MLX encode allocates ~0.9 GB on iOS for a program that peaks at 208 MB on macOS

🐛 Describe the bug

The same Whisper-small MLX .pte uses four to six times more memory on an iPhone than on a Mac, on the same ExecuTorch version and the same delegate. The gap is entirely in the encode method, and it is not precision specific.

Measurements

Whisper-small (multilingual), exported with ExecuTorch 1.4.1, two methods (encode, decode). The audio input is a dynamic dim resolved to its maximum, 480000 samples, in both harnesses.

macOS (executor_runner, whole-process peak RSS, one method per process), bf16:

Flat at 1, 5 and 20 executions, so nothing accumulates per run.

iPhone 16, iOS 26.6.1, Release build, process footprint sampled across each method's execute loop:

What this rules out

• Not a bf16 problem. int8 pays +680 MB for the same method. bf16's excess over int8 is about 250 MB, roughly its extra weight bytes (the two files differ by 195 MB).
• Not per-execution accumulation. The peak is flat in iteration count on both platforms.
• Not decode. It is free on device to within 0.06 MB.
• Not the exported graph. The identical file is the one that peaks at 208 MB on macOS.

That leaves the iOS MLX runtime's allocation behaviour for this method as the only remaining difference.

Caveats

macOS numbers are whole-process peak RSS under a bare runner; device numbers are footprint deltas inside a React Native app, so the two are not measured identically. Even comparing totals rather than deltas the gap stands (208 MB against 1372 MB). macOS Metal and iOS Metal also have different buffer residency behaviour, which may be the whole answer.

Practical impact

This is what makes the model unusable at the larger sizes. The published small bf16 build was withdrawn after being jetsam-killed at 3.80 GB on this device, though I could not reproduce that specific figure here: my runs peaked at 1.2 to 1.4 GB and survived.

Versions

• ExecuTorch 1.4.1
• Device: iPhone 16 (iPhone17,3), iOS 26.6.1, Release build
• Backend: MLXPartitioner with get_default_passes()
• Model: OpenAI Whisper small, compute dtype bf16 and int8, two-method program

Supersedes #22016, which framed this as a bf16-specific leak. That framing is wrong: both precisions pay the same encode cost.

cc @shoumikhin @cbilgin @metascroy