Cherry pick guidance fix into 0.11.1 release (#1872)

### Description

This PR cherry-picks the guidance fix PR into rel-0.11.1.

### Motivation and Context

This cherry-pick needs to be included for the 0.11.1 patch release.

Author: kunal-vaishnavi

src/constrained_logits_processor.cpp

@@ -229,6 +229,9 @@ void GuidanceLogitsProcessor::ResetWithoutCompute() {
     }
     llg_constraints_[i] = std::unique_ptr<LlgConstraint, LlgConstraintDeleter>(constraint_ptr);
   }
+  for (int i = 0; i < ff_tokens_batch_.size(); i++) {
+    ff_tokens_batch_[i].clear();
+  }
 }
 
 // Reset the masks and llguidance constraints and then recompute the mask

src/generators.cpp

@@ -444,7 +444,7 @@ void Generator::SetRuntimeOption(const char* key, const char* value) {
   state_->SetRunOption(key, value);
 }
 
-bool Generator::IsDone() const {
+bool Generator::IsDone() {
   ThrowErrorIfSessionTerminated(state_->session_terminated_);
   if (computed_logits_) {
     return false;
@@ -453,6 +453,10 @@ bool Generator::IsDone() const {
   bool is_done = search_->IsDone();
   if (is_done) {
     state_->Finalize(search_->GetSequenceLength());
+    if (guidance_logits_processor_) {
+      guidance_logits_processor_->Reset();
+      last_action_ = Action::standard;
+    }
   }
 
   return is_done;

src/generators.h

@@ -94,7 +94,7 @@ struct GeneratorParams : std::enable_shared_from_this<GeneratorParams>, LeakChec
 struct Generator : LeakChecked<Generator> {
   Generator(const Model& model, const GeneratorParams& params);
 
-  bool IsDone() const;
+  bool IsDone();
   void AppendTokens(cpu_span<const int32_t> input_ids);
   void GenerateNextToken();
   void RewindToLength(size_t new_length);  // Rewind state to new_length

src/ort_genai.h

@@ -444,7 +444,7 @@ struct OgaGenerator : OgaAbstract {
     return std::unique_ptr<OgaGenerator>(p);
   }
 
-  bool IsDone() const {
+  bool IsDone() {
     return OgaGenerator_IsDone(this);
   }
 

src/ort_genai_c.cpp

@@ -409,7 +409,7 @@ OgaResult* OgaCreateGenerator(const OgaModel* model, const OgaGeneratorParams* p
   OGA_CATCH
 }
 
-bool OGA_API_CALL OgaGenerator_IsDone(const OgaGenerator* generator) {
+bool OGA_API_CALL OgaGenerator_IsDone(OgaGenerator* generator) {
   return generator->IsDone();
 }
 

src/ort_genai_c.h

@@ -445,7 +445,7 @@ OGA_EXPORT void OGA_API_CALL OgaDestroyGenerator(OgaGenerator* generator);
  * \param[in] generator The generator to check if it is done with generating all sequences.
  * \return True if the generator has finished generating all the sequences, false otherwise.
  */
-OGA_EXPORT bool OGA_API_CALL OgaGenerator_IsDone(const OgaGenerator* generator);
+OGA_EXPORT bool OGA_API_CALL OgaGenerator_IsDone(OgaGenerator* generator);
 OGA_EXPORT bool OGA_API_CALL OgaGenerator_IsSessionTerminated(const OgaGenerator* generator);
 
 /**

src/python/python.cpp

@@ -256,7 +256,7 @@ struct PyGenerator {
     generator_->RewindTo(new_length);
   }
 
-  bool IsDone() const {
+  bool IsDone() {
     return generator_->IsDone();
   }