sie-server 0.3.1__tar.gz → 0.3.3__tar.gz

{sie_server-0.3.1 → sie_server-0.3.3}/.gitignore

@@ -16,6 +16,9 @@ eggs/
 .eggs/
 lib/
 lib64/
+# JS/TS projects under tools/ legitimately use a `lib/` directory.
+!tools/*/lib/
+!tools/*/lib/**
 parts/
 sdist/
 var/

{sie_server-0.3.1 → sie_server-0.3.3}/Dockerfile.cpu

@@ -18,14 +18,15 @@ ENV DEBIAN_FRONTEND=noninteractive \
     UV_LINK_MODE=copy \
     PIP_DISABLE_PIP_VERSION_CHECK=1
 
-# build-essential + git are builder-only; they do NOT leak into the runtime stage.
+# build-essential + git + zlib1g-dev are builder-only; they do NOT leak into the runtime stage.
 RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
     --mount=type=cache,target=/var/lib/apt/lists,sharing=locked \
     apt-get update && apt-get install -y --no-install-recommends \
     build-essential \
     ca-certificates \
     curl \
-    git
+    git \
+    zlib1g-dev
 
 ARG UV_VERSION=0.9.28
 RUN curl -LsSf https://astral.sh/uv/${UV_VERSION}/install.sh | sh \
@@ -39,7 +40,10 @@ COPY packages/sie_sdk/pyproject.toml /tmp/sie_sdk/pyproject.toml
 COPY packages/sie_server/pyproject.toml ./pyproject.toml
 
 # Stub source trees so pip accepts the editable installs during dep resolution.
-RUN mkdir -p /tmp/sie_sdk/src/sie_sdk src/sie_server \
+# Also create empty bundles/ and models/ — referenced by force-include in
+# pyproject.toml; hatchling resolves them at editable-metadata time even though
+# real contents only land in the `base` stage.
+RUN mkdir -p /tmp/sie_sdk/src/sie_sdk src/sie_server bundles models \
     && touch /tmp/sie_sdk/src/sie_sdk/__init__.py \
     && touch src/sie_server/__init__.py
 
@@ -151,13 +155,16 @@ ENV DEBIAN_FRONTEND=noninteractive \
     OMP_NUM_THREADS=4 \
     MKL_NUM_THREADS=4
 
-# Only the shared libs torch + pillow actually dlopen at runtime.
+# Only the shared libs torch + pillow + rtree actually dlopen at runtime.
+# libspatialindex-c6: rtree (docling dep) dlopens libspatialindex_c.so; the
+# rtree==1.4.1 wheel only bundles the C++ core, not the C wrapper.
 RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
     --mount=type=cache,target=/var/lib/apt/lists,sharing=locked \
     apt-get update && apt-get install -y --no-install-recommends \
     libgomp1 \
     libjpeg62-turbo \
-    libpng16-16
+    libpng16-16 \
+    libspatialindex-c6
 
 RUN groupadd -g 1000 sie && useradd -u 1000 -g sie -m sie
 

{sie_server-0.3.1 → sie_server-0.3.3}/Dockerfile.cuda11

@@ -41,7 +41,10 @@ WORKDIR /app
 COPY packages/sie_sdk/pyproject.toml /tmp/sie_sdk/pyproject.toml
 COPY packages/sie_server/pyproject.toml ./pyproject.toml
 
-RUN mkdir -p /tmp/sie_sdk/src/sie_sdk src/sie_server \
+# Stubs for editable metadata generation — bundles/ and models/ are referenced
+# by force-include in pyproject.toml and must exist; real contents are copied
+# in the base stage.
+RUN mkdir -p /tmp/sie_sdk/src/sie_sdk src/sie_server bundles models \
     && touch /tmp/sie_sdk/src/sie_sdk/__init__.py \
     && touch src/sie_server/__init__.py
 
