diff --git a/HeterogeneousCore/SonicCore/BuildFile.xml b/HeterogeneousCore/SonicCore/BuildFile.xml
index b0d5e2a08b98f..5208c91638f37 100644
--- a/HeterogeneousCore/SonicCore/BuildFile.xml
+++ b/HeterogeneousCore/SonicCore/BuildFile.xml
@@ -2,6 +2,7 @@
 <use name="FWCore/Concurrency"/>
 <use name="FWCore/MessageLogger"/>
 <use name="FWCore/ParameterSet"/>
+<use name="FWCore/PluginManager"/>
 <use name="FWCore/Utilities"/>
 <export>
   <lib name="1"/>
diff --git a/HeterogeneousCore/SonicCore/interface/RetryActionBase.h b/HeterogeneousCore/SonicCore/interface/RetryActionBase.h
new file mode 100644
index 0000000000000..e3fc0bbb8af9a
--- /dev/null
+++ b/HeterogeneousCore/SonicCore/interface/RetryActionBase.h
@@ -0,0 +1,35 @@
+#ifndef HeterogeneousCore_SonicCore_RetryActionBase
+#define HeterogeneousCore_SonicCore_RetryActionBase
+
+#include "FWCore/PluginManager/interface/PluginFactory.h"
+#include "FWCore/ParameterSet/interface/ParameterSet.h"
+#include "HeterogeneousCore/SonicCore/interface/SonicClientBase.h"
+#include <memory>
+#include <string>
+
+// Base class for retry actions
+class RetryActionBase {
+public:
+  RetryActionBase(const edm::ParameterSet& conf, SonicClientBase* client);
+  virtual ~RetryActionBase() = default;
+
+  bool shouldRetry() const { return shouldRetry_; }  // Getter for shouldRetry_
+
+  virtual void retry() = 0;  // Pure virtual function for execution logic
+  virtual void start() = 0;  // Pure virtual function for execution logic for initialization
+
+protected:
+  void eval();  // interface for calling evaluate in client
+
+protected:
+  SonicClientBase* client_;
+  bool shouldRetry_;  // Flag to track if further retries should happen
+};
+
+// Define the factory for creating retry actions
+using RetryActionFactory =
+    edmplugin::PluginFactory<RetryActionBase*(const edm::ParameterSet&, SonicClientBase* client)>;
+
+#endif
+
+#define DEFINE_RETRY_ACTION(type) DEFINE_EDM_PLUGIN(RetryActionFactory, type, #type);
diff --git a/HeterogeneousCore/SonicCore/interface/SonicClientBase.h b/HeterogeneousCore/SonicCore/interface/SonicClientBase.h
index 47caaae8b2052..45a089701ed12 100644
--- a/HeterogeneousCore/SonicCore/interface/SonicClientBase.h
+++ b/HeterogeneousCore/SonicCore/interface/SonicClientBase.h
@@ -9,12 +9,15 @@
 #include "HeterogeneousCore/SonicCore/interface/SonicDispatcherPseudoAsync.h"
 
 #include <string>
+#include <vector>
 #include <exception>
 #include <memory>
 #include <optional>
 
 enum class SonicMode { Sync = 1, Async = 2, PseudoAsync = 3 };
 
+class RetryActionBase;
+
 class SonicClientBase {
 public:
   //constructor
@@ -54,14 +57,23 @@ class SonicClientBase {
   SonicMode mode_;
   bool verbose_;
   std::unique_ptr<SonicDispatcher> dispatcher_;
-  unsigned allowedTries_, tries_;
+  unsigned totalTries_;
   std::optional<edm::WaitingTaskWithArenaHolder> holder_;
 
+  // Use a unique_ptr with a custom deleter to avoid incomplete type issues
+  struct RetryDeleter {
+    void operator()(RetryActionBase* ptr) const;
+  };
+
+  using RetryActionPtr = std::unique_ptr<RetryActionBase, RetryDeleter>;
+  std::vector<RetryActionPtr> retryActions_;
+
   //for logging/debugging
   std::string debugName_, clientName_, fullDebugName_;
 
   friend class SonicDispatcher;
   friend class SonicDispatcherPseudoAsync;
+  friend class RetryActionBase;
 };
 
