[misc][plugin] add plugin system implementation (#7426)

tests/conftest.py

@@ -1,7 +1,9 @@
 import contextlib
 import gc
+import json
 import os
 import sys
+import tempfile
 from collections import UserList
 from enum import Enum
 from typing import (Any, Callable, Dict, List, Optional, Tuple, TypedDict,
@@ -11,6 +13,7 @@ import pytest
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
+from huggingface_hub import snapshot_download
 from PIL import Image
 from transformers import (AutoModelForCausalLM, AutoModelForSeq2SeqLM,
                           AutoModelForVision2Seq, AutoTokenizer, BatchEncoding,
@@ -757,3 +760,26 @@ def num_gpus_available():
     in current process."""
 
     return cuda_device_count_stateless()
+
+
+temp_dir = tempfile.gettempdir()
+_dummy_path = os.path.join(temp_dir, "dummy_opt")
+
+
+@pytest.fixture
+def dummy_opt_path():
+    json_path = os.path.join(_dummy_path, "config.json")
+    if not os.path.exists(_dummy_path):
+        snapshot_download(repo_id="facebook/opt-125m",
+                          local_dir=_dummy_path,
+                          ignore_patterns=[
+                              "*.bin", "*.bin.index.json", "*.pt", "*.h5",
+                              "*.msgpack"
+                          ])
+        assert os.path.exists(json_path)
+        with open(json_path, "r") as f:
+            config = json.load(f)
+        config["architectures"] = ["MyOPTForCausalLM"]
+        with open(json_path, "w") as f:
+            json.dump(config, f)
+    return _dummy_path

tests/distributed/test_distributed_oot.py

@@ -0,0 +1,6 @@
+from ..entrypoints.openai.test_oot_registration import (
+    run_and_test_dummy_opt_api_server)
+
+
+def test_distributed_oot(dummy_opt_path: str):
+    run_and_test_dummy_opt_api_server(dummy_opt_path, tp=2)

tests/entrypoints/openai/test_oot_registration.py

@@ -1,94 +1,42 @@
-import sys
-import time
-
-import torch
-from openai import OpenAI, OpenAIError
-
-from vllm import ModelRegistry
-from vllm.model_executor.models.opt import OPTForCausalLM
-from vllm.model_executor.sampling_metadata import SamplingMetadata
-from vllm.utils import get_open_port
-
 from ...utils import VLLM_PATH, RemoteOpenAIServer
 
 chatml_jinja_path = VLLM_PATH / "examples/template_chatml.jinja"
 assert chatml_jinja_path.exists()
 
 
-class MyOPTForCausalLM(OPTForCausalLM):
-
-    def compute_logits(self, hidden_states: torch.Tensor,
-                       sampling_metadata: SamplingMetadata) -> torch.Tensor:
-        # this dummy model always predicts the first token
-        logits = super().compute_logits(hidden_states, sampling_metadata)
-        logits.zero_()
-        logits[:, 0] += 1.0
-        return logits
-
-
-def server_function(port: int):
-    # register our dummy model
-    ModelRegistry.register_model("OPTForCausalLM", MyOPTForCausalLM)
-
-    sys.argv = ["placeholder.py"] + [
-        "--model",
-        "facebook/opt-125m",
+def run_and_test_dummy_opt_api_server(model, tp=1):
+    # the model is registered through the plugin
+    server_args = [
         "--gpu-memory-utilization",
         "0.10",
         "--dtype",
         "float32",
-        "--api-key",
-        "token-abc123",
-        "--port",
-        str(port),
         "--chat-template",
         str(chatml_jinja_path),
+        "--load-format",
+        "dummy",
+        "-tp",
+        f"{tp}",
     ]
-
-    import runpy
-    runpy.run_module('vllm.entrypoints.openai.api_server', run_name='__main__')
-
-
-def test_oot_registration_for_api_server():
-    port = get_open_port()
-    ctx = torch.multiprocessing.get_context()
-    server = ctx.Process(target=server_function, args=(port, ))
-    server.start()
-
-    try:
-        client = OpenAI(
-            base_url=f"http://localhost:{port}/v1",
-            api_key="token-abc123",
+    with RemoteOpenAIServer(model, server_args) as server:
+        client = server.get_client()
+        completion = client.chat.completions.create(
+            model=model,
+            messages=[{
+                "role": "system",
+                "content": "You are a helpful assistant."
+            }, {
+                "role": "user",
+                "content": "Hello!"
+            }],
+            temperature=0,
         )
-        now = time.time()
-        while True:
-            try:
-                completion = client.chat.completions.create(
-                    model="facebook/opt-125m",
-                    messages=[{
-                        "role": "system",
-                        "content": "You are a helpful assistant."
-                    }, {
-                        "role": "user",
-                        "content": "Hello!"
-                    }],
-                    temperature=0,
-                )
-                break
-            except OpenAIError as e:
-                if "Connection error" in str(e):
-                    time.sleep(3)
-                    if time.time() - now > RemoteOpenAIServer.MAX_START_WAIT_S:
-                        msg = "Server did not start in time"
-                        raise RuntimeError(msg) from e
-                else:
-                    raise e
-    finally:
-        server.terminate()
+        generated_text = completion.choices[0].message.content
+        assert generated_text is not None
+        # make sure only the first token is generated
+        rest = generated_text.replace("<s>", "")
+        assert rest == ""
 
