sie-server 0.6.0__tar.gz → 0.6.2__tar.gz

{sie_server-0.6.0 → sie_server-0.6.2}/Dockerfile.cuda12

@@ -200,6 +200,7 @@ ENV DEBIAN_FRONTEND=noninteractive
 #   opencv-python wheel unconditionally dlopens an X11 + libGL + glib chain at
 #   import even in headless usage. Without these, every docling extract crashes
 #   with "ImportError: libxcb.so.1: cannot open shared object file" (issue #1028).
+# ninja-build: flashinfer invokes ninja during first-use SGLang kernel JIT.
 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 \
@@ -213,6 +214,7 @@ RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
     libnuma1 \
     libsm6 \
     libspatialindex-c6 \
+    ninja-build \
     libx11-6 \
     libxcb1 \
     libxext6

{sie_server-0.6.0 → sie_server-0.6.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sie-server
-Version: 0.6.0
+Version: 0.6.2
 Summary: Search Inference Engine - GPU inference server for search workloads
 License: Apache-2.0
 License-File: LICENSE

sie_server-0.6.2/models/Snowflake__snowflake-arctic-embed-l-v2.0.yaml

@@ -0,0 +1,28 @@
+sie_id: Snowflake/snowflake-arctic-embed-l-v2.0
+hf_id: Snowflake/snowflake-arctic-embed-l-v2.0
+inputs:
+  text: true
+  image: false
+  audio: false
+  video: false
+tasks:
+  encode:
+    dense:
+      dim: 1024
+    sparse: null
+    multivector: null
+  score: null
+  extract: null
+max_sequence_length: 8192
+profiles:
+  default:
+    max_batch_tokens: 16384
+    compute_precision: null
+    adapter_path: sie_server.adapters.xlm_roberta_flash:XLMRobertaFlashAdapter
+    adapter_options:
+      loadtime: {}
+      runtime:
+        pooling: cls
+        normalize: true
+        query_template: 'query: {text}'
+        doc_template: '{text}'

sie_server-0.6.2/models/mixedbread-ai__mxbai-embed-large-v1.yaml

@@ -0,0 +1,28 @@
+sie_id: mixedbread-ai/mxbai-embed-large-v1
+hf_id: mixedbread-ai/mxbai-embed-large-v1
+inputs:
+  text: true
+  image: false
+  audio: false
+  video: false
+tasks:
+  encode:
+    dense:
+      dim: 1024
+    sparse: null
+    multivector: null
+  score: null
+  extract: null
+max_sequence_length: 512
+profiles:
+  default:
+    max_batch_tokens: 16384
+    compute_precision: null
+    adapter_path: sie_server.adapters.bert_flash:BertFlashAdapter
+    adapter_options:
+      loadtime: {}
+      runtime:
+        pooling: cls
+        normalize: true
+        query_template: 'Represent this sentence for searching relevant passages: {text}'
+        doc_template: '{text}'

sie_server-0.6.2/models/nomic-ai__modernbert-embed-base.yaml

@@ -0,0 +1,28 @@
+sie_id: nomic-ai/modernbert-embed-base
+hf_id: nomic-ai/modernbert-embed-base
+inputs:
+  text: true
+  image: false
+  audio: false
+  video: false
+tasks:
+  encode:
+    dense:
+      dim: 768
+    sparse: null
+    multivector: null
+  score: null
+  extract: null
+max_sequence_length: 8192
+profiles:
+  default:
+    max_batch_tokens: 16384
+    compute_precision: bfloat16
+    adapter_path: sie_server.adapters.modernbert_flash:ModernBERTFlashAdapter
+    adapter_options:
+      loadtime: {}
+      runtime:
+        pooling: mean
+        normalize: true
+        query_template: 'search_query: {text}'
+        doc_template: 'search_document: {text}'

{sie_server-0.6.0 → sie_server-0.6.2}/openapi.json

@@ -3,7 +3,7 @@
   "info": {
     "title": "SIE Server",
     "description": "Search Inference Engine - GPU inference server for search workloads",
-    "version": "0.6.0"
+    "version": "0.6.2"
   },
   "paths": {
     "/": {

