use super::NanoChatGpt;
use crate::kits::support::{
construct::{
LmConstruct,
LmPolicy,
},
model_config::{
GenerationPolicy,
SamplingPolicy,
},
};
pub const CHECKPOINT: &str = "checkpoint";
pub const META: &str = "meta";
pub const TOKENIZER: &str = "tokenizer";
pub const TOKEN_BYTES: &str = "token_bytes";
pub type NanoChatPolicy = LmPolicy;
pub const NANOCHAT_SAMPLING: SamplingPolicy = SamplingPolicy {
temperature: 0.6,
top_k: Some(50),
top_p: None,
min_p: None,
};
pub fn nanochat_default_policy() -> NanoChatPolicy {
LmPolicy {
generation: GenerationPolicy {
sampling: Some(NANOCHAT_SAMPLING),
..Default::default()
},
}
}
pub type NanoChatConstruct = LmConstruct<NanoChatGpt>;
#[cfg(test)]
mod tests {
use std::path::Path;
use super::*;
use crate::{
burner::module::ModuleInit,
kits::language::nanochat::NanoChatGptContractConfig,
support::testing::cpu_device,
};
#[test]
fn test_files_by_key() {
let model: NanoChatGpt = NanoChatGptContractConfig::new()
.with_vocab_size(16)
.with_n_layer(1)
.with_n_head(2)
.with_n_kv_head(2)
.with_n_embed(8)
.with_init_seq_len(4)
.init(&cpu_device());
let bundle = NanoChatConstruct::new(model)
.with_max_positions(Some(4))
.with_file(TOKENIZER, "/x/tokenizer.pkl");
assert_eq!(bundle.max_positions, Some(4));
assert_eq!(bundle.tokenizer(), Some(Path::new("/x/tokenizer.pkl")));
assert_eq!(bundle.file(TOKEN_BYTES), None);
assert_eq!(
nanochat_default_policy().generation.sampling,
Some(NANOCHAT_SAMPLING)
);
}
}