Remove MPI dependency from api_cc library to avoid linking issues

Co-authored-by: njzjz <9496702+njzjz@users.noreply.github.com>

Author: Copilot

doc/development/pytorch-profiler.md

@@ -14,11 +14,15 @@ export DP_PYTORCH_PROFILER_OUTPUT_DIR=./profiler_results
 
 3. Check for profiler output in the specified directory:
 ```bash
-# For single-rank or non-MPI usage
 ls -la ./profiler_results/pytorch_profiler_trace.json
+```
 
-# For MPI usage, each rank gets its own file
-ls -la ./profiler_results/pytorch_profiler_trace_rank*.json
+For MPI applications, you can use different output directories per rank:
+```bash
+# Example for rank 0
+export DP_PYTORCH_PROFILER_OUTPUT_DIR=./profiler_results_rank0
+# Example for rank 1  
+export DP_PYTORCH_PROFILER_OUTPUT_DIR=./profiler_results_rank1
 ```
 
 ## Environment Variables
@@ -32,13 +36,11 @@ The profiler uses PyTorch's modern `torch::profiler` API and automatically:
 - Creates the output directory if it doesn't exist
 - Profiles all forward pass operations in DeepPotPT and DeepSpinPT
 - Saves profiling results to a JSON file when the object is destroyed
-- Automatically includes MPI rank in filename when MPI is available and initialized
 
 ## Output Files
 
-- **Single-rank or non-MPI usage**: `pytorch_profiler_trace.json`
-- **MPI usage**: `pytorch_profiler_trace_rank{rank}.json` (e.g., `pytorch_profiler_trace_rank0.json`, `pytorch_profiler_trace_rank1.json`)
+- **All usage**: `pytorch_profiler_trace.json`
 
-This ensures that each MPI rank saves its profiling data to a separate file, preventing conflicts in multi-rank simulations.
+For MPI applications, users can distinguish between ranks by setting different output directories per rank using the `DP_PYTORCH_PROFILER_OUTPUT_DIR` environment variable.
 
 This is intended for development and debugging purposes.

source/api_cc/include/common.h

@@ -173,8 +173,8 @@ void get_env_nthreads(int& num_intra_nthreads, int& num_inter_nthreads);
 void get_env_pytorch_profiler(bool& enable_profiler, std::string& output_dir);
 
 /**
- * @brief Get MPI rank if MPI is available and initialized, otherwise return -1.
- * @return The MPI rank or -1 if MPI is not available/initialized.
+ * @brief Get MPI rank. Currently disabled in api_cc to avoid MPI linking dependencies.
+ * @return Always returns -1. Users can distinguish ranks using different output directories.
  **/
 int get_mpi_rank();
 

source/api_cc/src/common.cc

@@ -11,12 +11,10 @@
 #include <sys/stat.h>
 #include <errno.h>
 
-// Try to include MPI if available - this will be a no-op if MPI is not available
-#ifdef __has_include
-  #if __has_include(<mpi.h>)
-    #include <mpi.h>
-  #endif
-#endif
+// Note: MPI rank detection has been removed from api_cc library
+// to avoid MPI linking dependencies. The profiler will use a generic
+// filename. Users can still distinguish between ranks by using different
+// output directories per rank if needed.
 
 #include "AtomMap.h"
 #include "device.h"
@@ -410,17 +408,10 @@ void deepmd::get_env_pytorch_profiler(bool& enable_profiler, std::string& output
 }
 
 int deepmd::get_mpi_rank() {
-  int rank = -1;  // Use -1 to indicate MPI not available/initialized
-  // Try to detect MPI at runtime
-  #ifdef MPI_H
-  int initialized = 0;
-  if (MPI_Initialized(&initialized) == MPI_SUCCESS && initialized) {
-    if (MPI_Comm_rank(MPI_COMM_WORLD, &rank) != MPI_SUCCESS) {
-      rank = -1;  // fallback to -1 if MPI_Comm_rank fails
-    }
-  }
-  #endif
-  return rank;
+  // MPI rank detection removed from api_cc to avoid MPI linking dependencies
+  // Always return -1 to indicate no MPI rank available
+  // Users can distinguish between ranks by using different output directories
+  return -1;
 }
 
 bool deepmd::create_directories(const std::string& path) {