Add instructions for running vLLM backend

README.md

@@ -0,0 +1,93 @@
+<!--
+# Copyright 2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions
+# are met:
+#  * Redistributions of source code must retain the above copyright
+#    notice, this list of conditions and the following disclaimer.
+#  * Redistributions in binary form must reproduce the above copyright
+#    notice, this list of conditions and the following disclaimer in the
+#    documentation and/or other materials provided with the distribution.
+#  * Neither the name of NVIDIA CORPORATION nor the names of its
+#    contributors may be used to endorse or promote products derived
+#    from this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
+# EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+# PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
+# CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+# EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+# PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+# PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
+# OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+-->
+
+[![License](https://img.shields.io/badge/License-BSD3-lightgrey.svg)](https://opensource.org/licenses/BSD-3-Clause)
+
+# vLLM Backend
+
+The Triton backend for [vLLM](https://github.com/vllm-project/vllm).
+You can learn more about Triton backends in the [backend
+repo](https://github.com/triton-inference-server/backend). Ask
+questions or report problems on the [issues
+page](https://github.com/triton-inference-server/server/issues).
+This backend is designed to run vLLM's
+[supported HuggingFace models](https://vllm.readthedocs.io/en/latest/models/supported_models.html).
+
+Where can I ask general questions about Triton and Triton backends?
+Be sure to read all the information below as well as the [general
+Triton documentation](https://github.com/triton-inference-server/server#triton-inference-server)
+available in the main [server](https://github.com/triton-inference-server/server)
+repo. If you don't find your answer there you can ask questions on the
+main Triton [issues page](https://github.com/triton-inference-server/server/issues).
+
+## Build the vLLM Backend
+
+As a Python-based backend, your Triton server just needs to have the (Python backend)[https://github.com/triton-inference-server/python_backend]
+built under `/opt/tritonserver/backends/python`. After that, you can save this in the backends folder as `/opt/tritonserver/backends/vllm`. The `model.py` file in the `src` directory should be in the vllm folder and will function as your Python-based backend. 
+
+In other words, there are no build steps. You only need to copy this to your Triton backends repository. If you use the official Triton vLLM container, this is already set up for you.
+
+The backend repository should look like this:
+```
+/opt/tritonserver/backends/
+`-- vllm
+    |-- model.py
+ -- python
+    |-- libtriton_python.so
+    |-- triton_python_backend_stub
+    |-- triton_python_backend_utils.py
+```
+
+
+## Using the vLLM Backend
+
+You can see an example model_repository in the `samples` folder.
+You can use this as is and change the model by changing the `model` value in `model.json`.
+You can change the GPU utilization and logging in that file as well.
+
+In the `samples` folder, you can also find a sample client, `client.py`.
+This client is meant to function similarly to the Triton
+(vLLM example)[https://github.com/triton-inference-server/tutorials/tree/main/Quick_Deploy/vLLM].
+By default, this will test `prompts.txt`, which we have included in the samples folder.
+
+
+## Important Notes
+
+* At present, Triton only supports one Python-based backend per server. If you try to start multiple vLLM models, you will get an error.
+
+### Running Multiple Instances of Triton Server
+
+Python-based backends use shared memory to transfer requests to the stub process. When running multiple instances of Triton Server on the same machine that use Python-based backend models, there would be shared memory region name conflicts that can result in segmentation faults or hangs. In order to avoid this issue, you need to specify different shm-region-prefix-name using the --backend-config flag.
+```
+# Triton instance 1
+tritonserver --model-repository=/models --backend-config=python,shm-region-prefix-name=prefix1
+
+# Triton instance 2
+tritonserver --model-repository=/models --backend-config=python,shm-region-prefix-name=prefix2
+```
+Note that the hangs would only occur if the /dev/shm is shared between the two instances of the server. If you run the servers in different containers that don't share this location, you don't need to specify shm-region-prefix-name.