 #endif
diff --git a/HeterogeneousCore/SonicCore/plugins/BuildFile.xml b/HeterogeneousCore/SonicCore/plugins/BuildFile.xml
new file mode 100644
index 0000000000000..0ecf2187a0f82
--- /dev/null
+++ b/HeterogeneousCore/SonicCore/plugins/BuildFile.xml
@@ -0,0 +1,6 @@
+<use name="FWCore/Framework"/>
+<use name="FWCore/PluginManager"/>
+<use name="FWCore/ParameterSet"/>
+<use name="HeterogeneousCore/SonicCore"/>
+<library file="*.cc" name="pluginHeterogeneousCoreSonicCore"/>
+
diff --git a/HeterogeneousCore/SonicCore/plugins/RetrySameServerAction.cc b/HeterogeneousCore/SonicCore/plugins/RetrySameServerAction.cc
new file mode 100644
index 0000000000000..9877013b93d5b
--- /dev/null
+++ b/HeterogeneousCore/SonicCore/plugins/RetrySameServerAction.cc
@@ -0,0 +1,30 @@
+#include "HeterogeneousCore/SonicCore/interface/RetryActionBase.h"
+#include "HeterogeneousCore/SonicCore/interface/SonicClientBase.h"
+
+class RetrySameServerAction : public RetryActionBase {
+public:
+  RetrySameServerAction(const edm::ParameterSet& pset, SonicClientBase* client)
+      : RetryActionBase(pset, client), allowedTries_(pset.getUntrackedParameter<unsigned>("allowedTries", 0)) {}
+
+  void start() override { tries_ = 0; };
+
+protected:
+  void retry() override;
+
+private:
+  unsigned allowedTries_, tries_;
+};
+
+void RetrySameServerAction::retry() {
+  ++tries_;
+  //if max retries has not been exceeded, call evaluate again
+  if (tries_ < allowedTries_) {
+    eval();
+    return;
+  } else {
+    shouldRetry_ = false;  // Flip flag when max retries are reached
+    edm::LogInfo("RetrySameServerAction") << "Max retry attempts reached. No further retries.";
+  }
+}
+
+DEFINE_RETRY_ACTION(RetrySameServerAction)
diff --git a/HeterogeneousCore/SonicCore/src/RetryActionBase.cc b/HeterogeneousCore/SonicCore/src/RetryActionBase.cc
new file mode 100644
index 0000000000000..41b9a6186da2b
--- /dev/null
+++ b/HeterogeneousCore/SonicCore/src/RetryActionBase.cc
@@ -0,0 +1,15 @@
+#include "HeterogeneousCore/SonicCore/interface/RetryActionBase.h"
+
+// Constructor implementation
+RetryActionBase::RetryActionBase(const edm::ParameterSet& conf, SonicClientBase* client)
+    : client_(client), shouldRetry_(true) {}
+
+void RetryActionBase::eval() {
+  if (client_) {
+    client_->evaluate();
+  } else {
+    edm::LogError("RetryActionBase") << "Client pointer is null, cannot evaluate.";
+  }
+}
+
+EDM_REGISTER_PLUGINFACTORY(RetryActionFactory, "RetryActionFactory");
diff --git a/HeterogeneousCore/SonicCore/src/SonicClientBase.cc b/HeterogeneousCore/SonicCore/src/SonicClientBase.cc
index 745c51f17aaf3..9949d9d1f2ea2 100644
--- a/HeterogeneousCore/SonicCore/src/SonicClientBase.cc
+++ b/HeterogeneousCore/SonicCore/src/SonicClientBase.cc
@@ -1,18 +1,33 @@
 #include "HeterogeneousCore/SonicCore/interface/SonicClientBase.h"
+#include "HeterogeneousCore/SonicCore/interface/RetryActionBase.h"
 #include "FWCore/Utilities/interface/Exception.h"
 #include "FWCore/ParameterSet/interface/allowedValues.h"
 
+// Custom deleter implementation
+void SonicClientBase::RetryDeleter::operator()(RetryActionBase* ptr) const { delete ptr; }
+
 SonicClientBase::SonicClientBase(const edm::ParameterSet& params,
                                  const std::string& debugName,
                                  const std::string& clientName)
-    : allowedTries_(params.getUntrackedParameter<unsigned>("allowedTries", 0)),
-      debugName_(debugName),
-      clientName_(clientName),
-      fullDebugName_(debugName_) {
+    : debugName_(debugName), clientName_(clientName), fullDebugName_(debugName_) {
   if (!clientName_.empty())
     fullDebugName_ += ":" + clientName_;
 
+  const auto& retryPSetList = params.getParameter<std::vector<edm::ParameterSet>>("Retry");
   std::string modeName(params.getParameter<std::string>("mode"));
+
+  for (const auto& retryPSet : retryPSetList) {
+    const std::string& actionType = retryPSet.getParameter<std::string>("retryType");
+
+    auto retryAction = RetryActionFactory::get()->create(actionType, retryPSet, this);
+    if (retryAction) {
+      //Convert to  RetryActionPtr Type from raw pointer of retryAction
+      retryActions_.emplace_back(RetryActionPtr(retryAction.release()));
+    } else {
+      throw cms::Exception("Configuration") << "Unknown Retry type " << actionType << " for SonicClient: " << modeName;
+    }
+  }
+
   if (modeName == "Sync")
     setMode(SonicMode::Sync);
   else if (modeName == "Async")
@@ -40,24 +55,30 @@ void SonicClientBase::start(edm::WaitingTaskWithArenaHolder holder) {
   holder_ = std::move(holder);
 }
 
-void SonicClientBase::start() { tries_ = 0; }
+void SonicClientBase::start() {
+  totalTries_ = 0;
+  // initialize all actions
+  for (const auto& action : retryActions_) {
+    action->start();
+  }
+}
 
