refactor: remove unused variables and parameters

infini_train/src/autograd/comm.cc

@@ -18,7 +18,6 @@ Scatter::Scatter(const std::vector<Device> &target_gpus, int64_t dim,
 std::vector<std::shared_ptr<Tensor>> Scatter::Forward(const std::vector<std::shared_ptr<Tensor>> &input_tensors) {
     const auto &input = input_tensors[0];
     std::vector<std::shared_ptr<Tensor>> output_tensors;
-    auto device = input->GetDevice().type();
     output_tensors = pg_->Scatter(input, target_gpus_, dim_);
     return output_tensors;
 }
@@ -51,7 +50,7 @@ std::vector<std::shared_ptr<Tensor>> Gather::Forward(const std::vector<std::shar
     } else {
         unsqueezed_scalar_ = false;
     }
-    auto device = input_tensors[0]->GetDevice().type();
+
     return {pg_->Gather(input_tensors, target_device_, dim_)};
 }
 

infini_train/src/kernels/cpu/linear.cc

@@ -22,7 +22,6 @@ std::shared_ptr<Tensor> LinearForward(const std::shared_ptr<Tensor> &input, cons
 
     const auto &input_dims = input->Dims();
     CHECK_GE(input_dims.size(), 2);
-    const int64_t bs = std::accumulate(input_dims.rbegin() + 1, input_dims.rend(), 1, std::multiplies<int64_t>{});
     const int64_t in_features = *input_dims.rbegin();
 
     const auto &weight_dims = weight->Dims();

infini_train/src/kernels/cpu/reduction.cc

@@ -18,7 +18,6 @@ std::shared_ptr<Tensor> ReduceOpForward(const std::shared_ptr<Tensor> &input, co
     CHECK_LT(actual_dim, input_dims.size());
 
     std::vector<int64_t> output_dims = input_dims;
-    const int64_t reduce_size = input_dims[dim];
     if (keep_dim) {
         output_dims[actual_dim] = 1;
     } else {

infini_train/src/nn/modules/transformer/causal_self_attention.cc

@@ -88,7 +88,6 @@ CausalSelfAttention::ForwardStandard(const std::vector<std::shared_ptr<infini_tr
     auto tp_world_size = parallel::global::GetTensorParallelSize();
 
     const auto B = x[0]->Dims()[0];                  // bs
-    const auto C = x[0]->Dims()[2];                  // n_embd
     const int64_t head_dim = n_embd_ / n_head_;      // per-head dim (global)
     const int64_t local_C = n_embd_ / tp_world_size; // per-rank hidden
 

infini_train/src/nn/modules/transformer/mlp.cc

@@ -35,7 +35,6 @@ MLP::MLP(const TransformerConfig &config) : CloneableModule(kType) {
     }
 
     // Round up to multiple_of
-    int64_t before_round = ffn_hidden;
     ffn_hidden = (ffn_hidden + config.multiple_of - 1) / config.multiple_of * config.multiple_of;
 
     // c_fc: ColumnParallel (input full, output parallel)

infini_train/src/nn/parallel/ddp/reducer.cc

@@ -368,7 +368,6 @@ void Reducer::MarkBucketReady(size_t bucket_index) {
     }
     // From next_bucket_, launch ready buckets(pending==0) in turn
     while (next_bucket_ < buckets_.size() && buckets_[next_bucket_].pending == 0) {
-        auto &bucket = buckets_[next_bucket_];
         FinalizeBucketDense(next_bucket_);
         ++next_bucket_;
     }

infini_train/src/nn/parallel/tensor_parallel.cc

@@ -442,7 +442,6 @@ VocabParallelCrossEntropy::Forward(const std::vector<std::shared_ptr<Tensor>> &i
     auto logits = std::make_shared<Tensor>(input_tensors[0]->To(DataType::kFLOAT32));
     auto target = input_tensors[1];
 
-    auto dtype = logits->Dtype();
     auto device = logits->GetDevice();
 
     CHECK(target->Dtype() == DataType::kINT64) << "target must be int64";

infini_train/src/tensor.cc

@@ -786,7 +786,6 @@ void Tensor::Print(std::ostream &os) const {
     CHECK(dtype_ == DataType::kFLOAT32);
 
     const size_t num_elements = NumElements();
-    const size_t num_bytes = num_elements * sizeof(float);
 
     auto impl = core::GetDeviceGuardImpl(GetDevice().type());