% Generated by roxygen2: do not edit by hand

% Please edit documentation in R/ministral.R

\name{ministral}

\alias{ministral}

\alias{ministral_from_config}

\alias{ministral_from_pretrained}

\title{ministral}

\usage{

ministral(

vocab_size = 131072,

n_embd = 5120,

n_inter = 16384,

n_head = 32,

n_kv_head = 8,

head_dim = 128,

n_layer = 40,

max_pos = 262144,

rmsnorm_eps = 1e-05,

rope_base = 1e+09,

rope_factor = 16,

rope_beta_fast = 32,

rope_beta_slow = 1,

rope_original_max_pos = 16384,

rope_mscale = 1,

rope_mscale_all_dim = 1

)

ministral_from_config(identifier, revision = "main")

ministral_from_pretrained(identifier, revision = "main")

}

\arguments{

\item{vocab_size}{Vocabulary size.}

\item{n_embd}{Embedding dimension.}

\item{n_inter}{Intermediate size in MLP.}

\item{n_head}{Number of attention heads.}

\item{n_kv_head}{Number of key/value heads (for GQA).}

\item{head_dim}{Dimension of each attention head.}

\item{n_layer}{Number of transformer layers.}

\item{max_pos}{Maximum position embeddings.}

\item{rmsnorm_eps}{Epsilon for RMSNorm.}

\item{rope_base}{Base for rotary embeddings.}

\item{rope_factor}{YaRN scaling factor.}

\item{rope_beta_fast}{YaRN beta_fast parameter.}

\item{rope_beta_slow}{YaRN beta_slow parameter.}

\item{rope_original_max_pos}{Original max position embeddings for YaRN.}

\item{rope_mscale}{YaRN mscale parameter.}

\item{rope_mscale_all_dim}{YaRN mscale_all_dim parameter.}

\item{identifier}{HuggingFace model identifier.}

\item{revision}{HuggingFace model revision.}

}

\value{

An initialized \code{\link[torch:nn_module]{torch::nn_module()}}.

}

\description{

Initializes a Ministral-like model with YaRN RoPE and GQA

}

\section{Functions}{

\itemize{

\item \code{ministral_from_config()}: Initializes from HuggingFace config

\item \code{ministral_from_pretrained()}: Initializes and loads pretrained weights from HF Hub

}}