ray-embedding 0.12.0__tar.gz → 0.12.1__tar.gz

{ray_embedding-0.12.0 → ray_embedding-0.12.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ray-embedding
-Version: 0.12.0
+Version: 0.12.1
 Summary: Deploy SentenceTransformers embedding models to a ray cluster
 Author: Crispin Almodovar
 Author-email: 

{ray_embedding-0.12.0 → ray_embedding-0.12.1}/ray_embedding/deploy.py

@@ -1,12 +1,10 @@
 import os
-from typing import Dict, Any, Optional
+
+import torch
 from ray.serve import Application
-from ray.serve.handle import DeploymentHandle
 
 from ray_embedding.dto import AppConfig, ModelDeploymentConfig, DeployedModel
 from ray_embedding.embedding_model import EmbeddingModel
-import torch
-
 from ray_embedding.model_router import ModelRouter
 
 
@@ -19,6 +17,7 @@ def build_model(model_config: ModelDeploymentConfig) -> DeployedModel:
     matryoshka_dim = model_config.matryoshka_dim
     trust_remote_code = model_config.trust_remote_code or False
     model_kwargs = model_config.model_kwargs or {}
+
     if "torch_dtype" in model_kwargs:
         torch_dtype = model_kwargs["torch_dtype"].strip()
         if torch_dtype == "float16":

{ray_embedding-0.12.0 → ray_embedding-0.12.1}/ray_embedding/dto.py

@@ -21,6 +21,7 @@ class EmbeddingResponse(BaseModel):
 class ModelRouterConfig(BaseModel):
     deployment: str
     path_prefix: List[str] = []
+    max_concurrency: int = 32
 
 
 class ModelDeploymentConfig(BaseModel):

{ray_embedding-0.12.0 → ray_embedding-0.12.1}/ray_embedding/embedding_model.py

@@ -37,12 +37,14 @@ class EmbeddingModel:
         self.logger.info(f"Successfully initialized model {self.model} using device {self.torch_device}")
 
     async def __call__(self, text: Union[str, List[str]], dimensions: Optional[int] = None) -> List[List[float]]:
-        """Compute embeddings for the input text using the loaded model."""
-        if isinstance(text, str):
-            text = [text]
+        """Compute embeddings for the input text using the current model."""
+        if not text or (isinstance(text, list) and not all(text)):
+            raise ValueError("Input text is empty or invalid")
+
+        text = [text] if isinstance(text, str) else text
         truncate_dim = dimensions or self.matryoshka_dim
 
-        # Compute embeddings and convert to a PyTorch tensor on the GPU
+        # Compute embeddings in PyTorch format
         embeddings = self.embedding_model.encode(
             text, convert_to_tensor=True, normalize_embeddings=True, show_progress_bar=False,
         ).to(self.torch_device)
@@ -56,7 +58,6 @@ class EmbeddingModel:
         embeddings_list = embeddings.cpu().tolist()
         return embeddings_list
 
-
     def wait_for_cuda(self, wait: int = 10):
         if self.init_device == "cuda" and not torch.cuda.is_available():
             time.sleep(wait)
@@ -65,7 +66,7 @@ class EmbeddingModel:
     def check_health(self):
         if self.init_device == "cuda":
             # Even though CUDA was available at init time,
-            # CUDA can become unavailable - this is a known problem in AWS EC2
+            # CUDA can become unavailable - this is a known problem in AWS EC2+Docker
             # https://github.com/ray-project/ray/issues/49594
             try:
                 nvmlInit()

{ray_embedding-0.12.0 → ray_embedding-0.12.1}/ray_embedding/model_router.py

@@ -1,27 +1,29 @@
 import asyncio
 import logging
 import time
-from typing import Optional, Dict, List, Tuple, Any, Coroutine
+from typing import Optional, Dict, List, Tuple
 
 from fastapi import FastAPI, HTTPException
 from ray import serve
 from ray.serve.handle import DeploymentHandle
 
