[lora] Add load option to LoRA adapter API

engines/python/setup/djl_python/huggingface.py

@@ -302,6 +302,38 @@ def _streaming_inference(self, batch: List, request_input: RequestInput,
                                  self.hf_configs.device, **parameters))
         return outputs
 
+    def add_lora(self, lora_name: str, lora_alias: str, lora_path: str):
+        if not is_rolling_batch_enabled(self.hf_configs.rolling_batch):
+            raise NotImplementedError(
+                "LoRA adapter API is only supported for rolling batch.")
+
+        loaded = self.rolling_batch.add_lora(lora_name, lora_path)
+        if not loaded:
+            raise RuntimeError(f"Failed to load LoRA adapter {lora_alias}")
+        return loaded
+
+    def remove_lora(self, lora_name: str, lora_alias: str):
+        if not is_rolling_batch_enabled(self.hf_configs.rolling_batch):
+            raise NotImplementedError(
+                "LoRA adapter API is only supported for rolling batch.")
+
+        removed = self.rolling_batch.remove_lora(lora_name)
+        if not removed:
+            logging.info(
+                f"Remove LoRA adapter {lora_alias} returned false, the adapter may have already been evicted."
+            )
+        return removed
+
+    def pin_lora(self, lora_name: str, lora_alias: str):
+        if not is_rolling_batch_enabled(self.hf_configs.rolling_batch):
+            raise NotImplementedError(
+                "LoRA adapter API is only supported for rolling batch.")
+
+        pinned = self.rolling_batch.pin_lora(lora_name)
+        if not pinned:
+            raise RuntimeError(f"Failed to pin LoRA adapter {lora_alias}")
+        return pinned
+
     def get_pipeline(self, task: str, model_id_or_path: str, kwargs):
         # define tokenizer or feature extractor as kwargs to load it the pipeline correctly
         if task in {
@@ -496,41 +528,36 @@ def register_adapter(inputs: Input):
     """
     Registers lora adapter with the model.
     """
-    if not is_rolling_batch_enabled(_service.hf_configs.rolling_batch):
-        raise NotImplementedError(
-            "LoRA adapter API is only supported for rolling batch.")
-
     adapter_name = inputs.get_property("name")
     adapter_alias = inputs.get_property("alias") or adapter_name
     adapter_path = inputs.get_property("src")
+    adapter_load = inputs.get_as_string(
+        "load").lower() == "true" if inputs.contains_key("load") else True
     adapter_pin = inputs.get_as_string(
         "pin").lower() == "true" if inputs.contains_key("pin") else False
 
-    added = False
+    loaded = False
     try:
         if not os.path.exists(adapter_path):
             raise ValueError(
                 f"Only local LoRA models are supported. {adapter_path} is not a valid path"
             )
 
-        added = _service.rolling_batch.add_lora(adapter_name, adapter_path)
-        if not added:
-            raise RuntimeError(
-                f"Failed to register LoRA adapter {adapter_alias}")
+        if adapter_load:
+            loaded = _service.add_lora(adapter_name, adapter_alias,
+                                       adapter_path)
 
         if adapter_pin:
-            pinned = _service.rolling_batch.pin_lora(adapter_name)
-            if not pinned:
-                raise RuntimeError(
-                    f"Failed to pin LoRA adapter {adapter_alias}")
-
+            if not adapter_load:
+                raise RuntimeError("Need to set load=true to set pin=true")
+            _service.pin_lora(adapter_name, adapter_alias)
         _service.adapter_registry[adapter_name] = inputs
     except Exception as e:
-        if added:
+        if loaded:
             logging.info(
                 f"LoRA adapter {adapter_alias} was successfully added, but failed to pin, removing ..."
             )
-            _service.rolling_batch.remove_lora(adapter_name)
+            _service.remove_lora(adapter_name, adapter_alias)
         if any(msg in str(e)
                for msg in ("No free lora slots",
                            "greater than the number of GPU LoRA slots")):
@@ -546,13 +573,11 @@ def update_adapter(inputs: Input):
     """
     Updates lora adapter with the model.
     """
-    if not is_rolling_batch_enabled(_service.hf_configs.rolling_batch):
-        raise NotImplementedError(
-            "LoRA adapter API is only supported for rolling batch.")
-
     adapter_name = inputs.get_property("name")
     adapter_alias = inputs.get_property("alias") or adapter_name
     adapter_path = inputs.get_property("src")
+    adapter_load = inputs.get_as_string(
+        "load").lower() == "true" if inputs.contains_key("load") else True
     adapter_pin = inputs.get_as_string(
         "pin").lower() == "true" if inputs.contains_key("pin") else False
 
@@ -566,17 +591,26 @@ def update_adapter(inputs: Input):
             raise NotImplementedError(
                 f"Updating adapter path is not supported.")
 
