds: Any) -> None:
 
 
 def ssim_update(
-        metric: StructuralSimilarityIndexMeasure | MultiScaleStructuralSimilarityIndexMeasure,
-        preds: Any,
-        target: Any
+    metric: StructuralSimilarityIndexMeasure | MultiScaleStructuralSimilarityIndexMeasure, preds: Any, target: Any
 ) -> None:
     """
     Update handler for SSIM or MS-SSIM metric.
@@ -152,29 +170,22 @@ class TorchMetrics(Enum):
     """
     Enumeration of torchmetrics metrics for evaluation.
 
-    This enum provides a tuple per member (metric_factory, update_fn, call_type):
-    metric_factory builds the metric (typically a torchmetrics class, or
-    functools.partial when some constructor arguments are fixed); update_fn is
-    an optional custom update handler; call_type describes how inputs are paired
-    for the metric.
+    Each member value is a ``(metric_factory, update_fn, call_type)`` tuple.
 
     Parameters
     ----------
     value : tuple
-        Tuple holding metric_factory, update_fn, and call_type as described above.
+        ``(metric_factory, update_fn, call_type)`` for this enum member.
     names : str
-        The name of the enum member.
+        Enum member name.
     module : str
-        The module where the enum is defined.
+        Defining module name.
     qualname : str
-        The qualified name of the enum.
+        Qualified name of the enum class.
     type : type
-        The type of the enum.
+        Enum metaclass type.
     start : int
-        The start index for auto-numbering enum values.
-    boundary : enum.FlagBoundary or None
-        Boundary handling mode used by the Enum functional API for Flag and
-        IntFlag enums.
+        Auto-numbering start index for functional API enums.
     """
 
     fid = (FrechetInceptionDistance, fid_update, "gt_y")
@@ -246,6 +257,7 @@ def __new__(cls, metric_name: str, call_type: str = "", **kwargs) -> StatefulMet
         if metric_name == "clip_score" and call_type.startswith(PAIRWISE):
             from pruna.evaluation.metrics.metric_pairwise_clip import PairwiseClipScore
 
+            _strip_task_routing_kwargs(kwargs)
             return PairwiseClipScore(**kwargs)
         return super().__new__(cls)
 
@@ -259,6 +271,7 @@ def __init__(self, metric_name: str, call_type: str = "", **kwargs) -> None:
             If the metric name is not supported.
         """
         self.metric_name = metric_name
+        _strip_task_routing_kwargs(kwargs)
         super().__init__(kwargs.pop("device", None))
         try:
             self.metric = TorchMetrics[metric_name](**kwargs)

src/pruna/evaluation/metrics/vendor/NOTICE.oneig_llm2vec

@@ -0,0 +1,17 @@
+LLM2Vec (llm2vec package) vendored from OneIG-Benchmark.
+
+Source: https://github.com/OneIG-Bench/OneIG-Benchmark
+Commit: 41b49831e79e6dde5323618c164da1c4cf0f699d
+Path: scripts/utils/llm2clip/llm2vec/
+
+OneIG-Benchmark is licensed under the Apache License 2.0.
+See the project repository for full license text.
+
+``oneig_llm2vec/modeling_llama_encoder.py`` is derived from
+McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp (Hugging Face Hub);
+Pruna relaxes the upstream flash-attention-only constraint for CPU use.
+
+Pruna also includes two minimal compatibility fixes in
+``oneig_llm2vec/llm2vec.py``:
+- Preserve constructor-provided ``doc_max_length`` instead of hardcoding 512.
+- Honor ``convert_to_numpy=True`` in ``encode()`` by returning ``numpy.ndarray``.