fix(rng): use fixed seed for deterministic LoRA init

Replace std::random_device with 42 + omp_get_thread_num() to ensure reproducible LoRA initialization across runs.

Author: chen2021673

example/llama3/main.cc

@@ -175,10 +175,6 @@ void Train(const nn::parallel::Rank &rank) {
         model = std::make_shared<LLaMA3>(model_config);
     }
 
-    model->To(device);
-
-    utils::PrecisionChecker::BuildNameMap(model.get());
-
     // Apply LoRA using GetLoRAModel (in-place injection)
     bool lora_enabled = FLAGS_lora_rank > 0;
     if (lora_enabled) {
@@ -198,6 +194,10 @@ void Train(const nn::parallel::Rank &rank) {
         nn::lora::PrintLoRASummary(model, rank.GlobalRank());
     }
 
+    model->To(device);
+
+    utils::PrecisionChecker::BuildNameMap(model.get());
+
     LOG(INFO) << "Rank " << rank.GlobalRank() << ": Model loaded to device.";
 
     DataType dtype;

infini_train/src/nn/init.cc

@@ -22,8 +22,7 @@
 
 namespace infini_train::nn::init {
 namespace {
-static std::random_device rd;
-static std::mt19937 gen(rd());
+static std::mt19937 gen(42);
 } // namespace
 
 std::shared_ptr<Tensor> Normal(const std::shared_ptr<Tensor> &tensor, float mean, float std,
@@ -34,7 +33,7 @@ std::shared_ptr<Tensor> Normal(const std::shared_ptr<Tensor> &tensor, float mean
 #ifdef USE_OMP
 #pragma omp parallel
     {
-        std::mt19937 local_gen(std::random_device{}() + omp_get_thread_num());
+        std::mt19937 local_gen(42 + omp_get_thread_num());
         std::normal_distribution<float> local_dis(mean, std);
 #pragma omp for
         for (int i = 0; i < buffer.size(); ++i) {
@@ -126,7 +125,7 @@ std::shared_ptr<Tensor> Uniform(const std::shared_ptr<Tensor> &tensor, float a,
 #ifdef USE_OMP
 #pragma omp parallel
     {
-        std::mt19937 local_gen(std::random_device{}() + omp_get_thread_num());
+        std::mt19937 local_gen(42 + omp_get_thread_num());
         std::uniform_real_distribution<float> local_dis(a, b);
 #pragma omp for
         for (int i = 0; i < buffer.size(); ++i) {