-        old_adapter_pin = inputs.get_as_string(
-            "pin").lower() == "true" if inputs.contains_key("pin") else False
+        old_adapter_load = old_adapter.get_as_string("load").lower(
+        ) == "true" if old_adapter.contains_key("load") else True
+        if old_adapter_load != adapter_load:
+            if adapter_load:
+                _service.add_lora(adapter_name, adapter_alias, adapter_path)
+            else:
+                if adapter_pin:
+                    raise RuntimeError("Need to set pin=false to set load=false")
+                _service.remove_lora(adapter_name, adapter_alias)
+
+        old_adapter_pin = old_adapter.get_as_string("pin").lower(
+        ) == "true" if old_adapter.contains_key("pin") else False
         if old_adapter_pin and not adapter_pin:
             raise NotImplementedError(f"Unpin adapter is not supported.")
 
-        if adapter_pin:
-            pinned = _service.rolling_batch.pin_lora(adapter_name)
-            if not pinned:
-                raise RuntimeError(
-                    f"Failed to pin LoRA adapter {adapter_alias}")
-
+        if old_adapter_pin != adapter_pin:
+            if adapter_pin:
+                if not adapter_load:
+                    raise RuntimeError("Need to set load=true to set pin=true")
+                _service.pin_lora(adapter_name, adapter_alias)
         _service.adapter_registry[adapter_name] = inputs
     except Exception as e:
         if any(msg in str(e)
@@ -593,23 +627,14 @@ def unregister_adapter(inputs: Input):
     """
     Unregisters lora adapter from the model.
     """
-    if not is_rolling_batch_enabled(_service.hf_configs.rolling_batch):
-        raise NotImplementedError(
-            "LoRA adapter API is only supported for rolling batch.")
-
     adapter_name = inputs.get_property("name")
     adapter_alias = inputs.get_property("alias") or adapter_name
 
     if adapter_name not in _service.adapter_registry:
         raise ValueError(f"Adapter {adapter_alias} not registered.")
 
     try:
-        removed = _service.rolling_batch.remove_lora(adapter_name)
-        if not removed:
-            logging.info(
-                f"Remove LoRA adapter {adapter_alias} returned false, the adapter may have already been evicted."
-            )
-
+        _service.remove_lora(adapter_name, adapter_alias)
         del _service.adapter_registry[adapter_name]
     except Exception as e:
         return Output().error("remove_adapter_error", message=str(e))

engines/python/setup/djl_python/rolling_batch/lmi_dist_rolling_batch.py

@@ -283,5 +283,5 @@ def pin_lora(self, lora_name):
         # 2) An adapter is not evicted, call add_lora() is not necessary.
         # But since whether an adapter is evicted is not exposed outside of engine,
         # and add_lora() in this case will take negligible time, we will still call add_lora().
-        self.engine.add_lora(lora_request)
-        return self.engine.pin_lora(lora_request.lora_int_id)
+        return self.engine.add_lora(lora_request) and self.engine.pin_lora(
+            lora_request.lora_int_id)

engines/python/setup/djl_python/rolling_batch/vllm_rolling_batch.py

@@ -198,5 +198,5 @@ def pin_lora(self, lora_name):
         # 2) An adapter is not evicted, call add_lora() is not necessary.
         # But since whether an adapter is evicted is not exposed outside of engine,
         # and add_lora() in this case will take negligible time, we will still call add_lora().
-        self.engine.add_lora(lora_request)
-        return self.engine.pin_lora(lora_request.lora_int_id)
+        return self.engine.add_lora(lora_request) and self.engine.pin_lora(
+            lora_request.lora_int_id)

serving/src/main/java/ai/djl/serving/http/AdapterManagementRequestHandler.java

@@ -168,7 +168,7 @@ private void handleListAdapters(
         for (int i = pagination.getPageToken(); i < pagination.getLast(); ++i) {
             String adapterName = keys.get(i);
             Adapter<Input, Output> adapter = modelInfo.getAdapter(adapterName);
-            list.addAdapter(adapter.getName(), adapter.getSrc(), adapter.isPin());
+            list.addAdapter(adapter.getName(), adapter.getSrc(), adapter.isLoad(), adapter.isPin());
         }
 