 void SonicClientBase::finish(bool success, std::exception_ptr eptr) {
   //retries are only allowed if no exception was raised
   if (!success and !eptr) {
-    ++tries_;
-    //if max retries has not been exceeded, call evaluate again
-    if (tries_ < allowedTries_) {
-      evaluate();
-      //avoid calling doneWaiting() twice
-      return;
-    }
-    //prepare an exception if exceeded
-    else {
-      edm::Exception ex(edm::errors::ExternalFailure);
-      ex << "SonicCallFailed: call failed after max " << tries_ << " tries";
-      eptr = make_exception_ptr(ex);
+    ++totalTries_;
+    for (const auto& action : retryActions_) {
+      if (action->shouldRetry()) {
+        action->retry();  // Call retry only if shouldRetry_ is true
+        return;
+      }
     }
+    //prepare an exception if no more retries left
+    edm::LogInfo("SonicClientBase") << "SonicCallFailed: call failed, no retry actions available after " << totalTries_
+                                    << " tries.";
+    edm::Exception ex(edm::errors::ExternalFailure);
+    ex << "SonicCallFailed: call failed, no retry actions available after  " << totalTries_ << " tries.";
+    eptr = make_exception_ptr(ex);
   }
   if (holder_) {
     holder_->doneWaiting(eptr);
@@ -74,7 +95,20 @@ void SonicClientBase::fillBasePSetDescription(edm::ParameterSetDescription& desc
   //restrict allowed values
   desc.ifValue(edm::ParameterDescription<std::string>("mode", "PseudoAsync", true),
                edm::allowedValues<std::string>("Sync", "Async", "PseudoAsync"));
-  if (allowRetry)
-    desc.addUntracked<unsigned>("allowedTries", 0);
+  if (allowRetry) {
+    // Defines the structure of each entry in the VPSet
+    edm::ParameterSetDescription retryDesc;
+    retryDesc.add<std::string>("retryType", "RetrySameServerAction");
+    retryDesc.addUntracked<unsigned>("allowedTries", 0);
+
+    // Define a default retry action
+    edm::ParameterSet defaultRetry;
+    defaultRetry.addParameter<std::string>("retryType", "RetrySameServerAction");
+    defaultRetry.addUntrackedParameter<unsigned>("allowedTries", 0);
+
+    // Add the VPSet with the default retry action
+    desc.addVPSet("Retry", retryDesc, {defaultRetry});
+  }
+  desc.add("sonicClientBase", desc);
   desc.addUntracked<bool>("verbose", false);
 }
diff --git a/HeterogeneousCore/SonicCore/test/DummyClient.h b/HeterogeneousCore/SonicCore/test/DummyClient.h
index ccef888ad9f7d..6504843926c0a 100644
--- a/HeterogeneousCore/SonicCore/test/DummyClient.h
+++ b/HeterogeneousCore/SonicCore/test/DummyClient.h
@@ -36,7 +36,7 @@ class DummyClient : public SonicClient<int> {
     this->output_ = this->input_ * factor_;
 
     //simulate a failure
-    if (this->tries_ < fails_)
+    if (this->totalTries_ < fails_)
       this->finish(false);
     else
       this->finish(true);
diff --git a/HeterogeneousCore/SonicCore/test/sonicTest_cfg.py b/HeterogeneousCore/SonicCore/test/sonicTest_cfg.py
index 2cc429138b85c..bf7b44cb01519 100644
--- a/HeterogeneousCore/SonicCore/test/sonicTest_cfg.py
+++ b/HeterogeneousCore/SonicCore/test/sonicTest_cfg.py
@@ -19,15 +19,19 @@
 
 process.options.numberOfThreads = 2
 process.options.numberOfStreams = 0
-
 process.dummySync = _moduleClass(_moduleName,
     input = cms.int32(1),
     Client = cms.PSet(
         mode = cms.string("Sync"),
         factor = cms.int32(-1),
         wait = cms.int32(10),
-        allowedTries = cms.untracked.uint32(0),
         fails = cms.uint32(0),
+        Retry = cms.VPSet(
+          cms.PSet(
+            retryType = cms.string('RetrySameServerAction'),
+            allowedTries = cms.untracked.uint32(0)
+          )
+        )
     ),
 )
 
@@ -37,8 +41,14 @@
         mode = cms.string("PseudoAsync"),
         factor = cms.int32(2),
         wait = cms.int32(10),
-        allowedTries = cms.untracked.uint32(0),
         fails = cms.uint32(0),
+        Retry = cms.VPSet(
+          cms.PSet(
+            retryType = cms.string('RetrySameServerAction'),
+            allowedTries = cms.untracked.uint32(0)
+          )
+        )
+
     ),
 )
 
@@ -48,32 +58,53 @@
         mode = cms.string("Async"),
         factor = cms.int32(5),
         wait = cms.int32(10),
-        allowedTries = cms.untracked.uint32(0),
         fails = cms.uint32(0),
+        Retry = cms.VPSet(
+          cms.PSet(
+            retryType = cms.string('RetrySameServerAction'),
+            allowedTries = cms.untracked.uint32(0)
+          )
+        )
     ),
 )
 
 process.dummySyncRetry = process.dummySync.clone(
     Client = dict(
         wait = 2,
-        allowedTries = 2,
         fails = 1,
+        Retry = cms.VPSet(
+          cms.PSet(
+            retryType = cms.string('RetrySameServerAction'),
+            allowedTries = cms.untracked.uint32(2)
+          )
+        )
+
     )
 )
 
 process.dummyPseudoAsyncRetry = process.dummyPseudoAsync.clone(
     Client = dict(
         wait = 2,
-        allowedTries = 2,
         fails = 1,
+        Retry = cms.VPSet(
+          cms.PSet(
+            retryType = cms.string('RetrySameServerAction'),
+            allowedTries = cms.untracked.uint32(2)
+          )
+        )
     )
 )
 
 process.dummyAsyncRetry = process.dummyAsync.clone(
     Client = dict(
         wait = 2,
-        allowedTries = 2,
         fails = 1,
+        Retry = cms.VPSet(
+          cms.PSet(
+            allowedTries = cms.untracked.uint32(2),
+            retryType = cms.string('RetrySameServerAction')
+          )
+        )
     )
 )
 