{sie_server-0.6.0 → sie_server-0.6.2}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "sie-server"
-version = "0.6.0"
+version = "0.6.2"
 description = "Search Inference Engine - GPU inference server for search workloads"
 requires-python = ">=3.12,<3.13"
 license = { text = "Apache-2.0" }

{sie_server-0.6.0 → sie_server-0.6.2}/src/sie_server/adapters/qwen3_vl_embedding/__init__.py

@@ -98,6 +98,7 @@ class Qwen3VLEmbeddingAdapter(BaseAdapter):
         max_seq_length: int | None = None,
         pooling: str = "last",
         default_instruction: str = _DEFAULT_INSTRUCTION,
+        dense_dim: int | None = None,
     ) -> None:
         if pooling not in _SUPPORTED_POOLING:
             msg = f"Unsupported pooling '{pooling}', must be one of {_SUPPORTED_POOLING}"
@@ -110,11 +111,12 @@ class Qwen3VLEmbeddingAdapter(BaseAdapter):
         self._max_seq_length = max_seq_length
         self._pooling = pooling
         self._default_instruction = default_instruction
+        self._configured_dense_dim = dense_dim
 
         self._model: Qwen3VLForConditionalGeneration | None = None
         self._processor: AutoProcessor | None = None
         self._device: str | None = None
-        self._dense_dim: int | None = 2048
+        self._dense_dim: int | None = dense_dim or 2048
 
     def load(self, device: str) -> None:
         from transformers import AutoProcessor, Qwen3VLForConditionalGeneration
@@ -162,12 +164,22 @@ class Qwen3VLEmbeddingAdapter(BaseAdapter):
         # Qwen3VLConfig stores the text model hidden size under text_config
         cfg = self._model.config
         if hasattr(cfg, "hidden_size"):
-            self._dense_dim = cfg.hidden_size
+            self._dense_dim = self._validate_or_set_dense_dim(cfg.hidden_size)
         elif hasattr(cfg, "text_config") and hasattr(cfg.text_config, "hidden_size"):
-            self._dense_dim = cfg.text_config.hidden_size
+            self._dense_dim = self._validate_or_set_dense_dim(cfg.text_config.hidden_size)
         else:
             logger.warning("Could not determine hidden_size from config, defaulting to 2048")
-            self._dense_dim = 2048
+            self._dense_dim = self._validate_or_set_dense_dim(2048)
+
+    def _validate_or_set_dense_dim(self, observed_dim: int) -> int:
+        """Validate observed model width against configured dense_dim."""
+        if self._configured_dense_dim is not None and observed_dim != self._configured_dense_dim:
+            msg = (
+                "Qwen3-VL embedding dimension mismatch: "
+                f"configured dense_dim={self._configured_dense_dim}, model hidden_size={observed_dim}"
+            )
+            raise ValueError(msg)
+        return observed_dim
 
     def _resolve_dtype(self) -> torch.dtype:
         if not self._device or not str(self._device).startswith("cuda"):

{sie_server-0.6.0 → sie_server-0.6.2}/src/sie_server/cli.py

@@ -16,6 +16,7 @@ os.environ.setdefault("HF_HUB_ETAG_TIMEOUT", "30")
 import hashlib
 import logging
 from pathlib import Path
+from typing import Annotated
 
 import typer
 from sie_sdk.bundle_utils import find_bundle_for_models, match_bundle_models
@@ -230,6 +231,15 @@ def serve(
     tracing: bool = typer.Option(default=False, help="Enable OpenTelemetry tracing (exports to localhost:4317)"),
     instrumentation: bool = typer.Option(False, "--instrumentation", "-i", help="Enable batch instrumentation logging"),
     verbose: bool = typer.Option(False, "--verbose", "-v", help="Enable verbose logging"),
+    log_level: Annotated[
+        str,
+        typer.Option(
+            "--log-level",
+            "-l",
+            envvar="SIE_LOG_LEVEL",
+            help="Log level (DEBUG, INFO, WARNING, ERROR). Helm sets SIE_LOG_LEVEL for workers.",
+        ),
+    ] = "info",
     preload: str | None = typer.Option(None, "--preload", help="Comma-separated model names to preload at startup"),
     json_logs: bool = typer.Option(False, "--json-logs", help="Enable structured JSON logging (for Loki)"),
 ) -> None:
@@ -239,7 +249,7 @@ def serve(
     from sie_server.core.logging import configure_logging
 