@@ -149,6 +152,8 @@ FROM nvidia/cuda:11.8.0-base-ubuntu22.04 AS runtime
 
 ENV DEBIAN_FRONTEND=noninteractive
 
+# libspatialindex-c6: rtree (docling dep) dlopens libspatialindex_c.so; the
+# rtree==1.4.1 wheel only bundles the C++ core, not the C wrapper.
 RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
     --mount=type=cache,target=/var/lib/apt/lists,sharing=locked \
     apt-get update && apt-get install -y --no-install-recommends \
@@ -156,7 +161,8 @@ RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
     gcc \
     libc6-dev \
     libgomp1 \
-    libnuma1
+    libnuma1 \
+    libspatialindex-c6
 
 RUN groupadd -g 1000 sie && useradd -u 1000 -g sie -m sie
 

{sie_server-0.3.1 → sie_server-0.3.3}/Dockerfile.cuda12

@@ -44,7 +44,10 @@ WORKDIR /app
 COPY packages/sie_sdk/pyproject.toml /tmp/sie_sdk/pyproject.toml
 COPY packages/sie_server/pyproject.toml ./pyproject.toml
 
-RUN mkdir -p /tmp/sie_sdk/src/sie_sdk src/sie_server \
+# Stubs for editable metadata generation — bundles/ and models/ are referenced
+# by force-include in pyproject.toml and must exist; real contents are copied
+# in the base stage.
+RUN mkdir -p /tmp/sie_sdk/src/sie_sdk src/sie_server bundles models \
     && touch /tmp/sie_sdk/src/sie_sdk/__init__.py \
     && touch src/sie_server/__init__.py
 
@@ -170,6 +173,8 @@ ENV DEBIAN_FRONTEND=noninteractive
 # libnuma1: required by sgl_kernel (SGLang bundle); import fails with a
 #   misleading SM-arch error without it.
 # libgomp1: torch OpenMP runtime.
+# libspatialindex-c6: rtree (docling dep) dlopens libspatialindex_c.so; the
+#   rtree==1.4.1 wheel only bundles the C++ core, not the C wrapper.
 RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
     --mount=type=cache,target=/var/lib/apt/lists,sharing=locked \
     apt-get update && apt-get install -y --no-install-recommends \
@@ -177,7 +182,8 @@ RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
     gcc \
     libc6-dev \
     libgomp1 \
-    libnuma1
+    libnuma1 \
+    libspatialindex-c6
 
 RUN groupadd -g 1000 sie && useradd -u 1000 -g sie -m sie
 

{sie_server-0.3.1 → sie_server-0.3.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sie-server
-Version: 0.3.1
+Version: 0.3.3
 Summary: Search Inference Engine - GPU inference server for search workloads
 License: Apache-2.0
 License-File: LICENSE

{sie_server-0.3.1 → sie_server-0.3.3}/bundles/default.yaml

@@ -12,6 +12,7 @@ adapters:
 - sie_server.adapters.colbert_rotary_flash
 - sie_server.adapters.colpali
 - sie_server.adapters.colqwen2
+- sie_server.adapters.colqwen3
 - sie_server.adapters.grounding_dino
 - sie_server.adapters.gte_sparse_flash
 - sie_server.adapters.jina_flash_cross_encoder
@@ -44,7 +45,7 @@ adapters:
 - sie_server.adapters.paddleocr_vl
 deps:
   # Most flash adapters; sentence_transformer needs >=4.57
-  transformers: '>=4.57'
+  transformers: '>=4.57,<5'
   # Pin torch — flash-attn's unpinned torch dep can pull in 2.10+ which breaks torchvision
   torch: '>=2.9,<2.10'
   # bge_m3, bge_m3_flash, nemo_colembed, nomic_flash, xlm_roberta_flash
@@ -56,7 +57,7 @@ deps:
   # bge_m3_flag
   FlagEmbedding: '>=1.3'
   # grounding_dino
-  pillow: ''
+  pillow: '>=11,<12'
   # grounding_dino
   requests: ''
   # gliner, gliner_bi

{sie_server-0.3.1 → sie_server-0.3.3}/bundles/transformers5.yaml

@@ -13,5 +13,8 @@ deps:
   transformers: '>=5.0'
   # Pin torch — flash-attn's unpinned torch dep can pull in 2.10+ which breaks torchvision
   torch: '>=2.9,<2.10'
-  huggingface-hub: '>=0.20'
+  # No huggingface-hub override: transformers 5.x's own metadata
+  # pins it to >=1.3,<2.0. An override here was previously
+  # unsatisfiable against that pin and broke `uv run` resolution.
+  # Let transformers drive the version.
   pillow: ''