diff --git a/HeterogeneousCore/SonicTriton/BuildFile.xml b/HeterogeneousCore/SonicTriton/BuildFile.xml
index b93d51e711e87..4af38d69d89e9 100644
--- a/HeterogeneousCore/SonicTriton/BuildFile.xml
+++ b/HeterogeneousCore/SonicTriton/BuildFile.xml
@@ -10,6 +10,7 @@
 <iftool name="cuda">
   <use name="cuda"/>
 </iftool>
+
 <export>
-  <lib   name="1"/>
+  <lib name="1"/>
 </export>
diff --git a/HeterogeneousCore/SonicTriton/interface/RetryActionDiffServer.h b/HeterogeneousCore/SonicTriton/interface/RetryActionDiffServer.h
new file mode 100644
index 0000000000000..af7720b90cb0b
--- /dev/null
+++ b/HeterogeneousCore/SonicTriton/interface/RetryActionDiffServer.h
@@ -0,0 +1,32 @@
+#ifndef HeterogeneousCore_SonicTriton_RetryActionDiffServer_h
+#define HeterogeneousCore_SonicTriton_RetryActionDiffServer_h
+
+#include "HeterogeneousCore/SonicCore/interface/RetryActionBase.h"
+
+/**
+ * @class RetryActionDiffServer
+ * @brief A concrete implementation of RetryActionBase that attempts to retry an inference
+ * request on a different Triton server.
+ *
+ * This class provides a fallback mechanism. If an initial inference request fails
+ * (e.g., due to server unavailability or a model-specific error), this action will be
+ * triggered. It queries the central TritonService to select an alternative server (e.g.,
+ * the fallback server when available) and instructs the TritonClient to reconnect to
+ * that server for the retry attempt. This action is designed for one-time use per
+ * inference call; after the retry attempt, it disables itself until the next `start()`
+ * call.
+ */
+ 
+class RetryActionDiffServer : public RetryActionBase {
+public:
+  RetryActionDiffServer(const edm::ParameterSet& conf, SonicClientBase* client);
+  ~RetryActionDiffServer() override = default;
+
+  void retry() override;
+  void start() override;
+
+private:
+}; 
+
+#endif
+
diff --git a/HeterogeneousCore/SonicTriton/interface/TritonClient.h b/HeterogeneousCore/SonicTriton/interface/TritonClient.h
index df8f9b559427c..9e21b646508e9 100644
--- a/HeterogeneousCore/SonicTriton/interface/TritonClient.h
+++ b/HeterogeneousCore/SonicTriton/interface/TritonClient.h
@@ -1,102 +1,107 @@
-#ifndef HeterogeneousCore_SonicTriton_TritonClient
-#define HeterogeneousCore_SonicTriton_TritonClient
-
-#include "FWCore/ParameterSet/interface/ParameterSet.h"
-#include "FWCore/ParameterSet/interface/ParameterSetDescription.h"
-#include "FWCore/ServiceRegistry/interface/ServiceToken.h"
-#include "HeterogeneousCore/SonicCore/interface/SonicClient.h"
-#include "HeterogeneousCore/SonicTriton/interface/TritonData.h"
-#include "HeterogeneousCore/SonicTriton/interface/TritonService.h"
-
-#include <map>
-#include <vector>
-#include <string>
-#include <exception>
-#include <unordered_map>
-
-#include "grpc_client.h"
-#include "grpc_service.pb.h"
-
-enum class TritonBatchMode { Rectangular = 1, Ragged = 2 };
-
-class TritonClient : public SonicClient<TritonInputMap, TritonOutputMap> {
-public:
-  struct ServerSideStats {
-    uint64_t inference_count_;
-    uint64_t execution_count_;
-    uint64_t success_count_;
-    uint64_t cumm_time_ns_;
-    uint64_t queue_time_ns_;
-    uint64_t compute_input_time_ns_;
-    uint64_t compute_infer_time_ns_;
-    uint64_t compute_output_time_ns_;
-  };
-
-  //constructor
-  TritonClient(const edm::ParameterSet& params, const std::string& debugName);
-
-  //destructor
-  ~TritonClient() override;
-
-  //accessors
-  unsigned batchSize() const;
-  TritonBatchMode batchMode() const { return batchMode_; }
-  bool verbose() const { return verbose_; }
-  bool useSharedMemory() const { return useSharedMemory_; }
-  void setUseSharedMemory(bool useShm) { useSharedMemory_ = useShm; }
-  bool setBatchSize(unsigned bsize);
-  void setBatchMode(TritonBatchMode batchMode);
-  void resetBatchMode();
-  void reset() override;
-  TritonServerType serverType() const { return serverType_; }
-  bool isLocal() const { return isLocal_; }
-
-  //for fillDescriptions
-  static void fillPSetDescription(edm::ParameterSetDescription& iDesc);
-
-protected:
-  //helpers
-  bool noOuterDim() const { return noOuterDim_; }
-  unsigned outerDim() const { return outerDim_; }
-  unsigned nEntries() const;
-  void getResults(const std::vector<std::shared_ptr<triton::client::InferResult>>& results);
-  void evaluate() override;
-  template <typename F>
-  bool handle_exception(F&& call);
-
-  void reportServerSideStats(const ServerSideStats& stats) const;
-  ServerSideStats summarizeServerStats(const inference::ModelStatistics& start_status,
-                                       const inference::ModelStatistics& end_status) const;
-
-  inference::ModelStatistics getServerSideStatus() const;
-
-  //members
-  unsigned maxOuterDim_;
-  unsigned outerDim_;
-  bool noOuterDim_;
-  unsigned nEntries_;
-  TritonBatchMode batchMode_;
-  bool manualBatchMode_;
-  bool verbose_;
-  bool useSharedMemory_;
-  TritonServerType serverType_;
-  bool isLocal_;
-  grpc_compression_algorithm compressionAlgo_;
-  triton::client::Headers headers_;
-
-  std::unique_ptr<triton::client::InferenceServerGrpcClient> client_;
-  //stores timeout, model name and version
-  std::vector<triton::client::InferOptions> options_;
-  edm::ServiceToken token_;
-
-private:
-  friend TritonInputData;
-  friend TritonOutputData;
-
-  //private accessors only used by data
-  auto client() { return client_.get(); }
-  void addEntry(unsigned entry);
-  void resizeEntries(unsigned entry);
-};
-
-#endif
+#ifndef HeterogeneousCore_SonicTriton_TritonClient
+#define HeterogeneousCore_SonicTriton_TritonClient
+
+#include "FWCore/ParameterSet/interface/ParameterSet.h"
+#include "FWCore/ParameterSet/interface/ParameterSetDescription.h"
+#include "FWCore/ServiceRegistry/interface/ServiceToken.h"
+#include "HeterogeneousCore/SonicCore/interface/SonicClient.h"
+#include "HeterogeneousCore/SonicTriton/interface/TritonData.h"
+#include "HeterogeneousCore/SonicTriton/interface/TritonService.h"
+
+#include <map>
+#include <vector>
+#include <string>
+#include <exception>
+#include <unordered_map>
+
+#include "grpc_client.h"
+#include "grpc_service.pb.h"
+
+enum class TritonBatchMode { Rectangular = 1, Ragged = 2 };
+
+class TritonClient : public SonicClient<TritonInputMap, TritonOutputMap> {
+public:
+  struct ServerSideStats {
+    uint64_t inference_count_;
+    uint64_t execution_count_;
+    uint64_t success_count_;
+    uint64_t cumm_time_ns_;
+    uint64_t queue_time_ns_;
+    uint64_t compute_input_time_ns_;
+    uint64_t compute_infer_time_ns_;
+    uint64_t compute_output_time_ns_;
+  };
+
+  //constructor
+  TritonClient(const edm::ParameterSet& params, const std::string& debugName);
+
+  //destructor
+  ~TritonClient() override;
+
+  //accessors
+  unsigned batchSize() const;
+  TritonBatchMode batchMode() const { return batchMode_; }
+  bool verbose() const { return verbose_; }
+  bool useSharedMemory() const { return useSharedMemory_; }
+  void setUseSharedMemory(bool useShm) { useSharedMemory_ = useShm; }
+  bool setBatchSize(unsigned bsize);
+  void setBatchMode(TritonBatchMode batchMode);
+  void resetBatchMode();
+  void reset() override;
+  TritonServerType serverType() const { return serverType_; }
+  bool isLocal() const { return isLocal_; }
+  virtual void connectToServer(const std::string& url);
+  virtual void updateServer(std::string serverName);
+
+  //for fillDescriptions
+  static void fillPSetDescription(edm::ParameterSetDescription& iDesc);
+
+protected:
+  // Protected default constructor for unit testing (no framework services)
+  TritonClient();
+
+  //helpers
+  bool noOuterDim() const { return noOuterDim_; }
+  unsigned outerDim() const { return outerDim_; }
+  unsigned nEntries() const;
+  void getResults(const std::vector<std::shared_ptr<triton::client::InferResult>>& results);
+  virtual void evaluate() override;
+  template <typename F>
+  bool handle_exception(F&& call);
+
+  void reportServerSideStats(const ServerSideStats& stats) const;
+  ServerSideStats summarizeServerStats(const inference::ModelStatistics& start_status,
+                                       const inference::ModelStatistics& end_status) const;
+
+  inference::ModelStatistics getServerSideStatus() const;
+
+  //members
+  unsigned maxOuterDim_;
+  unsigned outerDim_;
+  bool noOuterDim_;
+  unsigned nEntries_;
+  TritonBatchMode batchMode_;
+  bool manualBatchMode_;
+  bool verbose_;
+  bool useSharedMemory_;
+  TritonServerType serverType_;
+  bool isLocal_;
+  grpc_compression_algorithm compressionAlgo_;
+  triton::client::Headers headers_;
+
+  std::unique_ptr<triton::client::InferenceServerGrpcClient> client_;
+  //stores timeout, model name and version
+  std::vector<triton::client::InferOptions> options_;
+  edm::ServiceToken token_;
+
+private:
+  friend TritonInputData;
+  friend TritonOutputData;
+
+  //private accessors only used by data
+  auto client() { return client_.get(); }
+  void addEntry(unsigned entry);
+  void resizeEntries(unsigned entry);
+};
+
+#endif
diff --git a/HeterogeneousCore/SonicTriton/src/RetryActionDiffServer.cc b/HeterogeneousCore/SonicTriton/src/RetryActionDiffServer.cc
new file mode 100644
index 0000000000000..c3f931ae567ba
--- /dev/null
+++ b/HeterogeneousCore/SonicTriton/src/RetryActionDiffServer.cc
@@ -0,0 +1,36 @@
+#include "HeterogeneousCore/SonicTriton/interface/RetryActionDiffServer.h"
+#include "HeterogeneousCore/SonicTriton/interface/TritonClient.h"
+#include "HeterogeneousCore/SonicTriton/interface/TritonService.h"
+#include "FWCore/MessageLogger/interface/MessageLogger.h"
+#include "FWCore/ServiceRegistry/interface/Service.h"
+
+RetryActionDiffServer::RetryActionDiffServer(
+  const edm::ParameterSet& conf, 
+  SonicClientBase* client
+): RetryActionBase(conf, client) {}
+
+void RetryActionDiffServer::start() {
+  this->shouldRetry_ = true;
+}
+
+void RetryActionDiffServer::retry() {
+  if (!this->shouldRetry_) {
+    this->shouldRetry_ = false;
+    edm::LogInfo("RetryActionDiffServer") << "Retry not armed; skipping.";
+    return;
+  }
+
+  try {
+    auto* tritonClient = static_cast<TritonClient*>(client_);
+    edm::LogInfo("RetryActionDiffServer") << "Attempting retry by switching to fallback server";
+    tritonClient->updateServer(TritonService::Server::fallbackName);
+    eval();
+  } catch (const std::exception& e) {
+    edm::LogError("RetryActionDiffServer") 
+      << "Failed to retry with alternative server: "
+      << e.what();
+  }
+  this->shouldRetry_ = false;
+}
+
+DEFINE_RETRY_ACTION(RetryActionDiffServer);
\ No newline at end of file
diff --git a/HeterogeneousCore/SonicTriton/src/TritonClient.cc b/HeterogeneousCore/SonicTriton/src/TritonClient.cc
index ddcdff83448d0..404c84455e87d 100644
--- a/HeterogeneousCore/SonicTriton/src/TritonClient.cc
+++ b/HeterogeneousCore/SonicTriton/src/TritonClient.cc
@@ -28,6 +28,19 @@
 namespace tc = triton::client;
 
