Fix guidance bug

CMakeLists.txt

@@ -235,6 +235,31 @@ if(USE_GUIDANCE)
     # bcrypt is needed for the rust std lib
     target_link_libraries(onnxruntime-genai PRIVATE bcrypt)
   endif()
+  if(MSVC)
+    # The Rust llguidance static library is always compiled against the release MSVC CRT
+    # (Rust has no debug CRT concept). The .lib embeds /DEFAULTLIB directives for the release
+    # CRT and /NODEFAULTLIB directives that suppress the debug CRT (msvcrtd, ucrtd, vcruntimed).
+    # In Debug builds, C++ code (e.g. onnxruntime-extensions) is compiled with /MDd and references
+    # debug-only CRT functions like _CrtDbgReport (in ucrtd.lib). Because the Rust .lib suppresses
+    # ucrtd.lib via its embedded /NODEFAULTLIB, _CrtDbgReport becomes unresolved.
+    #
+    # Fix: explicitly add the debug CRT import libraries in Debug builds. Explicitly specified
+    # libraries are not affected by /NODEFAULTLIB directives (those only suppress /DEFAULTLIB
+    # auto-linking). Also suppress the conflicting release CRT to avoid LNK4098 warnings.
+    target_link_libraries(onnxruntime-genai PRIVATE
+      $<$<CONFIG:Debug>:msvcrtd.lib>
+      $<$<CONFIG:Debug>:ucrtd.lib>
+      $<$<CONFIG:Debug>:vcruntimed.lib>
+    )
+    target_link_options(onnxruntime-genai PRIVATE
+      $<$<CONFIG:Debug>:/NODEFAULTLIB:msvcrt.lib>
+      $<$<CONFIG:Debug>:/NODEFAULTLIB:ucrt.lib>
+      $<$<CONFIG:Debug>:/NODEFAULTLIB:vcruntime.lib>
+      $<$<CONFIG:Debug>:/NODEFAULTLIB:libcmt.lib>
+      $<$<CONFIG:Debug>:/NODEFAULTLIB:libucrt.lib>
+      $<$<CONFIG:Debug>:/NODEFAULTLIB:libvcruntime.lib>
+    )
+  endif()
 endif()
 
 if(CMAKE_GENERATOR_TOOLSET MATCHES "Visual Studio")

src/constrained_logits_processor.cpp

@@ -260,8 +260,10 @@ std::unique_ptr<ConstrainedLogitsProcessor> CreateGuidanceLogitsProcessor(const
   if (!state.params_->guidance_type.empty() && !state.params_->guidance_data.empty()) {
 #if USE_GUIDANCE
     return std::make_unique<GuidanceLogitsProcessor>(state);
+#else
+    if (g_log.enabled)
+      Log("warning", "No supported ConstrainedLogitsProcessor found. To use guidance, build with use_guidance=true");
 #endif
-    Log("warning", "No supported ConstrainedLogitsProcessor found. To use guidance, build with use_guidance=true");
   }
   return nullptr;
 }

src/generators.cpp

@@ -459,7 +459,10 @@ void Generator::ComputeLogits(DeviceSpan<int32_t> next_tokens) {
   if (computed_logits_)
     throw std::runtime_error("ComputeLogits called again without calling AppendTokens or GenerateNextToken first");
 
-  if (last_action_ == Action::generated && guidance_logits_processor_) {
+  // search_->GetSequenceLength() != next_tokens.size() implies that this is not the first time ComputeLogits
+  // is being called (i.e. we're not computing logits for the initial input tokens), so we need to commit
+  // tokens to the guidance logits processor before running the model.
+  if (guidance_logits_processor_ && search_->GetSequenceLength() != next_tokens.size()) {
     auto next_tokens_span = next_tokens.CopyDeviceToCpu();
     guidance_logits_processor_->CommitTokens(next_tokens_span);
   }
@@ -472,7 +475,7 @@ void Generator::ComputeLogits(DeviceSpan<int32_t> next_tokens) {
   }
   SetLogits(logits);
 
-  if (last_action_ == Action::generated && guidance_logits_processor_) {
+  if (guidance_logits_processor_ && search_->GetSequenceLength() != next_tokens.size()) {
     auto ff_tokens = guidance_logits_processor_->GetFFTokens(0);
     if (!ff_tokens.empty()) {
       // process fast-forward tokens
@@ -514,7 +517,7 @@ bool Generator::IsDone() {
   if (is_done) {
     state_->Finalize(search_->GetSequenceLength());
     if (guidance_logits_processor_) {
-      guidance_logits_processor_->Reset();
+      guidance_logits_processor_->ResetWithoutCompute();
       last_action_ = Action::standard;
     }
   }
@@ -619,7 +622,7 @@ void Generator::RewindToLength(size_t new_length) {
   search_->RewindTo(new_length);
   state_->RewindTo(new_length);
   if (guidance_logits_processor_) {
-    guidance_logits_processor_->Reset();
+    guidance_logits_processor_->ResetWithoutCompute();
   }
   computed_logits_ = false;
   last_action_ = Action::rewound;

src/models/model.cpp

@@ -66,7 +66,7 @@ class DirGuard {
   }
 
   ~DirGuard() {
-    if (CHDIR(original_dir_.c_str()) != 0) {
+    if (CHDIR(original_dir_.c_str()) != 0 && g_log.enabled) {
       Log("warning", "Failed to change back to original directory: " + original_dir_.string());
     }
   }

src/models/qwen_vl_model.cpp

@@ -196,10 +196,11 @@ void Qwen2_5_VL_PipelineState::InjectVisionEmbeddings(const std::string& embeddi
 
   // Warn if there's a mismatch between image tokens and vision features
   if (image_embed_consumed_ != static_cast<size_t>(num_vision_tokens)) {
-    Log("warning", "Vision embedding mismatch: consumed " + std::to_string(image_embed_consumed_) +
-                       " of " + std::to_string(num_vision_tokens) + " available vision tokens. " +
-                       "This may indicate a mismatch between the number of image placeholders in the prompt " +
-                       "and the number of images provided.");
+    if (g_log.enabled)
+      Log("warning", "Vision embedding mismatch: consumed " + std::to_string(image_embed_consumed_) +
+                         " of " + std::to_string(num_vision_tokens) + " available vision tokens. " +
+                         "This may indicate a mismatch between the number of image placeholders in the prompt " +
+                         "and the number of images provided.");
   }
 }