[AlvYuste]

Problem

loadTensorflowModel({ url: 'file://...' }) reads the whole model into memory before creating the interpreter.

• Android: URL(path).readBytes() copies the file into a ByteArrayOutputStream, which grows to about twice the model's size in the Java heap. A ~200 MB model (a Perch v2 audio classifier, in our case) throws OutOfMemoryError against the common 256 MB heap limit:

  java.lang.OutOfMemoryError: Failed to allocate a 203376824 byte allocation with 44346608 free bytes and 42MB until OOM, target footprint 268435456, growth limit 268435456
      at java.util.Arrays.copyOf(Arrays.java:4275)
      at java.io.ByteArrayOutputStream.toByteArray(ByteArrayOutputStream.java:211)
      at com.margelo.nitro.tflite.HybridAssetLoader$loadAsset$1.invokeSuspend(...)
  

• iOS: the model is copied twice, once into Data and once into the ArrayBuffer, so peak memory is about twice the model size.

Fix

Local file:// models are memory-mapped and wrapped in an ArrayBuffer without a copy:

• Android maps the file READ_ONLY and returns ArrayBuffer.wrap(mappedByteBuffer). TFLite only reads the model data. A copy-on-write mapping would need a writable channel. The mapping stays valid after the channel closes, and the wrapped direct buffer keeps it alive.
• iOS mmaps the file MAP_PRIVATE and returns ArrayBuffer.wrap(dataWithoutCopy:size:onDelete:), which munmaps it when the buffer is deleted. url.path also decodes percent-encoded paths such as Application%20Support.

http(s):// URLs and Android res/raw resources keep their current code paths.

Testing

I ran the same change as a patch on 3.0.1 in our app. 3.0.1's Android loader has no res/raw branch; otherwise the code is identical.

Device	Build	Before	After
Motorola edge 50 neo (256 MB heap limit)	Release, R8	OutOfMemoryError loading a 203 MB model	Loads; inference at about 830 ms per pass
iPhone 13 Pro Max	Debug dev client	Loads	Loads; inference at about 245 ms per pass

I also built this branch against main's example app: Android assembleDebug and the iOS Debug simulator build both pass, with no warnings from either loader. On Xcode 27.1 the iOS build needs IPHONEOS_DEPLOYMENT_TARGET=15.1 to get past the TensorFlowLiteC pods' 12.0 target, which is a problem on main too. The example's harness tests load models from Metro over HTTP, so they don't exercise the file:// path. Details.

🤖 Generated with Claude Code

[AlvYuste]

Built this branch (2f5e775) against main's example app:

• Android: ./gradlew assembleDebug --build-cache (same as build-android.yml): BUILD SUCCESSFUL, no warnings from HybridAssetLoader.kt.
• iOS: Debug iphonesimulator build of TfliteExample (same flags as build-ios.yml): BUILD SUCCEEDED, no warnings from HybridAssetLoader.swift.

Two things on main, unrelated to this change:

• With Xcode 27.1, the iOS build fails before compiling anything, because the TensorFlowLiteC pod targets have IPHONEOS_DEPLOYMENT_TARGET = 12.0 and the supported minimum is 15.0. I passed IPHONEOS_DEPLOYMENT_TARGET=15.1 on the xcodebuild command line to get past it.
• example/ios/Podfile.lock is stale since chore: upgrade react-native-nitro-modules #204: it still pins NitroModules 0.35.2 and react-native-fast-tflite 3.0.0, so pod install rewrites it.

The example's harness tests load models through require(), which in debug resolves to a Metro HTTP URL, so they don't reach the new file:// mmap path. That path was exercised on physical devices with a ~200 MB model: a Motorola edge 50 neo (256 MB heap limit, where readBytes() OOMs before this change) and an iPhone 13 Pro Max.

[DaveyEke]

Hey @AlvYuste! Many thanks for your PR. It contains some really nice improvements. I opened a follow up PR cleaning this one up.

See it here: #210