 namespace {
+  // Minimal ParameterSet to satisfy SonicClientBase requirements during unit tests
+  edm::ParameterSet makeMinimalSonicParamsForTest() {
+    edm::ParameterSet params;
+    params.addParameter<std::string>("mode", "PseudoAsync");
+
+    edm::ParameterSet defaultRetry;
+    defaultRetry.addParameter<std::string>("retryType", "RetrySameServerAction");
+    defaultRetry.addUntrackedParameter<unsigned>("allowedTries", 0u);
+    std::vector<edm::ParameterSet> retryVec{defaultRetry};
+    params.addParameter<std::vector<edm::ParameterSet>>("Retry", retryVec);
+
+    return params;
+  }
   grpc_compression_algorithm getCompressionAlgo(const std::string& name) {
     if (name.empty() or name.compare("none") == 0)
       return grpc_compression_algorithm::GRPC_COMPRESS_NONE;
@@ -61,7 +74,7 @@ TritonClient::TritonClient(const edm::ParameterSet& params, const std::string& d
       useSharedMemory_(params.getUntrackedParameter<bool>("useSharedMemory")),
       compressionAlgo_(getCompressionAlgo(params.getUntrackedParameter<std::string>("compression"))) {
   options_.emplace_back(params.getParameter<std::string>("modelName"));
-  //get appropriate server for this model
+
   edm::Service<TritonService> ts;
 
