ray-embedding 0.11.8__py3-none-any.whl → 0.12.0__py3-none-any.whl

ray_embedding/deploy.py

@@ -48,7 +48,8 @@ def build_model(model_config: ModelDeploymentConfig) -> DeployedModel:
 def build_app(args: AppConfig) -> Application:
     model_router, models = args.model_router, args.models
     assert model_router and models
+    assert model_router.path_prefix
 
     deployed_models = {model_config.served_model_name: build_model(model_config) for model_config in models}
-    router = ModelRouter.options(name=model_router.deployment).bind(deployed_models)
+    router = ModelRouter.options(name=model_router.deployment).bind(deployed_models, model_router.path_prefix)
     return router

ray_embedding/dto.py

@@ -20,6 +20,7 @@ class EmbeddingResponse(BaseModel):
 
 class ModelRouterConfig(BaseModel):
     deployment: str
+    path_prefix: List[str] = []
 
 
 class ModelDeploymentConfig(BaseModel):

ray_embedding/embedding_model.py

@@ -34,9 +34,9 @@ class EmbeddingModel:
                                                    trust_remote_code=self.trust_remote_code,
                                                    model_kwargs=self.model_kwargs)
 
-        self.logger.info(f"Successfully initialized embedding model {self.model} using device {self.torch_device}")
+        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):
+    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]
@@ -53,8 +53,8 @@ class EmbeddingModel:
             embeddings = embeddings / torch.norm(embeddings, dim=1, keepdim=True)
 
         # Move all embeddings to CPU at once before conversion
-        embeddings = embeddings.cpu().tolist()
-        return embeddings
+        embeddings_list = embeddings.cpu().tolist()
+        return embeddings_list
 
 
     def wait_for_cuda(self, wait: int = 10):

ray_embedding/model_router.py

@@ -1,10 +1,11 @@
 import asyncio
 import logging
 import time
-from typing import Optional, Dict, List
+from typing import Optional, Dict, List, Tuple, Any, Coroutine
 
 from fastapi import FastAPI, HTTPException
 from ray import serve
+from ray.serve.handle import DeploymentHandle
 
 from ray_embedding.dto import EmbeddingResponse, EmbeddingRequest, DeployedModel
 
@@ -13,11 +14,14 @@ 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]):
+    def __init__(self, deployed_models: Dict[str, DeployedModel], path_prefix: List[str]):
         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.available_models = [
             {"id": str(item),
              "object": "model",
@@ -28,7 +32,6 @@ class ModelRouter:
         self.logger.info(f"Successfully registered models: {self.available_models}")
 
     async def _compute_embeddings_from_resized_batches(self, model: str, inputs: List[str], dimensions: Optional[int] = None):
-        assert model in self.deployed_models
         deployed_model = self.deployed_models[model]
         model_handle = deployed_model.deployment_handle
         batch_size = deployed_model.batch_size
@@ -38,33 +41,49 @@ class ModelRouter:
         batches = [inputs[i:i+batch_size] for i in range(0, len(inputs), batch_size)]
         if len(inputs) > batch_size:
             self.logger.info(f"Original input (length {len(inputs)} was resized "
-                             f"to {len(batches)} mini-batches of length {batch_size}")
+                             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]
         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):
-                retries = 0
-                while retries < num_retries:
-                    try:
-                        all_results[i] = await model_handle.remote(batches[i], dimensions)
-                    except Exception as e:
-                        self.logger.warning(e)
-                    finally:
-                        retries += 1
-                    if not isinstance(all_results[i], Exception):
-                        break
-
-                if retries >= num_retries and isinstance(all_results[i], Exception):
-                    raise all_results[i]
-
-        # Flatten the results because all_results is a list of lists
+                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}.")
+
+                all_results[i] = result_retried
+
+        # Flatten the results because `all_results` is a list of lists
         self.logger.info(f"Successfully computed embeddings from {len(batches)} mini-batches")
         return [emb for result in all_results for emb in result]
 
