fix: move FillMask vector allocs into rank-0 block, use PRIVATE for internal CMake dep

Author: windreamer

src/turbomind/generation/CMakeLists.txt

@@ -15,10 +15,9 @@
 cmake_minimum_required(VERSION 3.11)
 
 add_library(guided_decoding STATIC guided_decoding.cc)
-target_link_libraries(guided_decoding PUBLIC
-    apply_token_bitmask_inplace_cuda
-    xgrammar
-    core)
+target_link_libraries(guided_decoding
+    PUBLIC xgrammar core
+    PRIVATE apply_token_bitmask_inplace_cuda)
 set_property(TARGET guided_decoding PROPERTY POSITION_INDEPENDENT_CODE ON)
 
 add_library(generation STATIC

src/turbomind/generation/guided_decoding.cc

@@ -66,12 +66,12 @@ void GuidedDecoding::FillMask(int phase, TensorMap& env)
                            nullptr,
                            0};
 
-        std::vector<xgrammar::GrammarMatcher> active_matchers;
-        std::vector<int32_t>                  active_indices;
-        active_matchers.reserve(gs);
-        active_indices.reserve(gs);
-
         if (tp_group_->rank() == 0) {
+            std::vector<xgrammar::GrammarMatcher> active_matchers;
+            std::vector<int32_t>                  active_indices;
+            active_matchers.reserve(gs);
+            active_indices.reserve(gs);
+
             for (int i = 0; i < gs; ++i) {
                 if (const auto& m = d.matchers[i]; m && !m->IsTerminated()) {
                     active_matchers.emplace_back(*m);