{sie_server-0.3.1 → sie_server-0.3.3}/models/answerdotai__answerai-colbert-small-v1.yaml

@@ -11,7 +11,7 @@ tasks:
     sparse: null
     multivector:
       dim: 96
-  score: null
+  score: {}
   extract: null
 max_sequence_length: 512
 profiles:

{sie_server-0.3.1 → sie_server-0.3.3}/models/colbert-ir__colbertv2.0.yaml

@@ -11,7 +11,7 @@ tasks:
     sparse: null
     multivector:
       dim: 128
-  score: null
+  score: {}
   extract: null
 max_sequence_length: 512
 profiles:

{sie_server-0.3.1 → sie_server-0.3.3}/models/docling.yaml

@@ -18,3 +18,11 @@ profiles:
     adapter_options:
       loadtime: {}
       runtime: {}
+  ocr:
+    max_batch_tokens: 1
+    compute_precision: null
+    adapter_path: sie_server.adapters.docling:DoclingAdapter
+    adapter_options:
+      loadtime: {}
+      runtime:
+        ocr: true

{sie_server-0.3.1 → sie_server-0.3.3}/models/jinaai__jina-colbert-v2.yaml

@@ -11,7 +11,7 @@ tasks:
     sparse: null
     multivector:
       dim: 128
-  score: null
+  score: {}
   extract: null
 max_sequence_length: 8192
 profiles:

{sie_server-0.3.1 → sie_server-0.3.3}/models/lightonai__GTE-ModernColBERT-v1.yaml

@@ -11,7 +11,7 @@ tasks:
     sparse: null
     multivector:
       dim: 128
-  score: null
+  score: {}
   extract: null
 max_sequence_length: 8192
 profiles:

{sie_server-0.3.1 → sie_server-0.3.3}/models/lightonai__Reason-ModernColBERT.yaml

@@ -11,7 +11,7 @@ tasks:
     sparse: null
     multivector:
       dim: 128
-  score: null
+  score: {}
   extract: null
 max_sequence_length: 8192
 profiles:

{sie_server-0.3.1 → sie_server-0.3.3}/models/mixedbread-ai__mxbai-colbert-large-v1.yaml

@@ -11,7 +11,7 @@ tasks:
     sparse: null
     multivector:
       dim: 128
-  score: null
+  score: {}
   extract: null
 max_sequence_length: 512
 profiles:

{sie_server-0.3.1 → sie_server-0.3.3}/models/mixedbread-ai__mxbai-edge-colbert-v0-32m.yaml

@@ -11,7 +11,7 @@ tasks:
     sparse: null
     multivector:
       dim: 64
-  score: null
+  score: {}
   extract: null
 max_sequence_length: 8192
 profiles:

{sie_server-0.3.1 → sie_server-0.3.3}/models/nvidia__llama-embed-nemotron-8b.yaml

@@ -18,10 +18,19 @@ profiles:
   default:
     max_batch_tokens: 8192
     compute_precision: bfloat16
-    adapter_path: sie_server.adapters.sglang:SGLangEmbeddingAdapter
+    # Was sie_server.adapters.sglang:SGLangEmbeddingAdapter, but
+    # SGLang's generic transformers fallback hits
+    # `assert get_embedding is False` for this Llama-based arch
+    # (no model-specific embedding implementation registered). Route
+    # through PyTorchEmbeddingAdapter like NV-Embed-v2 (also Llama-
+    # based 7B+ embedder); the heavy lane has 96 GiB so non-paged
+    # attention is fine. May be reconciled with a colleague's
+    # in-flight SGLang/arch fix later.
+    adapter_path: sie_server.adapters.pytorch_embedding:PyTorchEmbeddingAdapter
     adapter_options:
       loadtime:
-        mem_fraction_static: 0.85
+        trust_remote_code: true
+        attn_implementation: eager
       runtime:
         pooling: last_token
         normalize: true

