From 6bbb72e7880a01cf4ae33a85f6d0e9f4e5c45c11 Mon Sep 17 00:00:00 2001
From: Animesh Kumar <kumaranimesh097@gmail.com>
Date: Sun, 7 May 2023 08:04:32 +0530
Subject: [PATCH] make the desired changes as said here ->
 https://github.com/X-PLUG/mPLUG-Owl/issues/22#issue-1698528874

---
 apex_22.01_pp/csrc/mlp.cpp | 4 ++--
 interface.py               | 2 +-
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/apex_22.01_pp/csrc/mlp.cpp b/apex_22.01_pp/csrc/mlp.cpp
index 68cf174..830d606 100644
--- a/apex_22.01_pp/csrc/mlp.cpp
+++ b/apex_22.01_pp/csrc/mlp.cpp
@@ -62,7 +62,7 @@ std::vector<at::Tensor> mlp_forward(int use_bias, int activation, std::vector<at
 
   // create output/workspace tensor
   auto out = at::empty({batch_size, output_features.back()}, inputs[0].type());
-  auto reserved_space = at::empty({reserved_size}, inputs[0].type());
+  auto reserved_space = at::empty({static_cast<long>(reserved_size)}, inputs[0].type());
   // allocate fixed 4MB workspace for cublaslt for now, and this gets at least 4 MB
   auto lt_workspace = at::empty({1 << 22}, inputs[0].type());
 
@@ -135,7 +135,7 @@ std::vector<at::Tensor> mlp_backward(
         get_mlp_bp_workspace_in_bytes<scalar_t>(batch_size, num_layers, output_features.data());
 
     // auto work_space = at::empty({work_size*4}, at::kByte);
-    auto work_space = at::empty({work_size / sizeof(scalar_t)}, inputs[0].type());
+    auto work_space = at::empty({static_cast<long>(work_size / sizeof(scalar_t))}, inputs[0].type());
 
     auto result = mlp_bp<scalar_t>(
         inputs[0].data_ptr<scalar_t>(),
diff --git a/interface.py b/interface.py
index 8ecb5f8..fdfb7c9 100644
--- a/interface.py
+++ b/interface.py
@@ -30,8 +30,8 @@ def get_model(checkpoint_path=None, tokenizer_path=None, peft_config=None, devic
     tokenizer.eod_id = tokenizer.eos_token_id
     img_processor = ImageProcessor()
 
-    model = model.to(device)
     model = model.to(dtype)
+    model = model.to(device)
     return model, tokenizer, img_processor