backends/coreml: LFM2.5 1.2B PTE compiles to mlmodelc but ANE init fails (ANECCompile FAILED)

[msluszniak]

Problem

After applying the workaround in #19634, an LFM2.5 1.2B CoreML PTE loads via executorch.runtime.Runtime.load_program(...) and all metadata methods (get_eos_ids, get_max_seq_len, use_kv_cache, …) succeed. prog.load_method("forward") then fails:

[ETCoreMLModelManager.mm:495] Successfully got compiled model ...
[ETCoreMLModelAnalyzer.mm:68] [Core ML] Failed to create model profiler.
    Failed to build the model execution plan using a model architecture file '.../model.mil'
[coreml_backend_delegate.mm:324] CoreMLBackend: Failed to init the model.
[method.cpp:114] Init failed for backend CoreMLBackend: 0x23
E5RT encountered an STL exception. msg = MILCompilerForANE error:
    failed to compile ANE model using ANEF. Error=_ANECompiler : ANECCompile() FAILED.

MIL → mlmodelc compilation succeeds; the ANE-specific execution-plan build fails. Reproduces via executorch.runtime on macOS and on iPhone 17 Pro / iOS 26.4.2 (surfaces in react-native-executorch as code: 35 "Failed to load LLM runner"), so it's a CoreML/ANE-side issue rather than a runtime one. compute_units: cpu_only and cpu_and_gpu succeed, but XNNPACK already covers the CPU case at higher throughput — the value of CoreML for this model is the ANE.

Reproduces with two different quantisation modes, ruling out a quantiser-specific cause:

1. unquantised fp16 (no quantization: block)
2. weight-only 4-bit via the documented torchao quantize_ path (qmode: 4w, see docs/source/backends/coreml/coreml-quantization.md)

So the failure is in lowering LFM2's short-conv conv_state mutation (self.conv_state.copy_(new_state) in examples/models/lfm2/short_conv.py, which decomposes to slice_copy + index_put) to an ANE-compatible MIL representation. The same model graph works on cpu_and_gpu.

Reproduce

# Apply workaround from #19634 first.

cat > examples/models/lfm2/config/lfm2_coreml_4w.yaml <<'EOF'
base:
  metadata: '{"get_bos_id": 1, "get_eos_ids":[7]}'
model:
  use_kv_cache: True
  enable_dynamic_shape: False
  dtype_override: fp32
quantization:
  qmode: 4w
  group_size: 32
backend:
  coreml:
    enabled: True
    ios: 18
    enable_state: True
    preserve_sdpa: True
    compute_units: cpu_and_ne
EOF

python -m extension.llm.export.export_llm \
  --config examples/models/lfm2/config/lfm2_coreml_4w.yaml \
  +base.model_class=lfm2_5_1_2b \
  +base.params=examples/models/lfm2/config/lfm2_5_1_2b_config.json \
  +export.max_seq_length=2048 \
  +export.max_context_length=2048 \
  +export.output_name=lfm2_coreml_4w.pte

python -c "
from executorch.runtime import Runtime
prog = Runtime.get().load_program('lfm2_coreml_4w.pte')
prog.load_method('get_eos_ids')   # OK
prog.load_method('forward')       # fails with backend init 0x23 + ANECCompile FAILED
"

Asks

1. Is the short-conv .copy_(...) mutation pattern expected to lower to ANE-compatible MIL? If not, what's the recommended rewrite of examples/models/lfm2/short_conv.py to produce an ANE-friendly graph?
2. Is there a documented way to identify, before compilation, which ops in a model will block ANE compilation?

cc @kimishpatel @YifanShenSZ @cymbalrush @metascroy

[metascroy]

The ANE has a lot of quirks, and generally we've only found static definitions w/ buffer mutation expressed as IO to work really well with ANE. See here for example: https://github.com/pytorch/executorch/tree/main/examples/apple/coreml/llama

The ANE compiler error is a bug in CoreML iteself. The intended behavior is that CoreML falls back to CPU/GPU where ANE lacks support, but there are cases where this doesn't happen.

I'd recommend tagging the problematic ops to run outside of CoreML on the CPU using the partitioner, or rewriting your model to express state mutation as IO.

On knowing what ops are supported on ANE before compilation: I think you need at least modelc compilation to know. See #19252