sie_server-0.3.3/models/nvidia__nemotron-colembed-vl-4b-v2.yaml

@@ -0,0 +1,43 @@
+sie_id: nvidia/nemotron-colembed-vl-4b-v2
+hf_id: nvidia/nemotron-colembed-vl-4b-v2
+inputs:
+  text: true
+  image: true
+  audio: false
+  video: false
+tasks:
+  encode:
+    dense: null
+    sparse: null
+    multivector:
+      dim: 2560
+  score: null
+  extract: null
+max_sequence_length: 8192
+profiles:
+  default:
+    max_batch_tokens: 4096
+    compute_precision: bfloat16
+    adapter_path: sie_server.adapters.nemo_colembed:NemoColEmbedAdapter
+    adapter_options:
+      loadtime:
+        muvera_config:
+          num_repetitions: 40
+          num_simhash_projections: 6
+          projection_dim: null
+          final_projection_dim: 10240
+        token_dim: 2560
+      runtime:
+        normalize: true
+        batch_size: 4
+  muvera:
+    extends: default
+    adapter_options:
+      runtime:
+        normalize: true
+        batch_size: 4
+        output_types:
+        - dense
+        output_similarity:
+          dense: dot
+        muvera: {}

sie_server-0.3.3/models/tomoroai__tomoro-colqwen3-embed-4b.yaml

@@ -0,0 +1,42 @@
+sie_id: TomoroAI/tomoro-colqwen3-embed-4b
+hf_id: TomoroAI/tomoro-colqwen3-embed-4b
+inputs:
+  text: true
+  image: true
+  audio: false
+  video: false
+tasks:
+  encode:
+    dense: null
+    sparse: null
+    multivector:
+      dim: 320
+  score: null
+  extract: null
+max_sequence_length: 8192
+profiles:
+  default:
+    max_batch_tokens: 4096
+    compute_precision: bfloat16
+    adapter_path: sie_server.adapters.colqwen3:ColQwen3Adapter
+    adapter_options:
+      loadtime:
+        muvera_config:
+          num_repetitions: 40
+          num_simhash_projections: 6
+          projection_dim: null
+          final_projection_dim: 10240
+        token_dim: 320
+        trust_remote_code: true
+      runtime:
+        normalize: true
+  muvera:
+    extends: default
+    adapter_options:
+      runtime:
+        normalize: true
+        output_types:
+        - dense
+        output_similarity:
+          dense: dot
+        muvera: {}

{sie_server-0.3.1 → sie_server-0.3.3}/openapi.json

@@ -3,7 +3,7 @@
   "info": {
     "title": "SIE Server",
     "description": "Search Inference Engine - GPU inference server for search workloads",
-    "version": "0.3.1"
+    "version": "0.3.3"
   },
   "paths": {
     "/": {
@@ -1740,7 +1740,7 @@
               }
             ],
             "default": null,
-            "description": "Adapter-specific options",
+            "description": "Adapter-specific options. Recognized sub-keys include 'overflow_policy' (one of 'default', 'truncate_text', 'error'; default 'default') controlling how inputs exceeding the model's max_sequence_length are handled.",
             "title": "Options"
           }
         },
@@ -1763,6 +1763,23 @@
                 "location"
               ]
             }
+          },
+          {
+            "items": [
+              {
+                "text": "Apple Inc. was founded by Steve Jobs in Cupertino, California."
+              }
+            ],
+            "params": {
+              "labels": [
+                "person",
+                "organization",
+                "location"
+              ],
+              "options": {
+                "overflow_policy": "truncate_text"
+              }
+            }
           }
         ],
         "properties": {

{sie_server-0.3.1 → sie_server-0.3.3}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "sie-server"
-version = "0.3.1"
+version = "0.3.3"
 description = "Search Inference Engine - GPU inference server for search workloads"
 requires-python = ">=3.12,<3.13"
 license = { text = "Apache-2.0" }
@@ -80,6 +80,10 @@ build-backend = "hatchling.build"
 [tool.hatch.build.targets.wheel]
 packages = ["src/sie_server"]
 
+[tool.hatch.build.targets.wheel.force-include]
+"models" = "sie_server/models"
+"bundles" = "sie_server/bundles"
+
 [tool.uv.sources]
 # Prebuilt flash-attn wheel for torch 2.9 + cu128 (official wheels only go up to torch 2.8)
 # Platform-specific: Linux x86_64 only. Non-Linux users should not install the flash-attn extra.