samples/client.py

@@ -0,0 +1,202 @@
+# Copyright 2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions
+# are met:
+#  * Redistributions of source code must retain the above copyright
+#    notice, this list of conditions and the following disclaimer.
+#  * Redistributions in binary form must reproduce the above copyright
+#    notice, this list of conditions and the following disclaimer in the
+#    documentation and/or other materials provided with the distribution.
+#  * Neither the name of NVIDIA CORPORATION nor the names of its
+#    contributors may be used to endorse or promote products derived
+#    from this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
+# EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+# PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
+# CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+# EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+# PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+# PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
+# OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+import argparse
+import asyncio
+import queue
+import sys
+from os import system
+import json
+
+import numpy as np
+import tritonclient.grpc.aio as grpcclient
+from tritonclient.utils import *
+
+
+
+def create_request(prompt, stream, request_id, sampling_parameters, model_name, send_parameters_as_tensor=True):
+    inputs = []
+    prompt_data = np.array([prompt.encode("utf-8")], dtype=np.object_)
+    try:
+        inputs.append(grpcclient.InferInput("PROMPT", [1], "BYTES"))
+        inputs[-1].set_data_from_numpy(prompt_data)
+    except Exception as e:
+        print(f"Encountered an error {e}")
+
+    stream_data = np.array([stream], dtype=bool)
+    inputs.append(grpcclient.InferInput("STREAM", [1], "BOOL"))
+    inputs[-1].set_data_from_numpy(stream_data)
+
+    # Request parameters are not yet supported via BLS. Provide an
+    # optional mechanism to send serialized parameters as an input
+    # tensor until support is added
+
+    if send_parameters_as_tensor:
+        sampling_parameters_data = np.array(
+            [json.dumps(sampling_parameters).encode("utf-8")], dtype=np.object_
+        )
+        inputs.append(grpcclient.InferInput("SAMPLING_PARAMETERS", [1], "BYTES"))
+        inputs[-1].set_data_from_numpy(sampling_parameters_data)
+
+    # Add requested outputs
+    outputs = []
+    outputs.append(grpcclient.InferRequestedOutput("TEXT"))
+
+    # Issue the asynchronous sequence inference.
+    return {
+        "model_name": model_name,
+        "inputs": inputs,
+        "outputs": outputs,
+        "request_id": str(request_id),
+        "parameters": sampling_parameters
+    }
+
+
+async def main(FLAGS):
+    model_name = "vllm_opt"
+    sampling_parameters = {"temperature": "0.1", "top_p": "0.95"}
+    stream = FLAGS.streaming_mode
+    with open(FLAGS.input_prompts, "r") as file:
+        print(f"Loading inputs from `{FLAGS.input_prompts}`...")
+        prompts = file.readlines()
+
+    results_dict = {}
+
+    async with grpcclient.InferenceServerClient(
+        url=FLAGS.url, verbose=FLAGS.verbose
+    ) as triton_client:
+        # Request iterator that yields the next request
+        async def async_request_iterator():
+            try:
+                for iter in range(FLAGS.iterations):
+                    for i, prompt in enumerate(prompts):
+                        prompt_id = FLAGS.offset + (len(prompts) * iter) + i
+                        results_dict[str(prompt_id)] = []
+                        yield create_request(
+                            prompt, stream, prompt_id, sampling_parameters, model_name
+                        )
+            except Exception as error:
+                print(f"caught error in request iterator:  {error}")
+
+        try:
+            # Start streaming
+            response_iterator = triton_client.stream_infer(
+                inputs_iterator=async_request_iterator(),
+                stream_timeout=FLAGS.stream_timeout,
+            )
+            # Read response from the stream
+            async for response in response_iterator:
+                result, error = response
+                if error:
+                    print(f"Encountered error while processing: {error}")
+                else:
+                    output = result.as_numpy("TEXT")
+                    for i in output:
+                        results_dict[result.get_response().id].append(i)
+
+        except InferenceServerException as error:
+            print(error)
+            sys.exit(1)
+
+    with open(FLAGS.results_file, "w") as file:
+        for id in results_dict.keys():
+            for result in results_dict[id]:
+                file.write(result.decode("utf-8"))
+                file.write("\n")
+            file.write("\n=========\n\n")
+        print(f"Storing results into `{FLAGS.results_file}`...")
+
+    if FLAGS.verbose:
+        print(f"\nContents of `{FLAGS.results_file}` ===>")
+        system(f"cat {FLAGS.results_file}")
+
+    print("PASS: vLLM example")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "-v",
+        "--verbose",
+        action="store_true",
+        required=False,
+        default=False,
+        help="Enable verbose output",
+    )
+    parser.add_argument(
+        "-u",
+        "--url",
+        type=str,
+        required=False,
+        default="localhost:8001",
+        help="Inference server URL and it gRPC port. Default is localhost:8001.",
+    )
+    parser.add_argument(
+        "-t",
+        "--stream-timeout",
+        type=float,
+        required=False,
+        default=None,
+        help="Stream timeout in seconds. Default is None.",
+    )
+    parser.add_argument(
+        "--offset",
+        type=int,
+        required=False,
+        default=0,
+        help="Add offset to request IDs used",
+    )
+    parser.add_argument(
+        "--input-prompts",
+        type=str,
+        required=False,
+        default="prompts.txt",
+        help="Text file with input prompts",
+    )
+    parser.add_argument(
+        "--results-file",
+        type=str,
+        required=False,
+        default="results.txt",
+        help="The file with output results",
+    )
+    parser.add_argument(
+        "--iterations",
+        type=int,
+        required=False,
+        default=1,
+        help="Number of iterations through the prompts file",
+    )
+    parser.add_argument(
+        "-s",
+        "--streaming-mode",
+        action="store_true",
+        required=False,
+        default=False,
+        help="Enable streaming mode",
+    )
+    FLAGS = parser.parse_args()
+    asyncio.run(main(FLAGS))

