The published file was serialized by the PyPI executorch wheel. The MLX
delegate's FlatBuffer union order is the wire format, and the fork the shipped
MLXBackend runtime is built from (@ms/separate-backends-1.4.1) carries
CumMaxNode at union value 112, which upstream does not have. Every later opcode
shifts by one, so the file's 92 quantized matmuls were tagged 121, which the
runtime reads as ArgPartition, and loading failed with Error::InvalidProgram.
Verified on an iPhone 17 (iPhone17,3): loads in 96 ms, 32 ms per embed, and
cos(query, matching document) = 0.6792 against 0.6601 for the shipping XNNPACK
8da4w build and 0.6524 for the fp32 eager reference.
quantized duplicated precisions.json, which is authoritative for the
quantized/non-quantized partition. default expressed a per-repository
recommendation nothing consumed: the library resolves its own per-platform
default. Both are gone from the schema.
Adds library_name to the model card, and pipeline_tag where it was missing.
Brings the config and card licenses into agreement, and repairs the variant
declarations that did not match the files actually published here.
Converts fine (fp16, ~710 MB) but the .pte fails to LOAD with 0x23 in the portable tensor parser. Distinct from the coremltools#2825 dynamic-shape crash, which fails at execute.
