From 192c9ae24824652efb171d5d6a97e331c487d640 Mon Sep 17 00:00:00 2001
From: Jhen-Jie Hong <iainst0409@gmail.com>
Date: Sat, 2 Nov 2024 18:43:48 +0800
Subject: [PATCH] fix(cpp): skip register metal backend on os simulator

---
 cpp/common.cpp                 |  2 +-
 cpp/ggml-backend.cpp           |  3 +++
 cpp/llama.cpp                  | 10 +---------
 example/ios/Podfile.lock       |  4 ++--
 scripts/bootstrap.sh           |  1 +
 scripts/ggml-backend.cpp.patch | 14 ++++++++++++++
 scripts/llama.cpp.patch        | 25 -------------------------
 7 files changed, 22 insertions(+), 37 deletions(-)
 create mode 100644 scripts/ggml-backend.cpp.patch

diff --git a/cpp/common.cpp b/cpp/common.cpp
index 347dd0f..a7e4a46 100644
--- a/cpp/common.cpp
+++ b/cpp/common.cpp
@@ -985,7 +985,7 @@ struct llama_model_params common_model_params_to_llama(const common_params & par
     if (params.n_gpu_layers != -1) {
         mparams.n_gpu_layers = params.n_gpu_layers;
     }
-    
+
     mparams.vocab_only      = params.vocab_only;
     mparams.rpc_servers     = params.rpc_servers.c_str();
     mparams.main_gpu        = params.main_gpu;
diff --git a/cpp/ggml-backend.cpp b/cpp/ggml-backend.cpp
index 7451de3..d8cb76d 100644
--- a/cpp/ggml-backend.cpp
+++ b/cpp/ggml-backend.cpp
@@ -575,8 +575,11 @@ struct lm_ggml_backend_registry {
         register_backend(lm_ggml_backend_cuda_reg());
 #endif
 #ifdef LM_GGML_USE_METAL
+#include <TargetConditionals.h>
+#if !TARGET_OS_SIMULATOR
         register_backend(lm_ggml_backend_metal_reg());
 #endif
+#endif
 #ifdef LM_GGML_USE_SYCL
         register_backend(lm_ggml_backend_sycl_reg());
 #endif
diff --git a/cpp/llama.cpp b/cpp/llama.cpp
index d0c4f5d..6296abe 100644
--- a/cpp/llama.cpp
+++ b/cpp/llama.cpp
@@ -19097,9 +19097,7 @@ struct llama_model_params llama_model_default_params() {
 
 #ifdef LM_GGML_USE_METAL
     // note: we usually have plenty of VRAM, so by default offload all layers to the GPU
-    if (result.n_gpu_layers > 0) {
-        result.n_gpu_layers = 999;
-    }
+    result.n_gpu_layers = 999;
 #endif
 
     return result;
@@ -19302,13 +19300,7 @@ struct llama_model * llama_load_model_from_file(
                 break;
 
             case LM_GGML_BACKEND_DEVICE_TYPE_GPU:
-#ifdef LM_GGML_USE_METAL
-                if (params.n_gpu_layers > 0) {
-                    model->devices.push_back(dev);
-                }
-#else
                 model->devices.push_back(dev);
-#endif
                 break;
         }
     }
diff --git a/example/ios/Podfile.lock b/example/ios/Podfile.lock
index f9cf151..1827ea6 100644
--- a/example/ios/Podfile.lock
+++ b/example/ios/Podfile.lock
@@ -8,7 +8,7 @@ PODS:
     - hermes-engine/Pre-built (= 0.72.3)
   - hermes-engine/Pre-built (0.72.3)
   - libevent (2.1.12)
-  - llama-rn (0.3.10):
+  - llama-rn (0.3.11):
     - RCT-Folly
     - RCTRequired
     - RCTTypeSafety
@@ -1261,7 +1261,7 @@ SPEC CHECKSUMS:
   glog: 04b94705f318337d7ead9e6d17c019bd9b1f6b1b
   hermes-engine: 10fbd3f62405c41ea07e71973ea61e1878d07322
   libevent: 4049cae6c81cdb3654a443be001fb9bdceff7913
-  llama-rn: bb4447eb237c895e526b1adbb2b88700093005ac
+  llama-rn: 07a172eaad49cc56323f326e4c3ce3ec734b6a9e
   RCT-Folly: 424b8c9a7a0b9ab2886ffe9c3b041ef628fd4fb1
   RCTRequired: a2faf4bad4e438ca37b2040cb8f7799baa065c18
   RCTTypeSafety: cb09f3e4747b6d18331a15eb05271de7441ca0b3
diff --git a/scripts/bootstrap.sh b/scripts/bootstrap.sh
index 92521e9..b84fe66 100755
--- a/scripts/bootstrap.sh
+++ b/scripts/bootstrap.sh
@@ -119,6 +119,7 @@ patch -p0 -d ./cpp < ./scripts/common.cpp.patch
 patch -p0 -d ./cpp < ./scripts/log.cpp.patch
 patch -p0 -d ./cpp < ./scripts/llama.cpp.patch
 patch -p0 -d ./cpp < ./scripts/ggml-metal.m.patch
+patch -p0 -d ./cpp < ./scripts/ggml-backend.cpp.patch
 patch -p0 -d ./cpp < ./scripts/ggml.c.patch
 
 
diff --git a/scripts/ggml-backend.cpp.patch b/scripts/ggml-backend.cpp.patch
new file mode 100644
index 0000000..48757f4
--- /dev/null
+++ b/scripts/ggml-backend.cpp.patch
@@ -0,0 +1,14 @@
+--- ggml-backend.cpp.orig	2024-11-02 18:37:57
++++ ggml-backend.cpp	2024-11-02 18:39:36
+@@ -575,8 +575,11 @@
+         register_backend(lm_ggml_backend_cuda_reg());
+ #endif
+ #ifdef LM_GGML_USE_METAL
++#include <TargetConditionals.h>
++#if !TARGET_OS_SIMULATOR
+         register_backend(lm_ggml_backend_metal_reg());
+ #endif
++#endif
+ #ifdef LM_GGML_USE_SYCL
+         register_backend(lm_ggml_backend_sycl_reg());
+ #endif
diff --git a/scripts/llama.cpp.patch b/scripts/llama.cpp.patch
index a09f90e..efed972 100644
--- a/scripts/llama.cpp.patch
+++ b/scripts/llama.cpp.patch
@@ -39,28 +39,3 @@
                          strerror(errno));
              }
          }
-@@ -19086,7 +19097,9 @@
-
- #ifdef LM_GGML_USE_METAL
-     // note: we usually have plenty of VRAM, so by default offload all layers to the GPU
--    result.n_gpu_layers = 999;
-+    if (result.n_gpu_layers > 0) {
-+        result.n_gpu_layers = 999;
-+    }
- #endif
-
-     return result;
-@@ -19289,7 +19302,13 @@
-                 break;
-
-             case LM_GGML_BACKEND_DEVICE_TYPE_GPU:
-+#ifdef LM_GGML_USE_METAL
-+                if (params.n_gpu_layers > 0) {
-+                    model->devices.push_back(dev);
-+                }
-+#else
-                 model->devices.push_back(dev);
-+#endif
-                 break;
-         }
-     }