-from ray_embedding.dto import EmbeddingResponse, EmbeddingRequest, DeployedModel
+from ray_embedding.dto import DeployedModel, EmbeddingRequest, EmbeddingResponse
 
 web_api = FastAPI(title="Ray Embeddings - OpenAI-compatible API")
 
 @serve.deployment
 @serve.ingress(web_api)
 class ModelRouter:
-    def __init__(self, deployed_models: Dict[str, DeployedModel], path_prefix: List[str]):
+    def __init__(self, deployed_models: Dict[str, DeployedModel], path_prefix: List[str], max_concurrency: Optional[int] = 32):
         assert deployed_models, "models cannot be empty"
         assert path_prefix, "path_prefix cannot be empty"
 
         logging.basicConfig(level=logging.INFO)
         self.logger = logging.getLogger(self.__class__.__name__)
         self.deployed_models = deployed_models
-        self.path_prefix = path_prefix
+        self.path_prefix = [item.removeprefix("/").removesuffix("/") for item in path_prefix]
+        self.max_concurrency = max_concurrency
+        self.rate_limiter = asyncio.Semaphore(self.max_concurrency)
         self.available_models = [
             {"id": str(item),
              "object": "model",
@@ -43,17 +45,18 @@ class ModelRouter:
             self.logger.info(f"Original input (length {len(inputs)} was resized "
                              f"to {len(batches)} mini-batches, each with max length {batch_size}.")
 
-        # Call embedding model replicas in parallel
-        tasks = [model_handle.remote(batch, dimensions) for batch in batches]
+        # Call embedding model replicas in parallel (rate-limited)
+        tasks = [self._rate_limited_embedding_call(model_handle, batch, dimensions) for batch in batches]
         all_results = await asyncio.gather(*tasks, return_exceptions=True)
 
         # Retry any failed model calls
         for i, result in enumerate(all_results):
             if isinstance(result, Exception):
+                self.logger.warning(f"Retrying mini-batch {i} due to exception: {result}")
                 result_retried, retries = await self._retry_failed_embedding_call(model_handle, batches[i], dimensions,
                                                                                   num_retries)
-                if (retries >= num_retries and isinstance(result_retried, Exception)) or result_retried is None:
-                    raise result_retried or ValueError(f"Failed to compute `{model}` embeddings for mini-batch of size {batch_size}.")
+                if retries >= num_retries and (isinstance(result_retried, Exception) or result_retried is None):
+                    raise result_retried or ValueError(f"Failed to compute `{model}` embeddings for mini-batch {i} after {num_retries} retries.")
 
                 all_results[i] = result_retried
 
@@ -61,6 +64,10 @@ class ModelRouter:
         self.logger.info(f"Successfully computed embeddings from {len(batches)} mini-batches")
         return [emb for result in all_results for emb in result]
 
+    async def _rate_limited_embedding_call(self, model_handle: DeploymentHandle, batch: List[str], dimensions: int):
+        with self.rate_limiter:
+            return await model_handle.remote(batch, dimensions)
+
     async def _retry_failed_embedding_call(self,  model_handle: DeploymentHandle, batch: List[str],
                                            dimensions: Optional[int] = None, num_retries: Optional[int] = 2) \
             -> Tuple[List[List[float]] | Exception, int]:

{ray_embedding-0.12.0 → ray_embedding-0.12.1}/ray_embedding.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ray-embedding
-Version: 0.12.0
+Version: 0.12.1
 Summary: Deploy SentenceTransformers embedding models to a ray cluster
 Author: Crispin Almodovar
 Author-email: 

{ray_embedding-0.12.0 → ray_embedding-0.12.1}/setup.cfg

@@ -1,6 +1,6 @@
 [metadata]
 name = ray-embedding
-version = 0.12.0
+version = 0.12.1
 author = Crispin Almodovar
 author_email = 
 description = Deploy SentenceTransformers embedding models to a ray cluster