Add experimental custom routing for MoE

llama.cpp

@@ -2469,8 +2469,37 @@ static void llm_load_hparams(
     ml.get_key  (LLM_KV_ATTENTION_HEAD_COUNT, hparams.n_head);
     ml.get_key  (LLM_KV_BLOCK_COUNT,          hparams.n_layer);
     ml.get_key  (LLM_KV_EXPERT_COUNT,         hparams.n_expert,      false);
-    ml.get_key  (LLM_KV_EXPERT_USED_COUNT,    hparams.n_expert_used, false);
 
+    // Declare the custom expert used count variable and initialize it to 1
+    int CUSTOM_EXPERT_USED_COUNT = 1;
+    std::string filename = "experts.txt";
+
+    // Try to open the file for reading
+    std::ifstream infile(filename);
+
+    if (infile.is_open()) {
+        if (!(infile >> CUSTOM_EXPERT_USED_COUNT)) {
+            // If reading fails, set CUSTOM_EXPERT_USED_COUNT to an error value or handle the error as needed
+            printf("Error reading from file: %s\n", filename.c_str());
+            CUSTOM_EXPERT_USED_COUNT = 1; // Default value or error value
+        }
+        infile.close(); // Close the file after reading or failing to read
+    } else {
+        // The file doesn't exist or couldn't be opened for reading. Try creating it.
+        std::ofstream outfile(filename);
+        if (outfile.is_open()) {
+            outfile << CUSTOM_EXPERT_USED_COUNT; // Write 1 to the file
+            outfile.close(); // Close the file after writing
+        } else {
+            // If the file couldn't be opened for writing, print an error message
+            printf("Error creating file: %s\n", filename.c_str());
+        }
+    }
+
+    // Setter for the number of experts that will be used
+    hparams.n_expert_used = CUSTOM_EXPERT_USED_COUNT;
+
+    // Perform assertions to ensure valid parameters are being used
     GGML_ASSERT(hparams.n_expert <= LLAMA_MAX_EXPERTS);
     GGML_ASSERT(hparams.n_expert_used <= hparams.n_expert);
     if (hparams.n_expert > 0) {
@@ -2479,6 +2508,9 @@ static void llm_load_hparams(
         GGML_ASSERT(hparams.n_expert_used == 0);
     }
 
+    // Print out the number of experts that will be used
+    printf("-------------------------------------------------------\nNumber of experts that will be used per token (if MoE): %d\n-------------------------------------------------------\n", hparams.n_expert_used);
+
     // n_head_kv is optional, default to n_head
     hparams.n_head_kv = hparams.n_head;
     ml.get_key(LLM_KV_ATTENTION_HEAD_COUNT_KV, hparams.n_head_kv, false);