metal : create autorelease pool during library build (#4970)

* metal : create autorelease pool during library build

ggml-ci

* test : simplify

ggml-ci

.gitignore

@@ -105,3 +105,4 @@ poetry.toml
 /tests/test-tokenizer-1-bpe
 /tests/test-rope
 /tests/test-backend-ops
+/tests/test-autorelease

Makefile

@@ -9,7 +9,7 @@ TEST_TARGETS = \
 	tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt \
 	tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama          \
 	tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama tests/test-tokenizer-1-bpe tests/test-rope      \
-	tests/test-backend-ops
+	tests/test-backend-ops tests/test-autorelease
 
 # Code coverage output files
 COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report
@@ -747,3 +747,6 @@ tests/test-c.o: tests/test-c.c llama.h
 
 tests/test-backend-ops: tests/test-backend-ops.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+tests/test-autorelease: tests/test-autorelease.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

ci/run.sh

@@ -179,6 +179,8 @@ function gg_run_open_llama_3b_v2 {
 
     wiki_test_60="${path_wiki}/wiki.test-60.raw"
 
+    ./bin/test-autorelease ${model_f16}
+
     ./bin/quantize ${model_f16} ${model_q8_0} q8_0
     ./bin/quantize ${model_f16} ${model_q4_0} q4_0
     ./bin/quantize ${model_f16} ${model_q4_1} q4_1

ggml-metal.m

@@ -303,22 +303,21 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
                 return NULL;
             }
 
-            // dictionary of preprocessor macros
-            NSMutableDictionary * prep = [NSMutableDictionary dictionary];
+            @autoreleasepool {
+                // dictionary of preprocessor macros
+                NSMutableDictionary * prep = [NSMutableDictionary dictionary];
 
 #ifdef GGML_QKK_64
-            prep[@"QK_K"] = @(64);
+                prep[@"QK_K"] = @(64);
 #endif
 
-            MTLCompileOptions* options = [MTLCompileOptions new];
-            options.preprocessorMacros = prep;
+                MTLCompileOptions* options = [MTLCompileOptions new];
+                options.preprocessorMacros = prep;
 
-            //[options setFastMathEnabled:false];
+                //[options setFastMathEnabled:false];
 
-            ctx->library = [ctx->device newLibraryWithSource:src options:options error:&error];
-
-            [options release];
-            [prep release];
+                ctx->library = [ctx->device newLibraryWithSource:src options:options error:&error];
+            }
         }
 
         if (error) {

tests/CMakeLists.txt

@@ -49,6 +49,7 @@ llama_build_and_test_executable(test-llama-grammar.cpp)
 llama_build_and_test_executable(test-grad0.cpp)
 # llama_build_and_test_executable(test-opt.cpp) # SLOW
 llama_build_and_test_executable(test-backend-ops.cpp)
+llama_build_and_test_executable(test-autorelease.cpp)
 
 llama_build_and_test_executable(test-rope.cpp)
 

tests/test-autorelease.cpp

@@ -0,0 +1,28 @@
+// ref: https://github.com/ggerganov/llama.cpp/issues/4952#issuecomment-1892864763
+
+#include <cstdio>
+#include <string>
+#include <thread>
+
+#include "llama.h"
+
+// This creates a new context inside a pthread and then tries to exit cleanly.
+int main(int argc, char ** argv) {
+    if (argc < 2) {
+        printf("Usage: %s model.gguf\n", argv[0]);
+        return 0; // intentionally return success
+    }
+
+    const std::string fname = argv[1];
+
+    std::thread([&fname]() {
+        llama_backend_init(false);
+        auto * model = llama_load_model_from_file(fname.c_str(), llama_model_default_params());
+        auto * ctx = llama_new_context_with_model(model, llama_context_default_params());
+        llama_free(ctx);
+        llama_free_model(model);
+        llama_backend_free();
+    }).join();
+
+    return 0;
+}