remove comments

inference/models/mixtral.cc

@@ -324,11 +324,7 @@ void MIXTRAL::create_mixtral_model(FFModel &ff,
       aggregate_inputs[4 + expert_idx] = w2;
     }
 
-    // TODO uncomment, but is a nice-to-have at this point.. or try normalizing with softmax????
-//    Tensor topk_values_reduced = ff.reduce_sum(topk_values, {0}, true);
-//    topk_values = ff.divide(topk_values, topk_values_reduced);
-
-  // TODO have 2 fixed inputs instead of 4
+    // Those two inputs are quick fixes to make aggregate work. They are not used in inference.
     Tensor topk_values_DUMMY = ff.softmax(
         topk_values,
         -1,

src/ops/aggregate.cc

@@ -329,8 +329,6 @@ OpMeta *Aggregate::init_task(Task const *task,
 }
 
 void Aggregate::forward(FFModel const &ff) {
-  printf("running Aggregate::forward\n");
-  printf("\n\n\n\n Aggregate::init is running!!!!!!!!!! \n\n\n\n"); // Don't expect this to run
   ArgumentMap argmap;
   Context ctx = ff.config.lg_ctx;
   Runtime *runtime = ff.config.lg_hlr;
@@ -344,7 +342,6 @@ void Aggregate::forward(FFModel const &ff) {
                          0 /*mapper_id*/,
                          outputs[0]->machine_view.hash());
 
-  printf("Entered Aggregate::forward\n");
   // gate_preds
   launcher.add_region_requirement(RegionRequirement(inputs[0]->part,
                                                     0 /*projection id*/,
@@ -463,7 +460,6 @@ FutureMap Aggregate::inference(FFModel const &ff,
                                                     WRITE_ONLY,
                                                     EXCLUSIVE,
                                                     batch_outputs[0]->region));
-//  launcher.add_field(n + 2, FID_DATA);
   launcher.add_field(n + FIXED_ARG_CNT, FID_DATA);
 
   return runtime->execute_index_space(ctx, launcher);

src/ops/inc_multihead_self_attention.cc

@@ -736,7 +736,6 @@ bool IncMultiHeadSelfAttention::get_int_parameter(PMParameter para,
   }
 }
 
-// Just for benchmarking, don't need that
 bool IncMultiHeadSelfAttention::measure_operator_cost(
     Simulator *sim, MachineView const &mv, CostMetrics &cost_metrics) const {
   return false;

src/ops/inc_multihead_self_attention.cu

@@ -1517,7 +1517,6 @@ void IncMultiHeadSelfAttention::inference_kernel_wrapper(
   assert(input.data_type == output.data_type);
 
   if (input.data_type == DT_HALF) {
-    // calling input.get_inc_ptr() below would cause a legion error, "type mismatch get index space domain"
     Kernels::IncMultiHeadAttention::inference_kernel(
         m, bc, shard_id, input.get_half_ptr(), output.get_half_ptr(), stream);
   } else if (input.data_type == DT_FLOAT) {

src/runtime/request_manager.cc

@@ -26,7 +26,7 @@
 #include <nlohmann/json.hpp>
 #include <stack>
 #include <stdexcept>
-#include <cassert> // For assert
+#include <cassert>
 
 namespace FlexFlow {