-    generated_text = completion.choices[0].message.content
-    assert generated_text is not None
-    # make sure only the first token is generated
-    # TODO(youkaichao): Fix the test with plugin
-    rest = generated_text.replace("<s>", "")  # noqa
-    # assert rest == ""
+
+def test_oot_registration_for_api_server(dummy_opt_path: str):
+    run_and_test_dummy_opt_api_server(dummy_opt_path)

tests/models/test_oot_registration.py

@@ -1,32 +1,27 @@
-from typing import Optional
+import os
 
-import torch
+import pytest
 
-from vllm import LLM, ModelRegistry, SamplingParams
-from vllm.model_executor.models.opt import OPTForCausalLM
-from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm import LLM, SamplingParams
+
+# NOTE: the order of the tests is important
+# the first test does not load any plugins
+# the second test loads the plugin
+# they share the same process, so the plugin is loaded for the second test
 
 
-class MyOPTForCausalLM(OPTForCausalLM):
-
-    def compute_logits(
-        self,
-        hidden_states: torch.Tensor,
-        sampling_metadata: SamplingMetadata,
-    ) -> Optional[torch.Tensor]:
-        # this dummy model always predicts the first token
-        logits = super().compute_logits(hidden_states, sampling_metadata)
-        logits.zero_()
-        logits[:, 0] += 1.0
-        return logits
+def test_plugin(dummy_opt_path):
+    os.environ["VLLM_PLUGINS"] = ""
+    with pytest.raises(Exception) as excinfo:
+        LLM(model=dummy_opt_path, load_format="dummy")
+    assert "are not supported for now" in str(excinfo.value)
 
 
-def test_oot_registration():
-    # register our dummy model
-    ModelRegistry.register_model("OPTForCausalLM", MyOPTForCausalLM)
+def test_oot_registration(dummy_opt_path):
+    os.environ["VLLM_PLUGINS"] = "register_dummy_model"
     prompts = ["Hello, my name is", "The text does not matter"]
     sampling_params = SamplingParams(temperature=0)
-    llm = LLM(model="facebook/opt-125m")
+    llm = LLM(model=dummy_opt_path, load_format="dummy")
     first_token = llm.get_tokenizer().decode(0)
     outputs = llm.generate(prompts, sampling_params)
 

tests/plugins/vllm_add_dummy_model/setup.py

@@ -0,0 +1,9 @@
+from setuptools import setup
+
+setup(name='vllm_add_dummy_model',
+      version='0.1',
+      packages=['vllm_add_dummy_model'],
+      entry_points={
+          'vllm.general_plugins':
+          ["register_dummy_model = vllm_add_dummy_model:register"]
+      })

tests/plugins/vllm_add_dummy_model/vllm_add_dummy_model/__init__.py

@@ -0,0 +1,26 @@
+from typing import Optional
+
+import torch
+
+from vllm import ModelRegistry
+from vllm.model_executor.models.opt import OPTForCausalLM
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+
+
+class MyOPTForCausalLM(OPTForCausalLM):
+
+    def compute_logits(
+            self, hidden_states: torch.Tensor,
+            sampling_metadata: SamplingMetadata) -> Optional[torch.Tensor]:
+        # this dummy model always predicts the first token
+        logits = super().compute_logits(hidden_states, sampling_metadata)
+        if logits is not None:
+            logits.zero_()
+            logits[:, 0] += 1.0
+        return logits
+
+
+def register():
+    # register our dummy model
+    if "MyOPTForCausalLM" not in ModelRegistry.get_supported_archs():
+        ModelRegistry.register_model("MyOPTForCausalLM", MyOPTForCausalLM)