hermes-mal

hermes-mal is the single parser and data model for Hermes' Model Architecture Language (MAL). Both hermes-llm and hermes-train consume the same ModelDef, and hermes-mal-python is a thin binding over this crate.

MAL supports reusable attention, state-space, FFN, and block definitions as well as inline definitions:

attention local_gqa {
    num_heads: 16
    num_kv_heads: 4
    window_size: 2048
    position_encoding: rope { theta: 10000 }
}

ffn gated {
    hidden_dim: 4096
    activation: swiglu
}

block transformer {
    attention: local_gqa
    ffn: gated
    norm: rmsnorm { eps: 1e-5 }
}

model example {
    vocab_size: 32000
    max_seq_len: 4096
    hidden_size: 1024
    num_layers: 24
    block: transformer
}

Parse a source containing one model with parse_mal, retain every named definition with parse_mal_full, or use parse_mal_file for a local file:

let model = hermes_mal::parse_mal(source)?;
let layer = model.block_for_layer(0);
println!("heads: {}", layer.num_heads());
# Ok::<(), anyhow::Error>(())

Named references are resolved in source order. Duplicate names, undefined references, unknown properties, unsupported syntax, and a source passed to parse_mal with zero or multiple models are reported as errors.

The well-known/ directory is embedded into the crate. Use list_wellknown_models, get_wellknown_mal, and get_builtin_model to discover or load those definitions without filesystem access.

For heterogeneous models, pattern is repeated cyclically across num_layers. Use ModelDef::block_for_layer and the computed methods on BlockDef when inspecting such a model; the computed methods directly on ModelDef describe its homogeneous/default block.

Run the parser, built-in-model, parameter-estimation, and serialization regression suite with:

cargo test -p hermes-mal