   // We save the token to be able to notify the service in case of an exception in the evaluate method.
@@ -70,22 +83,9 @@ TritonClient::TritonClient(const edm::ParameterSet& params, const std::string& d
   // create the context.
   token_ = edm::ServiceRegistry::instance().presentToken();
 
-  const auto& server =
-      ts->serverInfo(options_[0].model_name_, params.getUntrackedParameter<std::string>("preferredServer"));
-  serverType_ = server.type;
-  edm::LogInfo("TritonDiscovery") << debugName_ << " assigned server: " << server.url;
-  //enforce sync mode for fallback CPU server to avoid contention
-  //todo: could enforce async mode otherwise (unless mode was specified by user?)
-  if (serverType_ == TritonServerType::LocalCPU)
-    setMode(SonicMode::Sync);
-  isLocal_ = serverType_ == TritonServerType::LocalCPU or serverType_ == TritonServerType::LocalGPU;
-
-  //connect to the server
-  TRITON_THROW_IF_ERROR(
-      tc::InferenceServerGrpcClient::Create(&client_, server.url, false, server.useSsl, server.sslOptions),
-      "TritonClient(): unable to create inference context",
-      isLocal_);
-
+  //Connect to server
+  updateServer(params.getUntrackedParameter<std::string>("preferredServer"));
+  
   //set options
   options_[0].model_version_ = params.getParameter<std::string>("modelVersion");
   options_[0].client_timeout_ = params.getUntrackedParameter<unsigned>("timeout");
@@ -369,7 +369,7 @@ void TritonClient::getResults(const std::vector<std::shared_ptr<tc::InferResult>
 //default case for sync and pseudo async
 void TritonClient::evaluate() {
   //undo previous signal from TritonException
-  if (tries_ > 0) {
+  if (totalTries_ > 0) {
     // If we are retrying then the evaluate method is called outside the frameworks TBB thread pool.
     // So we need to setup the service token for the current thread to access the service registry.
     edm::ServiceRegistry::Operate op(token_);
@@ -574,6 +574,26 @@ inference::ModelStatistics TritonClient::getServerSideStatus() const {
   return inference::ModelStatistics{};
 }
 