         NettyUtils.sendJsonResponse(ctx, list);

serving/src/main/java/ai/djl/serving/http/DescribeAdapterResponse.java

@@ -20,6 +20,7 @@
 public class DescribeAdapterResponse {
     private String name;
     private String src;
+    private boolean load;
     private boolean pin;
 
     /**
@@ -30,6 +31,7 @@ public class DescribeAdapterResponse {
     public DescribeAdapterResponse(Adapter<Input, Output> adapter) {
         this.name = adapter.getName();
         this.src = adapter.getSrc();
+        this.load = adapter.isLoad();
         this.pin = adapter.isPin();
     }
 
@@ -51,6 +53,15 @@ public String getSrc() {
         return src;
     }
 
+    /**
+     * Returns whether to load the adapter weights.
+     *
+     * @return whether to load the adapter weights
+     */
+    public boolean isLoad() {
+        return load;
+    }
+
     /**
      * Returns whether to pin the adapter.
      *

serving/src/main/java/ai/djl/serving/http/list/ListAdaptersResponse.java

@@ -58,21 +58,24 @@ public List<AdapterItem> getAdapters() {
      *
      * @param name the adapter name
      * @param src the adapter source
+     * @param load whether to load the adapter weights
      * @param pin whether to pin the adapter
      */
-    public void addAdapter(String name, String src, boolean pin) {
-        adapters.add(new AdapterItem(name, src, pin));
+    public void addAdapter(String name, String src, boolean load, boolean pin) {
+        adapters.add(new AdapterItem(name, src, load, pin));
     }
 
     /** A class that holds the adapter response. */
     public static final class AdapterItem {
         private String name;
         private String src;
+        private boolean load;
         private boolean pin;
 
-        private AdapterItem(String name, String src, boolean pin) {
+        private AdapterItem(String name, String src, boolean load, boolean pin) {
             this.name = name;
             this.src = src;
+            this.load = load;
             this.pin = pin;
         }
 

serving/src/test/java/ai/djl/serving/ModelServerTest.java

@@ -1002,12 +1002,25 @@ private void testRegisterAdapter(Channel channel, boolean registerModel, boolean
         testAdapterMissing();
 
         String strModelPrefix = modelPrefix ? "/models/adaptecho" : "";
-        url = strModelPrefix + "/adapters?name=adaptable&src=src&echooption=opt";
+        String adapterName = "adaptable";
+        url = strModelPrefix + "/adapters?name=" + adapterName + "&src=src&echooption=opt";
         request(channel, HttpMethod.POST, url);
         assertHttpOk();
 
         StatusResponse statusResp = JsonUtils.GSON.fromJson(result, StatusResponse.class);
-        assertEquals(statusResp.getStatus(), "Adapter adaptable registered");
+        assertEquals(statusResp.getStatus(), "Adapter " + adapterName + " registered");
+
+        // Assert adapter registered
+        url = strModelPrefix + "/adapters/" + adapterName;
+        request(channel, HttpMethod.GET, url);
+        assertHttpOk();
+
+        DescribeAdapterResponse resp =
+                JsonUtils.GSON.fromJson(result, DescribeAdapterResponse.class);
+        assertEquals(resp.getName(), adapterName);
+        assertEquals(resp.getSrc(), "src");
+        assertTrue(resp.isLoad());
+        assertFalse(resp.isPin());
     }
 
     private void testRegisterAdapterConflict() throws InterruptedException {
@@ -1070,17 +1083,27 @@ private void testRegisterAdapterHandlerError() throws InterruptedException {
     private void testUpdateAdapter(Channel channel, boolean modelPrefix)
             throws InterruptedException {
         logTestFunction();
+
+        String adapterName = "adaptable";
         String strModelPrefix = modelPrefix ? "/models/adaptecho" : "";
-        String url = strModelPrefix + "/adapters/adaptable/update?src=src1";
+        String url = strModelPrefix + "/adapters/" + adapterName + "/update?pin=true";
         request(channel, HttpMethod.POST, url);
         assertHttpOk();
 
