Add resolve_model_relative_to_config_file config option

README.md

@@ -130,6 +130,10 @@ Specifically,
 and
 [here](https://github.com/vllm-project/vllm/blob/ee8217e5bee5860469204ee57077a91138c9af02/vllm/engine/arg_utils.py#L201).
 
+When using local model files, specify the path to the model in the `model` field.
+By default relative paths are resolved relative to the working directory of the Triton server process.
+To specify a path relative to the `model.json` file, set the `resolve_model_relative_to_config_file` field to `true`.
+
 For multi-GPU support, EngineArgs like tensor_parallel_size can be specified in
 [model.json](samples/model_repository/vllm_model/1/model.json).
 

ci/L0_backend_vllm/vllm_backend/test.sh

@@ -49,6 +49,8 @@ function assert_curl_success {
 }
 
 rm -rf models && mkdir -p models
+
+# operational vllm model
 cp -r ${SAMPLE_MODELS_REPO}/vllm_model models/vllm_opt
 # `vllm_opt` model will be loaded on server start and stay loaded throughout
 # unittesting. To test vllm model load/unload we use a dedicated
@@ -58,10 +60,22 @@ cp -r ${SAMPLE_MODELS_REPO}/vllm_model models/vllm_opt
 sed -i 's/"gpu_memory_utilization": 0.5/"gpu_memory_utilization": 0.4/' models/vllm_opt/1/model.json
 cp -r models/vllm_opt models/vllm_load_test
 
+# python model
 mkdir -p models/add_sub/1/
 wget -P models/add_sub/1/ https://raw.githubusercontent.com/triton-inference-server/python_backend/main/examples/add_sub/model.py
 wget -P models/add_sub https://raw.githubusercontent.com/triton-inference-server/python_backend/main/examples/add_sub/config.pbtxt
 
+# local vllm model
+cp -r ${SAMPLE_MODELS_REPO}/vllm_model models/vllm_local
+sed -i 's/"facebook\/opt-125m"/"./local_model"/' models/vllm_local/1/model.json
+sed -i '/"model": /a  "resolve_model_relative_to_config_file": true,' models/vllm_local/1/model.json
+wget -P models/vllm_local/1/local_model https://huggingface.co/facebook/opt-125m/resolve/main/config.json
+wget -P models/vllm_local/1/local_model https://huggingface.co/facebook/opt-125m/resolve/main/merges.txt
+wget -P models/vllm_local/1/local_model https://huggingface.co/facebook/opt-125m/resolve/main/pytorch_model.bin
+wget -P models/vllm_local/1/local_model https://huggingface.co/facebook/opt-125m/resolve/main/special_tokens_map.json
+wget -P models/vllm_local/1/local_model https://huggingface.co/facebook/opt-125m/resolve/main/tokenizer_config.json
+wget -P models/vllm_local/1/local_model https://huggingface.co/facebook/opt-125m/resolve/main/vocab.json
+
 # Invalid model attribute
 cp -r ${SAMPLE_MODELS_REPO}/vllm_model models/vllm_invalid_1/
 sed -i 's/"disable_log_requests"/"invalid_attribute"/' models/vllm_invalid_1/1/model.json

ci/L0_backend_vllm/vllm_backend/vllm_backend_test.py

@@ -50,6 +50,7 @@ def setUp(self):
         self.python_model_name = "add_sub"
         self.ensemble_model_name = "ensemble_model"
         self.vllm_load_test = "vllm_load_test"
+        self.local_vllm_model_name = "vllm_local"
 
     def test_vllm_triton_backend(self):
         # Load both vllm and add_sub models
@@ -93,6 +94,31 @@ def test_vllm_triton_backend(self):
         )
         self.triton_client.unload_model(self.vllm_load_test)
         self.assertFalse(self.triton_client.is_model_ready(self.vllm_load_test))
+
+    def test_local_vllm_model(self):
+        # Load local vllm model
+        self.triton_client.load_model(self.local_vllm_model_name)
+        self.assertTrue(self.triton_client.is_model_ready(self.local_vllm_model_name))
+
+        # Test local vllm model
+        self._test_vllm_model(
+            prompts=PROMPTS,
+            sampling_parameters=SAMPLING_PARAMETERS,
+            stream=False,
+            send_parameters_as_tensor=True,
+            model_name=self.local_vllm_model_name,
+        )
+        self._test_vllm_model(
+            prompts=PROMPTS,
+            sampling_parameters=SAMPLING_PARAMETERS,
+            stream=False,
+            send_parameters_as_tensor=False,
+            model_name=self.local_vllm_model_name,
+        )
+
+        # Unload local vllm model
+        self.triton_client.unload_model(self.local_vllm_model_name)
+        self.assertFalse(self.triton_client.is_model_ready(self.local_vllm_model_name))
 
     def test_model_with_invalid_attributes(self):
         model_name = "vllm_invalid_1"

src/model.py

@@ -219,6 +219,20 @@ def init_engine(self):
         # Check for LoRA config and set it up if enabled
         self.setup_lora()
 
+        # Resolve the model path relative to the config file
+        if self.vllm_engine_config.pop("resolve_model_relative_to_config_file", False):
+            new_path = os.path.abspath(
+                os.path.join(
+                    pb_utils.get_model_dir(), self.vllm_engine_config["model"]
+                )
+            )
+            # Check if the resolved path is subdirectory of the model directory
+            if not new_path.startswith(pb_utils.get_model_dir()):
+                raise ValueError(
+                    f"Resolved model path '{new_path}' is not a subdirectory of the model directory '{pb_utils.get_model_dir()}'"
+                )
+            self.vllm_engine_config["model"] = new_path
+
         # Create an AsyncLLMEngine from the config from JSON
         aync_engine_args = AsyncEngineArgs(**self.vllm_engine_config)
         self.llm_engine = AsyncLLMEngine.from_engine_args(aync_engine_args)