The MIL scalar types used shorthand syntax (string("x"), int32(1)) that
only works on M4. Changed to the canonical verbose format that CoreML's
own compiler emits (tensor<string, []>("x"), tensor<int32, []>(1)).
Also targets program(1.0) with <ios16> instead of program(1.3)/<ios18>,
and simplifies buildInfo to just coremlc-version.
For conv-based kernels, adds runtime fp16 I/O fallback — M1/M2 ANE
doesn't support the cast op (fp32<->fp16), so on first compile failure
it retries with native fp16 inputs/outputs and does the conversion on
the CPU side. The fallback is persisted across exec() restarts.
Note: matmul and scaled_dot_product_attention ops still fail on M1/M2 —
these are M4+ ANE ops. The attention tests (test_ane_causal_attn,
test_ane_sdpa5, test_full_fused attention part) require M4 hardware.
Conv-based kernels (training, QKV projections, FFN) work on all generations.
Tested on M1 Pro, macOS 26.3 (Tahoe).
imperatormk
maderix