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

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)