-        url = strModelPrefix + "/adapters/adaptable/update?src=src";
-        request(channel, HttpMethod.POST, url);
+        StatusResponse statusResp = JsonUtils.GSON.fromJson(result, StatusResponse.class);
+        assertEquals(statusResp.getStatus(), "Adapter " + adapterName + " updated");
+
+        // Assert adapter updated
+        url = strModelPrefix + "/adapters/" + adapterName;
+        request(channel, HttpMethod.GET, url);
         assertHttpOk();
 
-        StatusResponse statusResp = JsonUtils.GSON.fromJson(result, StatusResponse.class);
-        assertEquals(statusResp.getStatus(), "Adapter adaptable updated");
+        DescribeAdapterResponse resp =
+                JsonUtils.GSON.fromJson(result, DescribeAdapterResponse.class);
+        assertEquals(resp.getName(), adapterName);
+        assertEquals(resp.getSrc(), "src");
+        assertTrue(resp.isLoad());
+        assertTrue(resp.isPin());
     }
 
     private void testUpdateAdapterModelNotFound() throws InterruptedException {
@@ -1117,7 +1140,11 @@ private void testUpdateAdapterHandlerError() throws InterruptedException {
         String modelName = "adaptecho";
         String adapterName = "adaptable";
         String strModelPrefix = "/models/" + modelName;
-        String url = strModelPrefix + "/adapters/" + adapterName + "/update?src=src1&error=true";
+        String url =
+                strModelPrefix
+                        + "/adapters/"
+                        + adapterName
+                        + "/update?src=src1&load=false&error=true";
         request(channel, HttpMethod.POST, url);
         channel.closeFuture().sync();
         channel.close().sync();
@@ -1135,6 +1162,8 @@ private void testUpdateAdapterHandlerError() throws InterruptedException {
                 JsonUtils.GSON.fromJson(result, DescribeAdapterResponse.class);
         assertEquals(resp.getName(), adapterName);
         assertEquals(resp.getSrc(), "src");
+        assertTrue(resp.isLoad());
+        assertFalse(resp.isPin());
     }
 
     private void testAdapterMissing() throws InterruptedException {
@@ -1290,6 +1319,8 @@ private void testDescribeAdapter(Channel channel, boolean modelPrefix)
                 JsonUtils.GSON.fromJson(result, DescribeAdapterResponse.class);
         assertEquals(resp.getName(), "adaptable");
         assertEquals(resp.getSrc(), "src");
+        assertTrue(resp.isLoad());
+        assertTrue(resp.isPin());
     }
 
     private void testDescribeAdapterModelNotFound() throws InterruptedException {
@@ -1337,12 +1368,24 @@ private void testUnregisterAdapter(Channel channel, boolean modelPrefix)
             throws InterruptedException {
         logTestFunction();
         String strModelPrefix = modelPrefix ? "/models/adaptecho" : "";
-        String url = strModelPrefix + "/adapters/adaptable";
+        String adapterName = "adaptable";
+        String url = strModelPrefix + "/adapters/" + adapterName;
         request(channel, HttpMethod.DELETE, url);
         assertHttpOk();
 
         StatusResponse statusResp = JsonUtils.GSON.fromJson(result, StatusResponse.class);
-        assertEquals(statusResp.getStatus(), "Adapter adaptable unregistered");
+        assertEquals(statusResp.getStatus(), "Adapter " + adapterName + " unregistered");
+
+        // Assert adapter unregistered
+        channel = connect(Connector.ConnectorType.MANAGEMENT);
+        assertNotNull(channel);
+
+        url = strModelPrefix + "/adapters";
+        request(channel, HttpMethod.GET, url);
+        assertHttpOk();
+
+        ListAdaptersResponse resp = JsonUtils.GSON.fromJson(result, ListAdaptersResponse.class);
+        assertFalse(resp.getAdapters().stream().anyMatch(a -> adapterName.equals(a.getName())));
     }
 
     private void testUnregisterAdapterModelNotFound() throws InterruptedException {

wlm/src/main/java/ai/djl/serving/wlm/Adapter.java

@@ -192,6 +192,15 @@ public void setOptions(Map<String, String> options) {
         this.options = options;
     }
 
+    /**
+     * Returns whether to load the adapter weights.
+     *
+     * @return whether to load the adapter weights
+     */
+    public boolean isLoad() {
+        return Boolean.parseBoolean(options.getOrDefault("load", "true"));
+    }
+
     /**
      * Returns whether to pin the adapter.
      *