cui-llama.rn 1.7.2 → 1.7.4

package/android/src/main/CMakeLists.txt

@@ -90,7 +90,7 @@ function(build_library target_name cpu_flags)
 
     target_link_libraries(${target_name} ${LOG_LIB} android)
 
-    target_compile_options(${target_name} PRIVATE -DLM_GGML_USE_CPU -DLM_GGML_USE_CPU_AARCH64 -pthread ${cpu_flags})
+    target_compile_options(${target_name} PRIVATE -DLM_GGML_USE_CPU -DLM_GGML_USE_CPU_AARCH64  -DRNLLAMA_USE_FD_FILE -pthread ${cpu_flags})
 
     if (${CMAKE_BUILD_TYPE} STREQUAL "Debug")
         target_compile_options(${target_name} PRIVATE -DRNLLAMA_ANDROID_ENABLE_LOGGING)
@@ -99,7 +99,7 @@ function(build_library target_name cpu_flags)
     # NOTE: If you want to debug the native code, you can uncomment if and endif
     # Note that it will be extremely slow
     # if (NOT ${CMAKE_BUILD_TYPE} STREQUAL "Debug")
-        target_compile_options(${target_name} PRIVATE -O3 -DNDEBUG -DRNLLAMA_USE_FD_FILE)
+        target_compile_options(${target_name} PRIVATE -O3 -DNDEBUG)
         target_compile_options(${target_name} PRIVATE -fvisibility=hidden -fvisibility-inlines-hidden)
         target_compile_options(${target_name} PRIVATE -ffunction-sections -fdata-sections)
 

package/android/src/main/java/com/rnllama/LlamaContext.java

@@ -69,7 +69,11 @@ public class LlamaContext {
     try {
       if (filepath.startsWith("content")) {
         Uri uri = Uri.parse(filepath);
-        reactContext.getApplicationContext().getContentResolver().takePersistableUriPermission(uri, Intent.FLAG_GRANT_READ_URI_PERMISSION);
+        try {
+          reactContext.getApplicationContext().getContentResolver().takePersistableUriPermission(uri, Intent.FLAG_GRANT_READ_URI_PERMISSION);
+        } catch (SecurityException e) {
+            Log.w(NAME, "Persistable permission not granted for URI: " + uri);
+        }
         fis = reactContext.getApplicationContext().getContentResolver().openInputStream(uri);
       } else {
         fis = new FileInputStream(filepath);
@@ -107,7 +111,11 @@ public class LlamaContext {
     }
 
     String modelName = params.getString("model");
-    
+
+    if(!isGGUF(modelName, reactContext)) {
+      throw new IllegalArgumentException("File is not in GGUF format");
+    }
+
     if (modelName.startsWith("content://")) {
       Uri uri = Uri.parse(modelName);
       try {
@@ -117,7 +125,6 @@ public class LlamaContext {
         Log.e(NAME, "Failed to convert to FD!");
       }
     }
-    
 
     // Check if file has GGUF magic numbers
     this.id = id;
@@ -442,10 +449,26 @@ public class LlamaContext {
     if (mmprojPath == null || mmprojPath.isEmpty()) {
       throw new IllegalArgumentException("mmproj_path is empty");
     }
+
+    if(!isGGUF(mmprojPath, this.reactContext)) {
+      throw new IllegalArgumentException("File is not in GGUF format");
+    }
+
     File file = new File(mmprojPath);
-    if (!file.exists()) {
+    if (!mmprojPath.startsWith("content") && !file.exists()) {
       throw new IllegalArgumentException("mmproj file does not exist: " + mmprojPath);
     }
+
+    if (mmprojPath.startsWith("content://")) {
+      Uri uri = Uri.parse(mmprojPath);
+      try {
+        ParcelFileDescriptor pfd = this.reactContext.getApplicationContext().getContentResolver().openFileDescriptor(uri, "r");
+        mmprojPath =  "" + pfd.getFd();
+      } catch (Exception e) {
+        Log.e(NAME, "Failed to convert to FD!");
+      }
+    }
+
     return initMultimodal(this.context, mmprojPath, mmprojUseGpu);
   }
 

package/android/src/main/java/com/rnllama/RNLlama.java

@@ -7,6 +7,7 @@ import android.os.Handler;
 import android.os.AsyncTask;
 import android.os.ParcelFileDescriptor;
 import android.net.Uri;
+import android.content.Intent;
 
 import com.facebook.react.bridge.Promise;
 import com.facebook.react.bridge.ReactApplicationContext;
@@ -85,6 +86,15 @@ public class RNLlama implements LifecycleEventListener {
 
