Merge branch 'inference' into update_register_interface

flexflow · Sep 25, 2024 · d7deb89 · d7deb89
2 parents 6ba0304 + 70e47b2
commit d7deb89
Show file tree

Hide file tree

Showing 3 changed files with 40 additions and 18 deletions.
diff --git a/src/ops/residual_layer_norm.cpp b/src/ops/residual_layer_norm.cpp
@@ -176,6 +176,8 @@ void ResidualLayerNorm::inference_kernel(ResidualLayerNormMeta const *m,
                      beta_ptr,
                      output_ptr);
 }
+
+#ifdef DEADCODE
 template <typename T>
 void save_inference_tensors(ResidualLayerNormMeta const *m) {
   if (m->inference_debugging) {
@@ -206,6 +208,7 @@ void save_inference_tensors(ResidualLayerNormMeta const *m) {
                 filename3.c_str());
   }
 }
+#endif
 
 /*static*/
 void ResidualLayerNorm::inference_kernel_wrapper(
@@ -314,15 +317,15 @@ void ResidualLayerNorm::inference_kernel_wrapper(
     }
   }
 
-  if (m->inference_debugging) {
-    if (m->input_type[0] == DT_FLOAT) {
-      save_inference_tensors<float>(m);
-    } else if (m->input_type[0] == DT_HALF) {
-      save_inference_tensors<half>(m);
-    } else {
-      assert(false && "unsupport datatype in layernorm");
-    }
-  }
+  // if (m->inference_debugging) {
+  //   if (m->input_type[0] == DT_FLOAT) {
+  //     save_inference_tensors<float>(m);
+  //   } else if (m->input_type[0] == DT_HALF) {
+  //     save_inference_tensors<half>(m);
+  //   } else {
+  //     assert(false && "unsupport datatype in layernorm");
+  //   }
+  // }
 
   if (m->profiling) {
     checkCUDA(hipEventRecord(t_end, stream));

diff --git a/src/ops/residual_layer_norm.cu b/src/ops/residual_layer_norm.cu
@@ -174,6 +174,8 @@ void ResidualLayerNorm::inference_kernel(ResidualLayerNormMeta const *m,
                    beta_ptr,
                    output_ptr);
 }
+
+#ifdef DEADCODE
 template <typename T>
 void save_inference_tensors(ResidualLayerNormMeta const *m) {
   if (m->inference_debugging) {
@@ -204,6 +206,7 @@ void save_inference_tensors(ResidualLayerNormMeta const *m) {
                 filename3.c_str());
   }
 }
+#endif
 
 /*static*/
 void ResidualLayerNorm::inference_kernel_wrapper(
@@ -312,15 +315,15 @@ void ResidualLayerNorm::inference_kernel_wrapper(
     }
   }
 
-  if (m->inference_debugging) {
-    if (m->input_type[0] == DT_FLOAT) {
-      save_inference_tensors<float>(m);
-    } else if (m->input_type[0] == DT_HALF) {
-      save_inference_tensors<half>(m);
-    } else {
-      assert(false && "unsupport datatype in layernorm");
-    }
-  }
+  // if (m->inference_debugging) {
+  //   if (m->input_type[0] == DT_FLOAT) {
+  //     save_inference_tensors<float>(m);
+  //   } else if (m->input_type[0] == DT_HALF) {
+  //     save_inference_tensors<half>(m);
+  //   } else {
+  //     assert(false && "unsupport datatype in layernorm");
+  //   }
+  // }
 
   if (m->profiling) {
     cudaEventRecord(t_end, stream);

diff --git a/src/runtime/cuda_helper.cu b/src/runtime/cuda_helper.cu
@@ -278,6 +278,10 @@ __host__ void
       host_ptr, ptr, sizeof(float) * num_elements, cudaMemcpyDeviceToHost));
   FILE *tensor_file;
   tensor_file = fopen(file_name, "w");
+  if (!tensor_file) {
+    fprintf(stderr, "Error %i creating file %s\n", errno, file_name);
+    assert(false);
+  }
   assert(tensor_file != NULL);
   for (unsigned i = 0; i < num_elements; i++) {
     if (i < num_elements - 1) {
@@ -299,6 +303,10 @@ __host__ void
       host_ptr, ptr, sizeof(half) * num_elements, cudaMemcpyDeviceToHost));
   FILE *tensor_file;
   tensor_file = fopen(file_name, "w");
+  if (!tensor_file) {
+    fprintf(stderr, "Error %i creating file %s\n", errno, file_name);
+    assert(false);
+  }
   assert(tensor_file != NULL);
   for (unsigned i = 0; i < num_elements; i++) {
     if (i < num_elements - 1) {
@@ -321,6 +329,10 @@ __host__ void save_tensor(int32_t const *ptr,
       host_ptr, ptr, sizeof(int32_t) * num_elements, cudaMemcpyDeviceToHost));
   FILE *tensor_file;
   tensor_file = fopen(file_name, "w");
+  if (!tensor_file) {
+    fprintf(stderr, "Error %i creating file %s\n", errno, file_name);
+    assert(false);
+  }
   assert(tensor_file != NULL);
   for (unsigned i = 0; i < num_elements; i++) {
     if (i < num_elements - 1) {
@@ -343,6 +355,10 @@ __host__ void save_tensor(int64_t const *ptr,
       host_ptr, ptr, sizeof(int64_t) * num_elements, cudaMemcpyDeviceToHost));
   FILE *tensor_file;
   tensor_file = fopen(file_name, "w");
+  if (!tensor_file) {
+    fprintf(stderr, "Error %i creating file %s\n", errno, file_name);
+    assert(false);
+  }
   assert(tensor_file != NULL);
   for (unsigned i = 0; i < num_elements; i++) {
     if (i < num_elements - 1) {