[RFC] Model metadata storage in PTE files via NamedData

[kirklandsign]

🚀 The feature, motivation and pitch

Motivation

PTE files currently carry model weights and delegate data, but lack a standard way to embed model-level metadata — tokenizer config, chat templates, architecture info, etc. Users must ship these as separate files or hardcode them in application code.
A self-contained PTE file that includes all metadata needed to run inference would simplify deployment and reduce integration errors.
PTE already has the infrastructure to support this: NamedData.

Proposal

Use the existing NamedData mechanism (flatbuffer Program.named_data) to store model metadata as key-value pairs. No schema changes needed.

Key naming convention

Follow a namespace.field pattern:

general.name             = "Llama-3.2-1B"
general.architecture     = "llama"
tokenizer.model          = "BPE"
tokenizer.vocab_size     = 128256
tokenizer.bos_token_id   = 128000
tokenizer.eos_token_id   = 128001
tokenizer.chat_template  = "{% for message in messages %}..."
context.length           = 8192

All metadata keys are prefixed with metadata. in the NamedData store to avoid collision with backend entries (e.g., XNNPACK constant tensors).

Value encoding

Values are stored as raw bytes:

• string → UTF-8 encoded bytes
• int → int64 little-endian
• float → float64 little-endian
• bytes → raw bytes

Python API (export side)

from executorch.extension.llm.export.metadata import add_metadata
edge_manager = to_edge_transform_and_lower(exported_program, ...)
add_metadata(edge_manager, {
    "tokenizer.model": "BPE",
    "tokenizer.vocab_size": 128256,
    "tokenizer.chat_template": template_str,
    "general.architecture": "llama",
})
et_program = edge_manager.to_executorch()

C++ API (runtime side)

auto* map = program->get_named_data_map();
auto result = map->get_data("metadata.tokenizer.chat_template");
std::string_view chat_template(
    static_cast<const char*>(result->data()), result->size());

Why NamedData over a new schema field?

• Zero schema changes — works with existing PTE format
• Already battle-tested — XNNPACK, AOTI, CoreML all use NamedData in production
• Linear key lookup is fine for ~20-50 metadata keys
• Prefix-based namespace isolation keeps metadata separate from backend data

Open questions

1. Should we define a standard set of well-known keys, or leave it fully user-defined?
2. Should EdgeProgramManager expose a public add_metadata() method instead of going through _named_data_store?
3. Do we need type tags in the stored bytes, or is type interpretation the caller's responsibility?

Alternatives

No response

Additional context

No response

RFC (Optional)

No response

cc @larryliu0820 @mergennachin @cccclai @helunwencser @jackzhxng