-    @web_api.post("/v1/embeddings", response_model=EmbeddingResponse)
-    async def compute_embeddings(self, request: EmbeddingRequest):
+    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]:
+
+        result_retried, retries = None, 0
+        while retries < num_retries:
+            try:
+                result_retried = await model_handle.remote(batch, dimensions)
+            except Exception as e:
+                result_retried = e
+                self.logger.warning(e)
+            finally:
+                retries += 1
+            if not isinstance(result_retried, Exception) and result_retried is not None:
+                break
+
+        return result_retried, retries
+
+    @web_api.post("/{path_prefix}/v1/embeddings", response_model=EmbeddingResponse)
+    async def compute_embeddings(self, path_prefix: str, request: EmbeddingRequest):
+        assert path_prefix in self.path_prefix, f"Invalid path prefix: {path_prefix}"
+        assert request.model in self.deployed_models, f"Invalid model: {request.model}"
+
         try:
             inputs = request.input if isinstance(request.input, list) else [request.input]
             self.logger.info(f"Computing embeddings for a batch of {len(inputs)} texts using model: {request.model}")
@@ -78,7 +97,8 @@ class ModelRouter:
             self.logger.error(f"Failed to create embeddings: {e}")
             raise HTTPException(status_code=500, detail=str(e))
 
-    @web_api.get("/v1/models")
-    async def list_models(self):
+    @web_api.get("/{path_prefix}/v1/models")
+    async def list_models(self, path_prefix: str):
         """Returns the list of available models in OpenAI-compatible format."""
-        return {"object": "list", "data": self.available_models}
+        assert path_prefix in self.path_prefix, f"Invalid path prefix: {path_prefix}"
+        return {"object": "list", "data": self.available_models}

{ray_embedding-0.11.8.dist-info → ray_embedding-0.12.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ray-embedding
-Version: 0.11.8
+Version: 0.12.0
 Summary: Deploy SentenceTransformers embedding models to a ray cluster
 Author: Crispin Almodovar
 Author-email: 

ray_embedding-0.12.0.dist-info/RECORD

@@ -0,0 +1,9 @@
+ray_embedding/__init__.py,sha256=YS5LAZfRIwwVvE3C9g7hsauvjgIkqKtHyxkwMFFfAGY,46
+ray_embedding/deploy.py,sha256=_p8LXLfRP2y9z4Jj1BbuVTuDYkuXZpGI_JkTEj_bMa4,2712
+ray_embedding/dto.py,sha256=NwS8EkeZZcfWDE6RFsLG0WtZtnc7onlr95llRSgnIQc,1432
+ray_embedding/embedding_model.py,sha256=cZe6voJTXYz0OKjjkFskC05V8frm3B-B3-Ae1j9GMKo,3408
+ray_embedding/model_router.py,sha256=04PN_ZptFwEGeSSaDVbCodbDMRMu2i-jeamC1Sb2v24,5145
+ray_embedding-0.12.0.dist-info/METADATA,sha256=jSrtclfqYRIfyd5e1KErc3DvDjXSB2TZEvXSyi-aGqk,1094
+ray_embedding-0.12.0.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+ray_embedding-0.12.0.dist-info/top_level.txt,sha256=ziCblpJq1YsrryshFqxTRuRMgNuO1_tgvAAkGShATNA,14
+ray_embedding-0.12.0.dist-info/RECORD,,

ray_embedding-0.11.8.dist-info/RECORD

@@ -1,9 +0,0 @@
-ray_embedding/__init__.py,sha256=YS5LAZfRIwwVvE3C9g7hsauvjgIkqKtHyxkwMFFfAGY,46
-ray_embedding/deploy.py,sha256=fHyxSGOcNTD_EFmPC2QYugvR77Ulr92bNU6Xk2Mwwf4,2649
-ray_embedding/dto.py,sha256=QlduDoqkFHaeF_KgsFeUKq2XWiPMmrgRPy_QjCTSCRE,1399
-ray_embedding/embedding_model.py,sha256=2wLk54BZIhHMCnwx5vneU0z4Y7EQs220BFNeLf-UQh4,3387
-ray_embedding/model_router.py,sha256=q7SgfOEyslbzKe-_gbIq02T2KLpUI2DmDivdmPf-_ts,3906
-ray_embedding-0.11.8.dist-info/METADATA,sha256=jRwJmh8p85_h7w8zuyLqYvS3nBE1f_YYzaR56rNOmXc,1094
-ray_embedding-0.11.8.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-ray_embedding-0.11.8.dist-info/top_level.txt,sha256=ziCblpJq1YsrryshFqxTRuRMgNuO1_tgvAAkGShATNA,14
-ray_embedding-0.11.8.dist-info/RECORD,,