     # Configure logging (supports JSON format for Loki compatibility)
-    configure_logging(verbose=verbose, json_format=json_logs or None)
+    configure_logging(verbose=verbose, json_format=json_logs or None, level_name=log_level)
 
     # Handle models directory - cloud URLs pass through, local paths resolve
     if is_cloud_path(models_dir):
@@ -334,6 +344,7 @@ def serve(
         typer.echo("Start Jaeger with: mise run jaeger")
 
     typer.echo(f"Starting SIE server on {host}:{port}")
+
     typer.echo(f"Models directory: {models_dir_resolved}")
     typer.echo(f"Device: {resolved_device}")
     if cluster_cache:
@@ -370,7 +381,11 @@ def serve(
         preload_models=preload_models,
     )
 
-    run_server(host=host, port=port, reload=reload, config=config)
+    uvicorn_log = "debug" if verbose else log_level.strip().lower()
+    if uvicorn_log not in ("critical", "error", "warning", "info", "debug", "trace"):
+        uvicorn_log = "info"
+
+    run_server(host=host, port=port, reload=reload, config=config, uvicorn_log_level=uvicorn_log)
 
 
 if __name__ == "__main__":

{sie_server-0.6.0 → sie_server-0.6.2}/src/sie_server/core/logging.py

@@ -69,15 +69,29 @@ class TextFormatter(logging.Formatter):
         )
 
 
-def configure_logging(*, verbose: bool = False, json_format: bool | None = None) -> None:
+def _resolve_log_level(*, verbose: bool, level_name: str | None) -> int:
+    """Pick root log level: ``--verbose`` wins, then explicit name, then ``SIE_LOG_LEVEL`` env."""
+    if verbose:
+        return logging.DEBUG
+    raw = (level_name or os.environ.get("SIE_LOG_LEVEL") or "INFO").strip()
+    mapping = logging.getLevelNamesMapping()
+    return mapping.get(raw.upper(), logging.INFO)
+
+
+def configure_logging(
+    *,
+    verbose: bool = False,
+    json_format: bool | None = None,
+    level_name: str | None = None,
+) -> None:
     """Configure logging for SIE server.
 
     Args:
-        verbose: Enable DEBUG level logging.
+        verbose: Enable DEBUG level logging (overrides ``level_name`` / ``SIE_LOG_LEVEL``).
         json_format: Use JSON format. If None, reads from SIE_LOG_JSON env var.
+        level_name: Log level name (e.g. ``DEBUG``, ``INFO``). When None, uses ``SIE_LOG_LEVEL``.
     """
-    # Determine log level
-    log_level = logging.DEBUG if verbose else logging.INFO
+    log_level = _resolve_log_level(verbose=verbose, level_name=level_name)
 
     # Determine format (env var takes precedence if json_format not explicitly set)
     if json_format is None:

{sie_server-0.6.0 → sie_server-0.6.2}/src/sie_server/main.py

@@ -28,7 +28,14 @@ def _create_app_from_env() -> FastAPI:
     return AppFactory.create_app(config)
 
 
-def run_server(host: str, port: int, reload: bool, config: AppStateConfig) -> None:
+def run_server(
+    host: str,
+    port: int,
+    reload: bool,
+    config: AppStateConfig,
+    *,
+    uvicorn_log_level: str = "info",
+) -> None:
     config.save_to_env_vars()
     uvicorn.run(
         "sie_server.main:_create_app_from_env",
@@ -38,4 +45,5 @@ def run_server(host: str, port: int, reload: bool, config: AppStateConfig) -> No
         factory=True,
         loop="uvloop",
         timeout_keep_alive=120,
+        log_level=uvicorn_log_level,
     )

{sie_server-0.6.0 → sie_server-0.6.2}/tests/core/test_loader.py

@@ -311,6 +311,20 @@ class TestLoadAdapter:
 
         assert type(adapter).__name__ == "BGEM3FlagAdapter"
 