{sie_server-0.3.1 → sie_server-0.3.3}/src/sie_server/adapters/_base_adapter.py

@@ -6,11 +6,15 @@ from typing import TYPE_CHECKING, Any, ClassVar, cast
 
 from sie_server.adapters._spec import AdapterSpec
 from sie_server.adapters._types import ERR_NOT_LOADED
+from sie_server.adapters._utils import grouped_score_pairs
 from sie_server.adapters.base import ModelAdapter, ModelCapabilities, ModelDims
 
 if TYPE_CHECKING:
     import torch
 
+    from sie_server.core.inference_output import ScoreOutput
+    from sie_server.types.inputs import Item
+
 logger = logging.getLogger(__name__)
 
 
@@ -57,7 +61,16 @@ class BaseAdapter(ModelAdapter):
             raise TypeError(msg)
 
         if "score" in spec.outputs:
-            if cls.score is ModelAdapter.score and cls.score_pairs is ModelAdapter.score_pairs:
+            # BaseAdapter ships a default score_pairs() that delegates to score().
+            # Treat that default as "not implemented" for validation purposes:
+            # subclasses must override either score() or score_pairs() so the
+            # default delegate doesn't bottom out in ModelAdapter.score().
+            score_overridden = cls.score is not ModelAdapter.score
+            score_pairs_overridden = cls.score_pairs not in (
+                ModelAdapter.score_pairs,
+                BaseAdapter.score_pairs,
+            )
+            if not score_overridden and not score_pairs_overridden:
                 msg = f"{cls.__name__} declares 'score' in outputs but does not implement score() or score_pairs()"
                 raise TypeError(msg)
 
@@ -117,6 +130,41 @@ class BaseAdapter(ModelAdapter):
             model_name=getattr(self, "_model_name_or_path", ""),
         )
 
+    # -- Default batched scoring ---------------------------------------------
+
+    def score_pairs(
+        self,
+        queries: list[Item],
+        docs: list[Item],
+        *,
+        instruction: str | None = None,
+        options: dict[str, Any] | None = None,
+    ) -> ScoreOutput:
+        """Default ``score_pairs()`` that batches via per-query grouping.
+
+        Groups parallel ``(query, doc)`` pairs by ``(text, id, instruction)``
+        so each unique query is encoded once and its docs are scored as a
+        single ``score()`` call. Subclasses with a more efficient native
+        cross-batch path (e.g. cross-encoders that pack queries and docs
+        into one transformer pass) should override this.
+
+        Per-call ``options`` are not supported by this default delegate
+        (it dispatches per-query and cannot route options into ``score()``
+        without subclass-specific knowledge). If ``options`` is a non-empty
+        mapping, this raises ``NotImplementedError`` to surface the
+        unsupported configuration; pass ``options=None`` (or ``{}``) or
+        override ``score_pairs()`` with an options-aware implementation.
+        """
+        if options:
+            msg = (
+                f"{type(self).__name__}.score_pairs(): per-call options are "
+                f"not supported by the default batching path "
+                f"(got options={options!r}). Override score_pairs() with an "
+                f"options-aware implementation."
+            )
+            raise NotImplementedError(msg)
+        return grouped_score_pairs(self.score, queries, docs, instruction=instruction)
+
     # -- Shared helpers ------------------------------------------------------
 
     def _check_loaded(self) -> None:

{sie_server-0.3.1 → sie_server-0.3.3}/src/sie_server/adapters/_utils.py

@@ -1,6 +1,10 @@
 from __future__ import annotations
 
-from typing import TYPE_CHECKING, Any
+from typing import TYPE_CHECKING, Any, Protocol
+
+import numpy as np
+
+from sie_server.core.inference_output import ScoreOutput
 
 if TYPE_CHECKING:
     import torch
@@ -8,6 +12,16 @@ if TYPE_CHECKING:
     from sie_server.types.inputs import Item
 