samples/model_repository/vllm_opt/1/model.json

@@ -0,0 +1,5 @@
+{
+    "model":"facebook/opt-125m",
+    "disable_log_requests": "true",
+    "gpu_memory_utilization": 0.5
+}

samples/model_repository/vllm_opt/config.pbtxt

@@ -0,0 +1,75 @@
+# Copyright 2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions
+# are met:
+#  * Redistributions of source code must retain the above copyright
+#    notice, this list of conditions and the following disclaimer.
+#  * Redistributions in binary form must reproduce the above copyright
+#    notice, this list of conditions and the following disclaimer in the
+#    documentation and/or other materials provided with the distribution.
+#  * Neither the name of NVIDIA CORPORATION nor the names of its
+#    contributors may be used to endorse or promote products derived
+#    from this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
+# EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+# PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
+# CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+# EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+# PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+# PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
+# OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+name: "vllm_opt"
+backend: "vllm"
+
+# Disabling batching in Triton, let vLLM handle the batching on its own.
+max_batch_size: 0
+
+# We need to use decoupled transaction policy for saturating
+# vLLM engine for max throughtput.
+# TODO [DLIS:5233]: Allow asychronous execution to lift this
+# restriction for cases there is exactly a single response to
+# a single request.
+model_transaction_policy {
+  decoupled: True
+}
+
+input [
+  {
+    name: "PROMPT"
+    data_type: TYPE_STRING
+    dims: [ 1 ]
+  },
+  {
+    name: "STREAM"
+    data_type: TYPE_BOOL
+    dims: [ 1 ]
+  },
+  {
+    name: "SAMPLING_PARAMETERS"
+    data_type: TYPE_STRING
+    dims: [ 1 ]
+    optional: true
+  }
+]
+
+output [
+  {
+    name: "TEXT"
+    data_type: TYPE_STRING
+    dims: [ -1 ]
+  }
+]
+
+# The usage of device is deferred to the vLLM engine
+instance_group [
+  {
+    count: 1
+    kind: KIND_MODEL
+  }
+]

samples/prompts.txt

@@ -0,0 +1,4 @@
+Hello, my name is
+The most dangerous animal is
+The capital of France is
+The future of AI is