def load_model

bend-ml-llama@0.2.0.0 · main.bend:1283

import Base
import bend-ml-tensor-array@0.1.14.0/main.bend as TA

1283# the whole model; the embedding gets one more level of bands than the layers (2^5 with 16 threads)
1284def load_model(+c: Cfg, +dir: String) -> IO(Model<c>):
1285  do IO<Model<c>>:
1286    threads : U32 <- IO.thread_count()
1287    +d : Nat = depth_for(threads)
1288    emb : TA.Bands<Cfg.vocab(c), Cfg.dim(c)> <- get_bands(dir, "emb", Nat.add(d, Nat.min(d, 1n)), Cfg.vocab(c), Cfg.dim(c))
1289    ls : List<&1, LW<c>> <- load_layers(Cfg.layers(c), c, dir, d, 0n, Nil{})
1290    lnf : List<&2, F32> <- get_list(dir, "lnf", Cfg.dim(c))
1291    return Model{d, emb, ls, lnf}