Struct GPTModel 

pub struct GPTModel { /* private fields */ }

[Listing 4.7] The GPT model architecture implementation

Implementations

impl GPTModel

New GPTModel constructor using ConfigV2

pub fn new_v2(cfg: ConfigV2, vb: VarBuilder<'_>) -> Result<Self>

impl GPTModel

pub fn new(cfg: Config, vb: VarBuilder<'_>) -> Result<Self>

Creates a new GPTModel

use candle_core::{Device, DType};
use candle_nn::{VarBuilder, VarMap};
use llms_from_scratch_rs::listings::ch04::{Config, GPTModel};

let dev = Device::cuda_if_available(0).unwrap();
let varmap = VarMap::new();
let vb = VarBuilder::from_varmap(&varmap, DType::F32, &dev);

let cfg = Config::gpt_sm_test();
let model = GPTModel::new(cfg, vb).unwrap();

pub fn from_fields( tok_emb: Embedding, pos_emb: Embedding, drop_emb: Dropout, trf_blocks: SequentialTransformers, final_norm: LayerNorm, out_head: Linear, ) -> Result<Self>

pub fn drop_emb(&self) -> &Dropout

pub fn tok_emb(&self) -> &Embedding

pub fn pos_emb(&self) -> &Embedding

pub fn trf_blocks(&self) -> &SequentialTransformers

pub fn final_norm(&self) -> &LayerNorm

pub fn out_head(&self) -> &Linear

pub fn forward(&self, xs: &Tensor) -> Result<Tensor>

Manual implementation of forward

Note: that blanket implementation of ModuleT when a type implements Module prevents having forward being overrided. Thus, this type is ModuleT but technically not Module.

pub fn set_out_head(&mut self, new_out_head: Linear)

Trait Implementations

impl Clone for GPTModel

fn clone(&self) -> GPTModel

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for GPTModel

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl GPT for GPTModel

fn context_size(&self) -> usize

impl ModuleT for GPTModel

fn forward_t(&self, xs: &Tensor, train: bool) -> Result<Tensor>