 
+class _ScoreFn(Protocol):
+    def __call__(
+        self,
+        query: Item,
+        items: list[Item],
+        *,
+        instruction: str | None = ...,
+    ) -> list[float]: ...
+
+
 # ---------------------------------------------------------------------------
 # RoPE utilities (eliminates 7 identical copies)
 # ---------------------------------------------------------------------------
@@ -140,3 +154,67 @@ def resolve_embedding_options(
         opts.get("query_template", default_query_template),
         opts.get("doc_template", default_doc_template),
     )
+
+
+# ---------------------------------------------------------------------------
+# Score-pair grouping (shared by ColBERT-family adapters)
+# ---------------------------------------------------------------------------
+
+
+def grouped_score_pairs(
+    score_fn: _ScoreFn,
+    queries: list[Item],
+    docs: list[Item],
+    *,
+    instruction: str | None = None,
+) -> ScoreOutput:
+    """Run a per-query ``score()`` callable over parallel (query, doc) pairs.
+
+    Groups pairs by ``(query.text, query.id, instruction)`` so each unique
+    query is encoded once and its docs are scored as one batch. Used by
+    ColBERT-family adapters to satisfy the worker's ``score_pairs()``
+    contract while reusing the optimized batched ``score()``.
+
+    Queries with ``text is None`` are not supported and raise ``ValueError``
+    (ColBERT scoring requires text). The grouping key is
+    ``(query.text, query.id or "", instruction or "")`` — two distinct
+    ``Item`` objects with identical text/id/instruction collapse to one
+    encoding pass.
+
+    Args:
+        score_fn: Bound ``adapter.score(query, items, *, instruction=None)``.
+        queries: Query items (parallel to docs).
+        docs: Document items to score.
+        instruction: Optional instruction passed through to ``score_fn``.
+
+    Returns:
+        ``ScoreOutput`` with one float per pair, in the original input order.
+
+    Raises:
+        ValueError: If ``queries`` and ``docs`` lengths differ, or any query
+            lacks text.
+    """
+    if len(queries) != len(docs):
+        msg = f"queries and docs must be parallel; got {len(queries)} vs {len(docs)}"
+        raise ValueError(msg)
+
+    if not docs:
+        return ScoreOutput(scores=np.zeros(0, dtype=np.float32), batch_size=0)
+
+    groups: dict[tuple[str, str, str], list[int]] = {}
+    for i, q in enumerate(queries):
+        if q.text is None:
+            msg = f"grouped_score_pairs requires queries[{i}].text; got None"
+            raise ValueError(msg)
+        key = (q.text, q.id or "", instruction or "")
+        groups.setdefault(key, []).append(i)
+
+    scores = np.zeros(len(docs), dtype=np.float32)
+    for indices in groups.values():
+        q = queries[indices[0]]
+        group_docs = [docs[i] for i in indices]
+        group_scores = score_fn(q, group_docs, instruction=instruction)
+        for idx, s in zip(indices, group_scores, strict=True):
+            scores[idx] = float(s)
+
+    return ScoreOutput(scores=scores, batch_size=len(docs))

{sie_server-0.3.1 → sie_server-0.3.3}/src/sie_server/adapters/base.py

@@ -127,6 +127,20 @@ class ModelAdapter(ABC):
             device: Device string (e.g., "cuda:0", "cpu").
         """
 
+    def warmup(self) -> None:
+        """Run a warmup forward pass on the loaded model.
+
+        Called by the model loader after ``load()`` has completed. The default
+        implementation is a no-op for adapters that do not need warmup. Adapters
+        that compile kernels on first call (e.g. flash-attention) or otherwise
+        benefit from a priming pass should override this and run a single
+        inference pass against a tiny synthetic input.
+
+        Splitting this from ``load()`` lets the cold-start instrumentation
+        attribute deserialize and warmup time separately.
+        """
+        return
+
     @abstractmethod
     def unload(self) -> None:
         """Unload the model and free resources.

{sie_server-0.3.1 → sie_server-0.3.3}/src/sie_server/adapters/bert_flash/__init__.py

@@ -172,6 +172,7 @@ class BertFlashAdapter(PEFTLoRAMixin, FlashBaseAdapter):
             self._max_seq_length,
         )
 
+    def warmup(self) -> None:
         # Warmup flash attention kernels
         logger.info("Warming up CUDA kernels...")
         warmup_items = [Item(text="warmup")]