projectdavid-orm 1.6.3__tar.gz → 1.8.0__tar.gz

{projectdavid_orm-1.6.3 → projectdavid_orm-1.8.0}/CHANGELOG.md

@@ -1,3 +1,17 @@
+# [1.8.0](https://github.com/project-david-ai/projectdavid-orm/compare/v1.7.0...v1.8.0) (2026-04-12)
+
+
+### Features
+
+* **inference_deployment:** add mm_processor_kwargs column ([ae55b96](https://github.com/project-david-ai/projectdavid-orm/commit/ae55b96acf0696b27b27afa24d9ab515bb35481c))
+
+# [1.7.0](https://github.com/project-david-ai/projectdavid-orm/compare/v1.6.3...v1.7.0) (2026-04-12)
+
+
+### Features
+
+* **inference_deployment:** add vLLM hyperparam columns ([6c6de67](https://github.com/project-david-ai/projectdavid-orm/commit/6c6de67eca69c7e71c9c3755765254fe1aeac25d))
+
 ## [1.6.3](https://github.com/project-david-ai/projectdavid-orm/compare/v1.6.2...v1.6.3) (2026-04-11)
 
 

{projectdavid_orm-1.6.3/src/projectdavid_orm.egg-info → projectdavid_orm-1.8.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: projectdavid-orm
-Version: 1.6.3
+Version: 1.8.0
 Summary: Shared SQLAlchemy ORM models for ProjectDavid services
 Author-email: "Francis N." <engineering@projectdavid.co.uk>
 License: PolyForm Noncommercial License 1.0.0

{projectdavid_orm-1.6.3 → projectdavid_orm-1.8.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "projectdavid-orm"
-version = "1.6.3"
+version = "1.8.0"
 description = "Shared SQLAlchemy ORM models for ProjectDavid services"
 readme = "README.md"
 authors = [

{projectdavid_orm-1.6.3 → projectdavid_orm-1.8.0}/src/projectdavid_orm/projectdavid_orm/models.py

@@ -886,22 +886,110 @@ class BaseModel(Base):
 
 
 class InferenceDeployment(Base):
+    """
+    Tracks a live vLLM deployment on Ray Serve.
+
+    Hyperparams stored here override environment-level defaults at deploy time.
+    The reconciler reads these fields and passes them directly to vLLM engine args.
+    None values fall back to the VLLM_DEFAULT_* env vars in inference_worker.py.
+
+    Vision kwargs resolution priority:
+      1. mm_processor_kwargs DB column  — set via activation API
+      2. _VISION_FAMILY_CONFIGS registry — inference_worker.py family defaults
+      3. vLLM built-in defaults         — no kwargs injected for unknown families
+    """
+
     __tablename__ = "inference_deployments"
-    id = Column(String(64), primary_key=True, index=True)
 
+    # --- Identity ---
+    id = Column(String(64), primary_key=True, index=True)
     node_id = Column(String(128), nullable=True)
-
     internal_hostname = Column(String(128), nullable=True)
+
+    # --- Foreign Keys ---
     base_model_id = Column(String(128), ForeignKey("base_models.id"))
     fine_tuned_model_id = Column(
         String(64), ForeignKey("fine_tuned_models.id"), nullable=True
     )
+
+    # --- Deployment State ---
     port = Column(Integer, default=8000)
     status = Column(SAEnum(StatusEnum), default=StatusEnum.active)
     current_throughput = Column(Float, default=0.0)
     last_seen = Column(BigInteger, default=lambda: int(time.time()))
 
-    tensor_parallel_size = Column(Integer, nullable=False, default=1)
+    # --- Parallelism ---
+    tensor_parallel_size = Column(
+        Integer,
+        nullable=False,
+        default=1,
+        comment="Number of GPUs for tensor parallelism. 1 = single GPU.",
+    )
+
+    # --- Memory & Context ---
+    gpu_memory_utilization = Column(
+        Float,
+        nullable=False,
+        default=0.90,
+        server_default="0.90",
+        comment="Fraction of GPU VRAM vLLM may allocate. Overrides VLLM_DEFAULT_GPU_MEM_UTIL.",
+    )
+    max_model_len = Column(
+        Integer,
+        nullable=True,
+        default=None,
+        comment="Max sequence length in tokens. Overrides VLLM_DEFAULT_MAX_MODEL_LEN. None = env default.",
+    )
+    max_num_seqs = Column(
+        Integer,
+        nullable=True,
+        default=None,
+        comment="Max concurrent sequences. Critical for vision — each image eats slots. None = vLLM default.",
+    )
+
+    # --- Quantization & Precision ---
+    quantization = Column(
+        String(32),
+        nullable=True,
+        default=None,
+        comment="Quantization scheme: 'awq', 'awq_marlin', 'gptq', 'bitsandbytes', or None for full precision.",
+    )
+    dtype = Column(
+        String(16),
+        nullable=True,
+        default=None,
+        comment="Compute dtype: 'float16', 'bfloat16', 'auto', or None to let vLLM decide.",
+    )
+
+    # --- Runtime Behaviour ---
+    enforce_eager = Column(
+        Boolean,
+        nullable=False,
+        default=False,
+        server_default="0",
+        comment="Disable CUDA graphs. Slower but useful for debugging OOM crashes.",
+    )
+
+    # --- Multimodal Limits ---
+    limit_mm_per_prompt = Column(
+        JSON,
+        nullable=True,
+        default=None,
+        comment='Per-modality token cap per request. e.g. {"image": 2, "video": 0}. None = family registry default.',
+    )
+    mm_processor_kwargs = Column(
+        JSON,
+        nullable=True,
+        default=None,
+        comment=(
+            "Processor-level kwargs passed to vLLM multimodal processor at engine init. "
+            "Overrides family registry defaults (_VISION_FAMILY_CONFIGS) when set. "
+            'Examples: {"min_pixels": 784, "max_pixels": 50176} for Qwen2.5-VL, '
+            '{"num_crops": 4} for Phi-3.5-Vision. '
+            "None = fall back to inference_worker.py family registry defaults."
+        ),
+    )
 
+    # --- Relationships (always last) ---
     base_model = relationship("BaseModel")
     fine_tuned_model = relationship("FineTunedModel")

{projectdavid_orm-1.6.3 → projectdavid_orm-1.8.0/src/projectdavid_orm.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: projectdavid-orm
-Version: 1.6.3
+Version: 1.8.0
 Summary: Shared SQLAlchemy ORM models for ProjectDavid services
 Author-email: "Francis N." <engineering@projectdavid.co.uk>
 License: PolyForm Noncommercial License 1.0.0