+void TritonClient::updateServer(std::string serverName){
+      //get appropriate server for this model
+      edm::Service<TritonService> ts;
+
+      const auto& server = ts->serverInfo(options_[0].model_name_, serverName);
+      serverType_ = server.type;
+      edm::LogInfo("TritonDiscovery") << debugName_ << " assigned server: " << server.url;
+      //enforce sync mode for fallback CPU server to avoid contention
+      //todo: could enforce async mode otherwise (unless mode was specified by user?)
+      if (serverType_ == TritonServerType::LocalCPU)
+        setMode(SonicMode::Sync);
+      isLocal_ = serverType_ == TritonServerType::LocalCPU or serverType_ == TritonServerType::LocalGPU;
+    
+      //connect to the server
+      TRITON_THROW_IF_ERROR(
+          tc::InferenceServerGrpcClient::Create(&client_, server.url, false, server.useSsl, server.sslOptions),
+          "TritonClient(): unable to create inference context",
+          isLocal_);
+}
+
 //for fillDescriptions
 void TritonClient::fillPSetDescription(edm::ParameterSetDescription& iDesc) {
   edm::ParameterSetDescription descClient;
@@ -591,3 +611,26 @@ void TritonClient::fillPSetDescription(edm::ParameterSetDescription& iDesc) {
   descClient.addUntracked<std::vector<std::string>>("outputs", {});
   iDesc.add<edm::ParameterSetDescription>("Client", descClient);
 }
+
+void TritonClient::connectToServer(const std::string& url) {
+  // Update client state for a generic remote server
+  serverType_ = TritonServerType::Remote;
+  isLocal_ = false;
+
+  edm::LogInfo("TritonDiscovery") << debugName_ << " connecting to server: " << url;
+
+  // Use default SSL options
+  triton::client::SslOptions sslOptions;
+  bool useSsl = false; // Assuming no SSL for direct URL connection
+
+  // Connect to the server
+  TRITON_THROW_IF_ERROR(
+      triton::client::InferenceServerGrpcClient::Create(&client_, url, false, useSsl, sslOptions),
+      "TritonClient::connectToServer(): unable to create inference context",
+      false // isLocal is false
+  );
+}
+
+//constructor for testing
+TritonClient::TritonClient() : SonicClient(makeMinimalSonicParamsForTest(), "TritonClient_test", "TritonClient") {}
+
diff --git a/HeterogeneousCore/SonicTriton/test/BuildFile.xml b/HeterogeneousCore/SonicTriton/test/BuildFile.xml
index e4ff7a0bb56f3..f6ab9108035f2 100644
--- a/HeterogeneousCore/SonicTriton/test/BuildFile.xml
+++ b/HeterogeneousCore/SonicTriton/test/BuildFile.xml
@@ -1,11 +1,22 @@
 <test name="TestHeterogeneousCoreSonicTritonProducerCPU" command="unittest.sh ${LOCALTOP} CPU"/>
 <test name="TestHeterogeneousCoreSonicTritonProducerGPU" command="unittest.sh ${LOCALTOP} GPU"/>
+<test name="TestHeterogeneousCoreSonicTritonRetryActionSame" command="cmsRun ${LOCALTOP}/src/HeterogeneousCore/SonicTriton/test/tritonTest_cfg.py --modules TritonGraphProducer --models gat_test --maxEvents 2 --unittest --device cpu --retryAction same"/>
+<test name="TestHeterogeneousCoreSonicTritonRetryActionDiff" command="cmsRun ${LOCALTOP}/src/HeterogeneousCore/SonicTriton/test/tritonTest_cfg.py --modules TritonGraphProducer --models gat_test --maxEvents 2 --unittest --device cpu --retryAction diff"/>
 <test name="TestHeterogeneousCoreSonicTritonVersionCheck" command="cmsTritonConfigTool versioncheck">
   <use name="cmsswdata"/>
 </test>
-<library file="*.cc" name="testHeterogenousCoreSonicTriton">
+<library file="TritonGraphModules.cc, TritonIdentityProducer.cc, TritonImageProducer.cc" name="testHeterogenousCoreSonicTriton">
   <flags EDM_PLUGIN="1"/>
   <use name="FWCore/ParameterSet"/>
   <use name="FWCore/Framework"/>
   <use name="HeterogeneousCore/SonicTriton"/>
 </library>
+
+<bin file="test_RetryActionDiffServer.cc" name="TestHeterogeneousCoreSonicTritonRetryActionDiffServer">
+  <use name="catch2"/>
+  <use name="FWCore/ParameterSet"/>
+  <use name="HeterogeneousCore/SonicTriton"/>
+</bin>
+
+<test name="TestHeterogeneousCoreSonicTritonRetryActionDiff_Log" command="retry_action_diff_log_test.sh ${LOCALTOP}"/>
+
diff --git a/HeterogeneousCore/SonicTriton/test/retry_action_diff_log_test.sh b/HeterogeneousCore/SonicTriton/test/retry_action_diff_log_test.sh
new file mode 100755
index 0000000000000..00e1cc1dce90a
--- /dev/null
+++ b/HeterogeneousCore/SonicTriton/test/retry_action_diff_log_test.sh
@@ -0,0 +1,22 @@
+#!/bin/bash
+
+LOCALTOP=$1
+
+tmpFile=$(mktemp -p ${LOCALTOP} RetryActionDiffLogXXXXXXXX.log)
+cmsRun ${LOCALTOP}/src/HeterogeneousCore/SonicTriton/test/tritonTest_cfg.py \
+  --modules TritonGraphProducer --models gat_test \
+  --maxEvents 2 --unittest --device cpu --retryAction diff --verbose \
+  > "$tmpFile" 2>&1
+status=$?
+
+if ! grep -q "Retry type: RetryActionDiffServer" "$tmpFile"; then
+  echo "Expected retry type log line not found" >&2
+  cat "$tmpFile"
+  rm -f "$tmpFile"
+  exit 1
+fi
+
+rm -f "$tmpFile"
+exit $status
+
+
diff --git a/HeterogeneousCore/SonicTriton/test/test_RetryActionDiffServer.cc b/HeterogeneousCore/SonicTriton/test/test_RetryActionDiffServer.cc
new file mode 100644
index 0000000000000..48305dd083216
--- /dev/null
+++ b/HeterogeneousCore/SonicTriton/test/test_RetryActionDiffServer.cc
@@ -0,0 +1,73 @@
+#include "catch.hpp"
+
+#include "HeterogeneousCore/SonicTriton/interface/RetryActionDiffServer.h"
+#include "HeterogeneousCore/SonicTriton/interface/TritonClient.h"
+#include "HeterogeneousCore/SonicTriton/interface/TritonService.h"
+#include "HeterogeneousCore/SonicCore/interface/RetryActionBase.h"
+
+#include "FWCore/ParameterSet/interface/ParameterSet.h"
+
+#include <string>
+
+// Test double for TritonClient to observe updateServer calls without framework/services
+class TestTritonClient : public TritonClient {
+public:
+  TestTritonClient() : TritonClient() {}
+
+  void connectToServer(const std::string& url) override { lastConnectedUrl = url; }
+
+  void updateServer(std::string serverName) override {
+    lastUpdatedServerName = std::move(serverName);
+  }
+
+  const std::string& lastUrl() const { return lastConnectedUrl; }
+  const std::string& lastServerName() const { return lastUpdatedServerName; }
+
+protected:
+  void evaluate() override {}
+
+private:
+  std::string lastConnectedUrl;
+  std::string lastUpdatedServerName;
+};
+
+TEST_CASE("RetryActionDiffServer switches to fallback via updateServer", "[RetryActionDiffServer]") {
+  edm::ParameterSet empty;
+  TestTritonClient client;
+
+  RetryActionDiffServer action(empty, static_cast<SonicClientBase*>(&client));
+
+  // start should arm the action
+  action.start();
+  REQUIRE(action.shouldRetry());
+
+  // retry should call updateServer with fallback name then disarm
+  action.retry();
+  REQUIRE(client.lastServerName() == TritonService::Server::fallbackName);
+
+  // second retry without re-arming should be a no-op: lastServerName unchanged
+  std::string afterFirst = client.lastServerName();
+  action.retry();
+  REQUIRE(client.lastServerName() == afterFirst);
+}
+
+// A client that throws during updateServer to exercise error handling path
+class ThrowingTritonClient : public TritonClient {
+public:
+  ThrowingTritonClient() : TritonClient() {}
+  void updateServer(std::string) override { throw std::runtime_error("updateServer failure"); }
+protected:
+  void evaluate() override {}
+};
+
+TEST_CASE("RetryActionDiffServer catches exceptions from updateServer", "[RetryActionDiffServer]") {
+  edm::ParameterSet empty;
+  ThrowingTritonClient client;
+  RetryActionDiffServer action(empty, static_cast<SonicClientBase*>(&client));
+  action.start();
+
+  // Should not throw despite client throwing internally; action disarms afterward
+  REQUIRE_NOTHROW(action.retry());
+}
+
+
diff --git a/HeterogeneousCore/SonicTriton/test/tritonTest_cfg.py b/HeterogeneousCore/SonicTriton/test/tritonTest_cfg.py
index 9cede0e496706..b5915ea1a05ce 100644
--- a/HeterogeneousCore/SonicTriton/test/tritonTest_cfg.py
+++ b/HeterogeneousCore/SonicTriton/test/tritonTest_cfg.py
@@ -45,6 +45,7 @@
 parser.add_argument("--device", default="auto", type=str.lower, choices=allowed_devices, help="specify device for fallback server")
 parser.add_argument("--container", default="apptainer", type=str.lower, choices=allowed_containers, help="specify container for fallback server")
 parser.add_argument("--tries", default=0, type=int, help="number of retries for failed request")
+parser.add_argument("--retryAction", default="same", type=str, choices=["same","diff"], help="retry policy: same server or different server")
 options = parser.parse_args()
 
