add fallback for m chip & fix compiler bugs (ggerganov#4)

CMakeLists.txt

@@ -37,7 +37,7 @@ endif()
 #
 
 if (APPLE)
-    set(LLAMA_METAL_DEFAULT ON)
+    set(LLAMA_METAL_DEFAULT OFF) # metal has not been supported on Apple Silicon yet
 else()
     set(LLAMA_METAL_DEFAULT OFF)
 endif()

ggml.c

@@ -146,7 +146,7 @@ void ggml_print_backtrace(void) {
 }
 #endif
 
-#define GGML_PERF
+// #define GGML_PERF
 #define GGML_DEBUG 0
 #define GGML_GELU_FP16
 #define GGML_GELU_QUICK_FP16
@@ -14436,6 +14436,7 @@ static void ggml_compute_forward_mul_mat_axpy_dense(
     // 计算剩余的元素个数
     int remainder = ne00 % 8;
 
+#if defined(__AVX2__)
     // 使用AVX指令进行向量化计算
     for (i = 0; i < ne00 - remainder; i += 8) {
         __m256 res_vec = _mm256_loadu_ps(res + i);  // 加载res中的8个浮点数
@@ -14448,10 +14449,11 @@ static void ggml_compute_forward_mul_mat_axpy_dense(
     for (i = ne00 - remainder; i < ne00; i++) {
         res[i] += tmp[i];
     }
-    // for (i = 0; i < dst->ne[0]; i++) {
-    //     res[i] += tmp[i];
-    // }
-
+#else
+    for (i = 0; i < dst->ne[0]; i++) {
+        res[i] += tmp[i];
+    }
+#endif
     atomic_flag_clear(&g_axpy_dense_lock);
 
 }
@@ -14586,6 +14588,7 @@ static void ggml_compute_forward_mul_mat_axpy(
         // 计算剩余的元素个数
         int remainder = ne00 % 8;
 
+#if defined(__AVX2__)
         // 使用AVX指令进行向量化计算
         for (i = 0; i < ne00 - remainder; i += 8) {
             __m256 res_vec = _mm256_loadu_ps(res + i);  // 加载res中的8个浮点数
@@ -14598,8 +14601,11 @@ static void ggml_compute_forward_mul_mat_axpy(
         for (i = ne00 - remainder; i < ne00; i++) {
             res[i] += tmp[i];
         }
-
-
+#else
+        for (i = 0; i < ne00; i++) {
+            res[i] += tmp[i];
+        }
+#endif
         atomic_flag_clear(&g_axpy_lock);
     }
 
@@ -14733,7 +14739,7 @@ static void ggml_compute_forward_mul_mat_axpy_q4_0(
 
         // 计算剩余的元素个数
         int remainder = ne00 % 8;
-
+#if defined(__AVX2__)
         // 使用AVX指令进行向量化计算
         for (i = 0; i < ne00 - remainder; i += 8)
         {
@@ -14748,6 +14754,11 @@ static void ggml_compute_forward_mul_mat_axpy_q4_0(
         {
             res[i] += tmp[i];
         }
+#else
+        for (i = 0; i < ne00; i++) {
+            res[i] += tmp[i];
+        }
+#endif
         atomic_flag_clear(&g_axpy_lock);
     }
 
@@ -14869,6 +14880,7 @@ static void ggml_compute_forward_mul_mat_axpy_head(
     // 计算剩余的元素个数
     int remainder = ne00 % 8;
 
+#if defined(__AVX2__)
     // 使用AVX指令进行向量化计算
     for (i = 0; i < ne00 - remainder; i += 8) {
         __m256 res_vec = _mm256_loadu_ps(res + i);  // 加载res中的8个浮点数
@@ -14881,10 +14893,11 @@ static void ggml_compute_forward_mul_mat_axpy_head(
     for (i = ne00 - remainder; i < ne00; i++) {
         res[i] += tmp[i];
     }
-    // for (i = 0; i < ne00; i++) {
-    //     res[i] = tmp[i];
-    // }
-
+#else
+    for (i = 0; i < ne00; i++) {
+        res[i] += tmp[i];
+    }
+#endif
     atomic_flag_clear(&g_axpy_head_lock);
 
 }

llama.cpp

@@ -2737,7 +2737,7 @@ struct llama_mlp_model_loader {
         offset = (offset + 31) & -32;
         file.seek(offset, SEEK_SET);
         // point to the mmaped mlp model file
-        mlp_tensor -> data = mapping -> addr + static_cast<std::streamoff>(offset);
+        mlp_tensor -> data = (void *) (static_cast<char *>(mapping -> addr) + offset);
         file.seek(tensor_data_size, SEEK_CUR);
         return mlp_tensor;
     }
@@ -2757,7 +2757,7 @@ struct llama_augmentation_model_loader {
         // const int64_t ggml_aux_tensor_size = 4 * (100 * 100 + 5120*40*4 * ggml_tensor_overhead() + (int64_t)13824*5120*40*4);
         int model_layer = model->layers.size();
         int ffn_dim = model->layers[0].ffn_up->ne[1];
-        const int64_t ggml_aux_tensor_size = 4 * (100 * 100 + model_layer*ffn_dim*sizeof(float) * ggml_tensor_overhead() );
+        const size_t ggml_aux_tensor_size = 4 * (100 * 100 + model_layer*ffn_dim*sizeof(float) * ggml_tensor_overhead() );
         printf("augmentation buffer: %ld\n", ggml_aux_tensor_size);
         struct ggml_init_params params = {
             /*.mem_size   =*/ ggml_aux_tensor_size,
@@ -2974,7 +2974,7 @@ static void llm_load_tensors(
     auto create_tensor = [&] (const std::string & name, const std::vector<int64_t> & ne, ggml_backend_type backend) -> ggml_tensor * {
         ggml_tensor * created_tensor = ml.create_tensor(ctx, name, ne, backend);
         if (created_tensor == nullptr) {
-            LLAMA_LOG_ERROR("%s: error: failed to create tensor '%s'\n", __func__, name);
+            LLAMA_LOG_ERROR("%s: error: failed to create tensor '%s'\n", __func__, name.c_str());
             return nullptr;
         }
         if (created_tensor->backend == GGML_BACKEND_GPU || created_tensor->backend == GGML_BACKEND_GPU_SPLIT) {