fix file sorting

build.sh

@@ -0,0 +1,8 @@
+#!/bin/bash
+#rm -rf build
+cmake -S . -B build -DLLAMA_CUBLAS=ON -DLLAMA_GGML_PERF=ON #-DLLAMA_RUN_WARMUP=OFF
+cmake --build build --config Release
+
+# if DLLAMA_GGML_PERF=ON
+# -> avg exec time of operator in prefill & decode stage.
+# -> ratio of active neurons located in the CPU

llama.cpp

@@ -2863,6 +2863,9 @@ struct llama_gpu_split_loader {
             }
         }
 
+        ggml_tensor * up = model->layers[0].ffn_up;
+        printf("\n>>> %ld, %ld\n", up->ne[0], up->ne[1]);
+
         const int64_t t_mlp_us = ggml_time_us() - t_start_mlp_us;
         LLAMA_LOG_INFO(" done (%.2f ms)\n", t_mlp_us / 1000.0);
 

powerinfer-py/powerinfer/export_split.py

@@ -18,7 +18,10 @@ def load_activation_weights(models_base: Path):
     # But for now, let's assume it is a plain directory of activation_{0, ... , n_layers - 1}.pt
     *_, files = next(os.walk(models_base))
     activation_files = [f for f in files if re.match(r"activation_\d+.pt", f)]
-    activation_files.sort()
+
+    layer_num = np.array([int(re.sub(f'[^0-9]', '', f)) for f in activation_files])
+    idx = np.argsort(layer_num)
+    activation_files = [activation_files[i] for i in idx]
     return [torch.load(models_base / f) for f in activation_files]
 
 def append_gpu_idx(gguf: GGUFWriter, i_layer: int, activation, select_count) -> None:

scripts/pg19_firstbook_128.txt

@@ -0,0 +1 @@
+Half-way down the Rue Saint-Denis, almost at the corner of the Rue du Petit-Lion, there stood formerly one of those delightful houses which enable historians to reconstruct old Paris by analogy. The threatening walls of this tumbledown abode seemed to have been decorated with hieroglyphics. For what other name could the passer-by give to the Xs and Vs which the horizontal or diagonal timbers traced on the front, outlined by little parallel cracks in the plaster? It was evident that every beam quivered in its mortices at the passing of the lightest vehicle. This venerable structure was crowned by a