 if len(options.params)>0:
@@ -108,6 +109,9 @@
     keepMsgs.append('TritonClient')
 if options.verbose or options.verboseService:
     keepMsgs.append('TritonService')
+if options.verbose:
+    # ensure RetryActionDiffServer messages are not suppressed if emitted
+    keepMsgs.append('RetryActionDiffServer')
 
 for im,module in enumerate(options.modules):
     model = options.models[im]
@@ -123,9 +127,20 @@
                 modelVersion = cms.string(""),
                 modelConfigPath = cms.FileInPath("HeterogeneousCore/SonicTriton/data/models/{}/config.pbtxt".format(model)),
                 verbose = cms.untracked.bool(options.verbose or options.verboseClient),
-                allowedTries = cms.untracked.uint32(options.tries),
                 useSharedMemory = cms.untracked.bool(not options.noShm),
                 compression = cms.untracked.string(options.compression),
+                Retry = (
+                  cms.VPSet(
+                    cms.PSet(
+                      retryType = cms.string('RetrySameServerAction'),
+                      allowedTries = cms.untracked.uint32(options.tries)
+                    )
+                  ) if options.retryAction == 'same' else cms.VPSet(
+                    cms.PSet(
+                      retryType = cms.string('RetryActionDiffServer')
+                    )
+                  )
+                )
             )
         )
     )
@@ -148,6 +163,8 @@
             processModule.edgeMax = cms.uint32(15000)
         processModule.brief = cms.bool(options.brief)
     process.p += processModule
+    if options.verbose:
+        print("Retry type:", ('RetrySameServerAction' if options.retryAction == 'same' else 'RetryActionDiffServer'))
     if options.verbose or options.verboseClient:
         keepMsgs.extend([module,module+':TritonClient'])
     if options.testother: