Merge pull request #41 from freelw/wangli_dev_20250617_2

#fix 35

Author: freelw

backends/gpu/metal/metal_ops.cpp

@@ -35,7 +35,15 @@ MetalOps::MetalOps() : commandBuffer(nullptr), cur_int_args(0), cur_float_args(0
     }
 
     bufferIntArgs = device->newBuffer(TOTAL_INT_ARGS * sizeof(int), MTL::ResourceStorageModeShared);
+    if (!bufferIntArgs) {
+        std::cerr << "Failed to create buffer for int args!" << std::endl;
+        throw std::runtime_error("Failed to create buffer for int args");
+    }
     bufferFloatArgs = device->newBuffer(TOTAL_FLOAT_ARGS * sizeof(float), MTL::ResourceStorageModeShared);
+    if (!bufferFloatArgs) {
+        std::cerr << "Failed to create buffer for float args!" << std::endl;
+        throw std::runtime_error("Failed to create buffer for float args");
+    }
     load_kernel_metal();
 
     addOps = new MetalKops("tensor_add_kernel", library);
@@ -115,10 +123,10 @@ void MetalOps::prepare() {
     cur_float_args = 0;
 }
 
-int calc_offset(const Tensor* t) {
+unsigned int calc_offset(const Tensor* t) {
     char* base = reinterpret_cast<char*>(reinterpret_cast<MTL::Buffer*>(t->get_storage()->ctx)->contents());
     char* pos = reinterpret_cast<char*>(t->get_data());
-    auto offset_res = pos - base;
+    unsigned int offset_res = pos - base;
     return offset_res;
 }
 
@@ -1428,6 +1436,10 @@ void MetalOps::mulSV(Tensor* dst, Tensor* src, float value) {
 
 void* MetalOps::alloc(size_t size, void** ctx) {
     MTL::Buffer* buffer = device->newBuffer(size, MTL::ResourceStorageModeShared);
+    if (!buffer) {
+        std::cerr << "Error allocating buffer of size " << size << std::endl;
+        abort();
+    }
     *ctx = (void*)buffer;
     return buffer->contents();
 }

lm.cpp

@@ -41,7 +41,8 @@ std::vector<uint> trim_or_padding(const std::vector<uint>& src, uint max_len, ui
     std::vector<uint> res = src;
     if (src.size() > max_len) {
         res.resize(max_len);
-    } else {
+    }
+    else {
         res.resize(max_len, pad_id);
     }
     return res;
@@ -98,7 +99,7 @@ int main(int argc, char* argv[]) {
 
     int opt;
     int epochs = 10;
-    int batch_size = 4;
+    int batch_size = 16;
     int gpu = 1;
     int max_words_cnt = 256;
     float lr = 0.001f;
@@ -225,6 +226,7 @@ int main(int argc, char* argv[]) {
     adam.clip_grad(1.0f);
     adam.step();
     graph::validateAllNodesRefCnt(0);
+    // printAllTensors();
     // printAllActions();
     allocMemAndInitTensors();
     std::cout << "Allocating memory  " << std::endl
@@ -264,7 +266,8 @@ int main(int argc, char* argv[]) {
             auto origin_size = src_token_ids.size();
             if (src_token_ids.size() < num_steps) {
                 src_token_ids.resize(num_steps, loader.get_pad_id());
-            } else if (src_token_ids.size() > num_steps) {
+            }
+            else if (src_token_ids.size() > num_steps) {
                 src_token_ids.erase(src_token_ids.begin(), src_token_ids.end() - num_steps);
             }
             auto cur_step = origin_size - 1;
@@ -302,15 +305,17 @@ int main(int argc, char* argv[]) {
                 if (cur_step >= num_steps - 1) {
                     src_token_ids.push_back(max_index);
                     src_token_ids.erase(src_token_ids.begin(), src_token_ids.end() - num_steps);
-                } else {
+                }
+                else {
                     src_token_ids[++cur_step] = max_index;
                 }
             }
             std::cout << std::endl;
             std::cout << "-----------------" << std::endl;
             ::free(res_buffer);
         }
-    } else {
+    }
+    else {
         init_dec_valid_lens_for_training(dec_valid_lens);
         signal(SIGINT, signal_callback_handler);
         int epoch = 0;