ray-embedding 0.9.10__py3-none-any.whl → 0.10.1__py3-none-any.whl

ray_embedding/embedding_model.py

@@ -24,10 +24,6 @@ web_api = FastAPI(title=f"Ray Embeddings - OpenAI-compatible API")
         "min_replicas": 0,
         "initial_replicas": 1,
         "max_replicas": 1,
-    },
-    user_config={
-        "max_batch_size": 8,
-        "batch_wait_timeout_s": 0.25,
     }
 )
 @serve.ingress(web_api)
@@ -56,12 +52,6 @@ class EmbeddingModel:
         ]
         self.logger.info(f"Successfully initialized embedding model {self.model} using device {self.torch_device}")
 
-    async def reconfigure(self, user_config: Dict):
-        assert "max_batch_size" in user_config and "batch_wait_timeout_s" in user_config, "Invalid user config"
-        self.logger.info(f"Reconfiguring dynamic batching parameters: {user_config}")
-        self.__create_embeddings_batch.set_max_batch_size(user_config["max_batch_size"])
-        self.__create_embeddings_batch.set_batch_wait_timeout_s(user_config["batch_wait_timeout_s"])
-
     @web_api.post("/v1/embeddings", response_model=EmbeddingResponse)
     async def create_embeddings(self, request: EmbeddingRequest):
         """Generate embeddings for the input text using the specified model."""
@@ -69,59 +59,35 @@ class EmbeddingModel:
             assert request.model == self.served_model_name, (
                 f"Model '{request.model}' is not supported. Use '{self.served_model_name}' instead."
             )
-            return await self.__create_embeddings_batch(request)
+            if isinstance(request.input, str):
+                request.input = [request.input]
+
+            truncate_dim = request.dimensions or self.matryoshka_dim
+
+            # Compute embeddings and convert to a PyTorch tensor on the GPU
+            embeddings = self.embedding_model.encode(
+                request.input, convert_to_tensor=True, normalize_embeddings=True, show_progress_bar=False,
+            ).to(self.torch_device)
+
+            if truncate_dim is not None:
+                # Truncate and re-normalize the embeddings
+                embeddings = embeddings[:, :truncate_dim]
+                embeddings = embeddings / torch.norm(embeddings, dim=1, keepdim=True)
+
+            # Move all embeddings to CPU at once before conversion
+            embeddings = embeddings.cpu().tolist()
+
+            # Convert embeddings to list format for response
+            response_data = [
+                {"index": idx, "embedding": emb}
+                for idx, emb in enumerate(embeddings)
+            ]
+            return EmbeddingResponse(object="list", data=response_data, model=request.model)
+
         except Exception as e:
             self.logger.error(e)
             raise HTTPException(status_code=500, detail=str(e))
 
-    @serve.batch(max_batch_size=8, batch_wait_timeout_s=0.25)
-    async def __create_embeddings_batch(self, requests_batch: List[EmbeddingRequest]) -> List[EmbeddingResponse]:
-        self_0 = self[0] if isinstance(self, list) else self  # Ray also passes an array of self refs; just take the first one
-        embedding_model, matryoshka_dim, torch_device = self_0.embedding_model, self_0.matryoshka_dim, self_0.torch_device
-
-        inputs, truncate_dims, num_inputs_list = [], [], []
-        for request in requests_batch:
-            input_text = request.input if isinstance(request.input, list) else [request.input] # Can be a list of texts
-            inputs.extend(input_text)
-            num_inputs_list.append(len(input_text))
-            truncate_dims.append(request.dimensions or matryoshka_dim)
-
-        embeddings = embedding_model.encode(
-            inputs, convert_to_tensor=True, normalize_embeddings=True, show_progress_bar=False,
-        ).to(torch_device)
-
-        model_name = requests_batch[0].model
-        truncate_needed = any(dim is not None for dim in truncate_dims)
-        results_batch, ix = [], 0
-
-        if truncate_needed:
-            for truncate_dim, num_inputs in zip(truncate_dims, num_inputs_list):
-                batch_embeddings = embeddings[ix: ix + num_inputs]
-                ix += num_inputs
-
-                if truncate_dim is not None:
-                    # Truncate and normalize using pytorch (faster)
-                    batch_embeddings = batch_embeddings[:, :truncate_dim]
-                    batch_embeddings = batch_embeddings / torch.norm(batch_embeddings, dim=1, keepdim=True)
-
-                batch_embeddings = batch_embeddings.cpu().tolist()
-                response_data = [
-                    {"index": emb_ix, "embedding": emb} for emb_ix, emb in enumerate(batch_embeddings)
-                ]
-                results_batch.append(EmbeddingResponse(object="list", data=response_data, model=model_name))
-        else:
-            embeddings_list = embeddings.cpu().tolist() # Move everything to CPU
-            for num_inputs in num_inputs_list:
-                batch_embeddings = embeddings_list[ix: ix + num_inputs]
-                ix += num_inputs
-
-                response_data = [
-                    {"index": emb_ix, "embedding": emb} for emb_ix, emb in enumerate(batch_embeddings)
-                ]
-                results_batch.append(EmbeddingResponse(object="list", data=response_data, model=model_name))
-
-        return results_batch
-
     @web_api.get("/v1/models")
     async def list_models(self):
         """Returns the list of available models in OpenAI-compatible format."""

{ray_embedding-0.9.10.dist-info → ray_embedding-0.10.1.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ray-embedding
-Version: 0.9.10
+Version: 0.10.1
 Summary: Deploy SentenceTransformers embedding models to a ray cluster
 Author: Crispin Almodovar
 Author-email: 

ray_embedding-0.10.1.dist-info/RECORD

@@ -0,0 +1,8 @@
+ray_embedding/__init__.py,sha256=OYJT0rVaaGzY613JqgfktsCgroDnBkGOHxR2FE9UtRU,49
+ray_embedding/deploy.py,sha256=oqOhMxS5UyZ4oGhfpL7kqvrxLO8QW41sF_FHbNSJe-w,1858
+ray_embedding/dto.py,sha256=e91ejZbM_NB9WTjF1YnfuV71cajYIh0vOX8oV_g2OwM,595
+ray_embedding/embedding_model.py,sha256=54QQ5iIAnOHtg_8nqGPIfqrZVwih6qKKeEOgmZweQeQ,3931
+ray_embedding-0.10.1.dist-info/METADATA,sha256=E9sZzIgqproOjSMt1ujPha_BhuBHLtd3ipTXD8vJSAU,605
+ray_embedding-0.10.1.dist-info/WHEEL,sha256=GHB6lJx2juba1wDgXDNlMTyM13ckjBMKf-OnwgKOCtA,91
+ray_embedding-0.10.1.dist-info/top_level.txt,sha256=ziCblpJq1YsrryshFqxTRuRMgNuO1_tgvAAkGShATNA,14
+ray_embedding-0.10.1.dist-info/RECORD,,

{ray_embedding-0.9.10.dist-info → ray_embedding-0.10.1.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: setuptools (79.0.1)
+Generator: setuptools (80.3.0)
 Root-Is-Purelib: true
 Tag: py3-none-any
 

ray_embedding-0.9.10.dist-info/RECORD

@@ -1,8 +0,0 @@
-ray_embedding/__init__.py,sha256=OYJT0rVaaGzY613JqgfktsCgroDnBkGOHxR2FE9UtRU,49
-ray_embedding/deploy.py,sha256=oqOhMxS5UyZ4oGhfpL7kqvrxLO8QW41sF_FHbNSJe-w,1858
-ray_embedding/dto.py,sha256=e91ejZbM_NB9WTjF1YnfuV71cajYIh0vOX8oV_g2OwM,595
-ray_embedding/embedding_model.py,sha256=DyhFO0kQnyoFq_VGJQYDEZKuvtwqhuXr83Vu23yb4ds,5966
-ray_embedding-0.9.10.dist-info/METADATA,sha256=aDU4KlY1_-mtJ80tPb9FGGAnB9o2VSjQtsUR3Xt60_k,605
-ray_embedding-0.9.10.dist-info/WHEEL,sha256=SmOxYU7pzNKBqASvQJ7DjX3XGUF92lrGhMb3R6_iiqI,91
-ray_embedding-0.9.10.dist-info/top_level.txt,sha256=ziCblpJq1YsrryshFqxTRuRMgNuO1_tgvAAkGShATNA,14
-ray_embedding-0.9.10.dist-info/RECORD,,