feat: rename dynamo decorator (#1133)

examples/multimodal/components/encode_worker.py

@@ -24,7 +24,7 @@ from transformers import AutoImageProcessor, LlavaForConditionalGeneration
 from utils.protocol import EncodeRequest, EncodeResponse
 from utils.vllm import parse_vllm_args
 
-from dynamo.sdk import dynamo_endpoint, service
+from dynamo.sdk import endpoint, service
 
 logger = logging.getLogger(__name__)
 
@@ -50,7 +50,7 @@ class EncodeWorker:
             self.MODEL_ID, device_map="auto", torch_dtype=torch.float16
         ).eval()
 
-    @dynamo_endpoint()
+    @endpoint()
     async def encode(self, request: EncodeRequest) -> AsyncIterator[EncodeResponse]:
         image = self.open_image(request.image_url)
         image_embeds = self.image_processor(images=image, return_tensors="pt")

examples/multimodal/components/frontend.py

@@ -20,7 +20,7 @@ from fastapi import FastAPI
 from fastapi.responses import StreamingResponse
 from utils.protocol import MultiModalRequest
 
-from dynamo.sdk import DYNAMO_IMAGE, depends, dynamo_api, service
+from dynamo.sdk import DYNAMO_IMAGE, api, depends, service
 
 logger = logging.getLogger(__name__)
 
@@ -37,7 +37,7 @@ logger = logging.getLogger(__name__)
 class Frontend:
     processor = depends(Processor)
 
-    @dynamo_api()
+    @api()
     async def generate(self, request: MultiModalRequest):
         async def content_generator():
             async for response in self.processor.generate(request.model_dump_json()):

examples/multimodal/components/prefill_worker.py

@@ -34,7 +34,7 @@ from vllm.entrypoints.openai.api_server import (
 from vllm.inputs.data import TokensPrompt
 from vllm.remote_prefill import RemotePrefillParams, RemotePrefillRequest
 
-from dynamo.sdk import async_on_start, depends, dynamo_context, dynamo_endpoint, service
+from dynamo.sdk import async_on_start, depends, dynamo_context, endpoint, service
 
 logger = logging.getLogger(__name__)
 
@@ -223,6 +223,6 @@ class PrefillWorker:
         ):
             yield
 
-    @dynamo_endpoint()
+    @endpoint()
     async def mock(self, req: RequestType):
         yield f"mock_response: {req}"

examples/multimodal/components/processor.py

@@ -31,7 +31,7 @@ from vllm.outputs import RequestOutput
 from vllm.transformers_utils.tokenizer import AnyTokenizer
 
 from dynamo.runtime import EtcdKvCache
-from dynamo.sdk import async_on_start, depends, dynamo_context, dynamo_endpoint, service
+from dynamo.sdk import async_on_start, depends, dynamo_context, endpoint, service
 
 logger = logging.getLogger(__name__)
 
@@ -195,7 +195,7 @@ class Processor(ProcessMixIn):
                 )
 
     # The generate endpoint will be used by the frontend to handle incoming requests.
-    @dynamo_endpoint()
+    @endpoint()
     async def generate(self, request: MultiModalRequest):
         # TODO: After having the multimodal support in OpenAI compatible frontend, we can use that directly and remove the custom endpoint.
         msg = {

examples/multimodal/components/worker.py

@@ -41,7 +41,7 @@ from vllm.inputs.data import TokensPrompt
 from vllm.remote_prefill import RemotePrefillParams, RemotePrefillRequest
 from vllm.sampling_params import RequestOutputKind
 
-from dynamo.sdk import async_on_start, depends, dynamo_context, dynamo_endpoint, service
+from dynamo.sdk import async_on_start, depends, dynamo_context, endpoint, service
 
 logger = logging.getLogger(__name__)
 
@@ -175,7 +175,7 @@ class VllmWorker:
 
         return callback
 
-    @dynamo_endpoint()
+    @endpoint()
     async def generate(self, request: vLLMMultimodalRequest):
         image_features = None
         if self.do_remote_prefill:

examples/sglang/components/decode_worker.py

@@ -21,7 +21,7 @@ import sglang as sgl
 from utils.protocol import DisaggPreprocessedRequest
 from utils.sglang import parse_sglang_args
 
-from dynamo.sdk import dynamo_endpoint, service
+from dynamo.sdk import endpoint, service
 
 logger = logging.getLogger(__name__)
 
@@ -42,7 +42,7 @@ class SGLangDecodeWorker:
 
         logger.warning("Decode worker initialized")
 
-    @dynamo_endpoint()
+    @endpoint()
     async def generate(self, req: DisaggPreprocessedRequest):
         g = await self.engine.async_generate(
             input_ids=req.request.token_ids,

examples/sglang/components/worker.py

@@ -36,7 +36,7 @@ from utils.protocol import DisaggPreprocessedRequest, PreprocessedRequest
 from utils.sglang import parse_sglang_args
 
 from dynamo.llm import ModelType, register_llm
-from dynamo.sdk import async_on_start, depends, dynamo_context, dynamo_endpoint, service
+from dynamo.sdk import async_on_start, depends, dynamo_context, endpoint, service
 
 logger = logging.getLogger(__name__)
 
@@ -112,7 +112,7 @@ class SGLangWorker:
             sampling_params["ignore_eos"] = request.stop_conditions.ignore_eos
         return sampling_params
 
-    @dynamo_endpoint()
+    @endpoint()
     async def generate(self, request: PreprocessedRequest):
         # TODO: maintain a mapping from SGLang's Ouput struct to LLMEngineOuput
         sampling_params = self._build_sampling_params(request)

examples/tensorrt_llm/components/kv_router.py

@@ -25,7 +25,7 @@ from common.protocol import Tokens
 from components.worker import TensorRTLLMWorker
 
 from dynamo.llm import AggregatedMetrics, KvIndexer, KvMetricsAggregator, OverlapScores
-from dynamo.sdk import async_on_start, depends, dynamo_context, dynamo_endpoint, service
+from dynamo.sdk import async_on_start, depends, dynamo_context, endpoint, service
 from dynamo.sdk.lib.config import ServiceConfig
 
 logger = logging.getLogger(__name__)
@@ -209,7 +209,7 @@ class Router:
 
         return best_worker_id, worker_scores.get(best_worker_id, 0.0)
 
-    @dynamo_endpoint()
+    @endpoint()
     async def generate(self, request: Tokens) -> AsyncIterator[WorkerId]:
         if self.indexer is None or self.metrics_aggregator is None:
             yield "_0.0"

examples/tensorrt_llm/components/prefill_worker.py

@@ -20,7 +20,7 @@ from common.parser import parse_tensorrt_llm_args
 from common.protocol import TRTLLMWorkerRequest
 from common.utils import ServerType
 
-from dynamo.sdk import async_on_start, dynamo_context, dynamo_endpoint, service
+from dynamo.sdk import async_on_start, dynamo_context, endpoint, service
 from dynamo.sdk.lib.config import ServiceConfig
 
 logger = logging.getLogger(__name__)
@@ -68,7 +68,7 @@ class TensorRTLLMPrefillWorker(BaseTensorrtLLMEngine):
         component = dynamo_context["component"]
         await self.kv_metrics_publisher.create_endpoint(component)
 
-    @dynamo_endpoint()
+    @endpoint()
     async def generate(self, request: TRTLLMWorkerRequest):
         async for response in super().generate(request):
             yield response

examples/tensorrt_llm/components/processor.py

@@ -27,7 +27,7 @@ from common.utils import RequestType
 from components.kv_router import Router
 from components.worker import TensorRTLLMWorker
 
-from dynamo.sdk import async_on_start, depends, dynamo_context, dynamo_endpoint, service
+from dynamo.sdk import async_on_start, depends, dynamo_context, endpoint, service
 from dynamo.sdk.lib.config import ServiceConfig
 
 logger = logging.getLogger(__name__)
@@ -143,7 +143,7 @@ class Processor(ChatProcessorMixin):
                 logger.debug(f"[preprocessor] Response: {response}")
                 yield json.loads(response)
 
-    @dynamo_endpoint(name="chat/completions")
+    @endpoint(name="chat/completions")
     async def generate_chat(self, raw_request: DynamoTRTLLMChatCompletionRequest):
         # max_tokens is deprecated, however if the max_tokens is provided instead
         # of max_completion_tokens, we will use the value as max_completion_tokens.
@@ -172,7 +172,7 @@ class Processor(ChatProcessorMixin):
         async for response in self._generate(raw_request, RequestType.CHAT):
             yield response
 
-    @dynamo_endpoint(name="completions")
+    @endpoint(name="completions")
     async def completions(self, raw_request: DynamoTRTLLMCompletionRequest):
         # min_tokens isn't currently propagated through the Rust OpenAI HTTP frontend,
         # and ignore_eos is passed through the 'nvext' field, so set both when found.

examples/tensorrt_llm/components/worker.py

@@ -21,7 +21,7 @@ from common.protocol import TRTLLMWorkerRequest
 from common.utils import ServerType
 from components.prefill_worker import TensorRTLLMPrefillWorker
 
-from dynamo.sdk import async_on_start, depends, dynamo_context, dynamo_endpoint, service
+from dynamo.sdk import async_on_start, depends, dynamo_context, endpoint, service
 from dynamo.sdk.lib.config import ServiceConfig
 
 logger = logging.getLogger(__name__)
@@ -91,7 +91,7 @@ class TensorRTLLMWorker(BaseTensorrtLLMEngine):
         component = dynamo_context["component"]
         await self._kv_metrics_publisher.create_endpoint(component)
 
-    @dynamo_endpoint()
+    @endpoint()
     async def generate(self, request: TRTLLMWorkerRequest):
         async for response in super().generate(request):
             yield response

examples/vllm_v1/components/simple_load_balancer.py

@@ -25,7 +25,7 @@ from vllm.inputs import TokensPrompt
 from vllm.sampling_params import SamplingParams
 
 from dynamo.llm import ModelType, register_llm
-from dynamo.sdk import async_on_start, depends, dynamo_context, dynamo_endpoint, service
+from dynamo.sdk import async_on_start, depends, dynamo_context, endpoint, service
 
 logger = logging.getLogger(__name__)
 
@@ -129,7 +129,7 @@ class SimpleLoadBalancer:
         ):
             yield MyRequestOutput.model_validate_json(decode_response.data())
 
-    @dynamo_endpoint()
+    @endpoint()
     async def generate(self, request: PreprocessedRequest):
         logger.debug(
             "Processor received completion request: %s", request.model_dump_json()

examples/vllm_v1/components/worker.py

@@ -27,7 +27,7 @@ from vllm.entrypoints.openai.api_server import (
     build_async_engine_client_from_engine_args,
 )
 
-from dynamo.sdk import async_on_start, dynamo_endpoint, service
+from dynamo.sdk import async_on_start, endpoint, service
 
 logger = logging.getLogger(__name__)
 
@@ -65,7 +65,7 @@ class VllmBaseWorker:
         finally:
             loop.stop()
 
-    @dynamo_endpoint()
+    @endpoint()
     async def generate(self, request: vLLMGenerateRequest):
         gen = self.engine_client.generate(
             prompt=request.prompt,