   public void modelInfo(final String model, final ReadableArray skip, final Promise promise) {
     final String modelPath = getContentFileDescriptor(model);
+    
+    if (model.startsWith("content")) {
+        Uri uri = Uri.parse(model);
+        try {
+          reactContext.getApplicationContext().getContentResolver().takePersistableUriPermission(uri, Intent.FLAG_GRANT_READ_URI_PERMISSION);
+        } catch (SecurityException e) {
+            Log.w(NAME, "Persistable permission not granted for URI: " + uri);
+        }
+    }
 
     new AsyncTask<Void, Void, WritableMap>() {
       private Exception exception;

package/cpp/tools/mtmd/clip.cpp

@@ -31,6 +31,12 @@
 #include <numeric>
 #include <functional>
 
+// rnllama additions
+#include <fcntl.h>
+#include <unistd.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+
 struct clip_logger_state g_logger_state = {LM_GGML_LOG_LEVEL_CONT, clip_log_callback_default, NULL};
 
 enum ffn_op_type {
@@ -2486,6 +2492,55 @@ struct clip_model_loader {
         {
             std::vector<uint8_t> read_buf;
 
+            // rnllama addition - we support usage of file descriptors
+            // Check if fname is an FD number (no '/' characters)
+            bool is_fd = (fname.find('/') == std::string::npos);
+
+            if (is_fd) {
+                // Routine for handling FD
+                int fd = -1;
+                try {
+                    fd = std::stoi(fname); // Convert string to integer FD
+                } catch (const std::invalid_argument& e) {
+                    throw std::runtime_error(string_format("%s: invalid FD number provided: %s\n", __func__, fname.c_str()));
+                } catch (const std::out_of_range& e) {
+                    throw std::runtime_error(string_format("%s: FD number out of range: %s\n", __func__, fname.c_str()));
+                }
+
+                lm_ggml_backend_buffer_type_t buft = lm_ggml_backend_get_default_buffer_type(ctx_clip.backend);
+                ctx_clip.buf.reset(lm_ggml_backend_alloc_ctx_tensors_from_buft(ctx_clip.ctx_data.get(), buft));
+                lm_ggml_backend_buffer_set_usage(ctx_clip.buf.get(), LM_GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
+
+                for (auto & t : tensors_to_load) {
+                    lm_ggml_tensor * cur = lm_ggml_get_tensor(ctx_clip.ctx_data.get(), t->name);
+                    const size_t offset = tensor_offset[t->name];
+
+                    if (lseek(fd, offset, SEEK_SET) == (off_t)-1) {
+                        throw std::runtime_error(string_format("%s: failed to seek for tensor %s (FD %d): %s\n", __func__, t->name, fd, strerror(errno)));
+                    }
+
+                    size_t num_bytes = lm_ggml_nbytes(cur);
+                    if (lm_ggml_backend_buft_is_host(buft)) {
+                        // for the CPU and Metal backend, we can read directly into the tensor
+                        ssize_t bytes_read = read(fd, reinterpret_cast<char *>(cur->data), num_bytes);
+                        if (bytes_read == -1 || static_cast<size_t>(bytes_read) != num_bytes) {
+                            throw std::runtime_error(string_format("%s: failed to read for tensor %s (FD %d): %s\n", __func__, t->name, fd, strerror(errno)));
+                        }
+                    } else {
+                        // read into a temporary buffer first, then copy to device memory
+                        read_buf.resize(num_bytes);
+                        ssize_t bytes_read = read(fd, reinterpret_cast<char *>(read_buf.data()), num_bytes);
+                        if (bytes_read == -1 || static_cast<size_t>(bytes_read) != num_bytes) {
+                            throw std::runtime_error(string_format("%s: failed to read for tensor %s (FD %d): %s\n", __func__, t->name, fd, strerror(errno)));
+                        }
+                        lm_ggml_backend_tensor_set(cur, read_buf.data(), 0, num_bytes);
+                    }
+                }
+                // Assuming the FD is managed externally and shouldn't be closed here.
+                LOG_DBG("%s: loaded %zu tensors from FD %s\n", __func__, tensors_to_load.size(), fname.c_str());
+
+            } else {
+            // The original ifstream routine for file paths
             auto fin = std::ifstream(fname, std::ios::binary);
             if (!fin) {
                 throw std::runtime_error(string_format("%s: failed to open %s\n", __func__, fname.c_str()));
@@ -2516,6 +2571,7 @@ struct clip_model_loader {
             fin.close();
 
             LOG_DBG("%s: loaded %zu tensors from %s\n", __func__, tensors_to_load.size(), fname.c_str());
+            }
         }
     }
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "cui-llama.rn",
-  "version": "1.7.2",
+  "version": "1.7.4",
   "description": "Fork of llama.rn for ChatterUI",
   "main": "lib/commonjs/index",
   "module": "lib/module/index",