+    def test_load_qwen3_vl_embedding_accepts_dense_dim(self, tmp_path: Path) -> None:
+        """Qwen3 VL embedding adapter accepts dense_dim supplied by loader."""
+        config = _make_config(
+            sie_id="Qwen/Qwen3-VL-Embedding-2B",
+            hf_id="Qwen/Qwen3-VL-Embedding-2B",
+            adapter_path="sie_server.adapters.qwen3_vl_embedding:Qwen3VLEmbeddingAdapter",
+            dense_dim=2048,
+        )
+
+        adapter = load_adapter(config, tmp_path, device="cpu")
+
+        assert type(adapter).__name__ == "Qwen3VLEmbeddingAdapter"
+        assert adapter.dims.dense == 2048
+
     def test_load_custom_adapter(self, tmp_path: Path) -> None:
         """Can load a custom adapter from file."""
         # Create custom adapter file

{sie_server-0.6.0 → sie_server-0.6.2}/tests/core/test_logging.py

@@ -113,6 +113,7 @@ class TestConfigureLogging:
     def test_json_format_explicit(self, monkeypatch: pytest.MonkeyPatch) -> None:
         """Test explicit JSON format configuration."""
         monkeypatch.delenv("SIE_LOG_JSON", raising=False)
+        monkeypatch.delenv("SIE_LOG_LEVEL", raising=False)
 
         configure_logging(json_format=True)
 
@@ -123,6 +124,7 @@ class TestConfigureLogging:
     def test_text_format_explicit(self, monkeypatch: pytest.MonkeyPatch) -> None:
         """Test explicit text format configuration."""
         monkeypatch.delenv("SIE_LOG_JSON", raising=False)
+        monkeypatch.delenv("SIE_LOG_LEVEL", raising=False)
 
         configure_logging(json_format=False)
 
@@ -133,6 +135,7 @@ class TestConfigureLogging:
     def test_json_format_from_env(self, monkeypatch: pytest.MonkeyPatch) -> None:
         """Test JSON format from environment variable."""
         monkeypatch.setenv("SIE_LOG_JSON", "true")
+        monkeypatch.delenv("SIE_LOG_LEVEL", raising=False)
 
         configure_logging()
 
@@ -143,8 +146,38 @@ class TestConfigureLogging:
     def test_verbose_sets_debug_level(self, monkeypatch: pytest.MonkeyPatch) -> None:
         """Test verbose flag sets DEBUG level."""
         monkeypatch.delenv("SIE_LOG_JSON", raising=False)
+        monkeypatch.setenv("SIE_LOG_LEVEL", "INFO")
 
         configure_logging(verbose=True, json_format=False)
 
         root = logging.getLogger()
         assert root.level == logging.DEBUG
+
+    def test_sie_log_level_from_env(self, monkeypatch: pytest.MonkeyPatch) -> None:
+        """SIE_LOG_LEVEL controls root level when not verbose."""
+        monkeypatch.delenv("SIE_LOG_JSON", raising=False)
+        monkeypatch.setenv("SIE_LOG_LEVEL", "DEBUG")
+
+        configure_logging(json_format=False)
+
+        root = logging.getLogger()
+        assert root.level == logging.DEBUG
+
+    def test_level_name_param_over_env(self, monkeypatch: pytest.MonkeyPatch) -> None:
+        """Explicit level_name wins over SIE_LOG_LEVEL."""
+        monkeypatch.delenv("SIE_LOG_JSON", raising=False)
+        monkeypatch.setenv("SIE_LOG_LEVEL", "DEBUG")
+
+        configure_logging(json_format=False, level_name="WARNING")
+
+        root = logging.getLogger()
+        assert root.level == logging.WARNING
+
+    def test_invalid_log_level_falls_back_to_info(self, monkeypatch: pytest.MonkeyPatch) -> None:
+        monkeypatch.delenv("SIE_LOG_JSON", raising=False)
+        monkeypatch.setenv("SIE_LOG_LEVEL", "not-a-real-level")
+
+        configure_logging(json_format=False)
+
+        root = logging.getLogger()
+        assert root.level == logging.INFO