2026-04-02 22:22:52 +08:00
|
|
|
use crate::models::common::MultiModalData;
|
|
|
|
|
use crate::models::common::generate::{
|
|
|
|
|
GenerationContext, generate_generic, generate_stream_generic,
|
|
|
|
|
};
|
2026-03-31 12:30:12 +08:00
|
|
|
use crate::params::chat::{
|
2025-10-27 10:55:40 +08:00
|
|
|
ChatCompletionChunkResponse, ChatCompletionParameters, ChatCompletionResponse,
|
|
|
|
|
};
|
2026-04-02 22:22:52 +08:00
|
|
|
use anyhow::Result;
|
|
|
|
|
use candle_core::{DType, Device};
|
2025-09-25 12:09:25 +08:00
|
|
|
use candle_nn::VarBuilder;
|
|
|
|
|
use rocket::futures::Stream;
|
|
|
|
|
|
2025-10-15 21:03:49 +08:00
|
|
|
use crate::models::minicpm4::config::MiniCPM4Config;
|
|
|
|
|
use crate::models::minicpm4::model::MiniCPMModel;
|
|
|
|
|
// use crate::models::GenerateStream;
|
2026-04-02 22:22:52 +08:00
|
|
|
use crate::utils::{find_type_files, get_device, get_dtype};
|
2025-10-15 21:03:49 +08:00
|
|
|
use crate::{chat_template::ChatTemplate, models::GenerateModel, tokenizer::TokenizerModel};
|
|
|
|
|
|
2025-09-25 12:09:25 +08:00
|
|
|
pub struct MiniCPMGenerateModel<'a> {
|
|
|
|
|
chat_template: ChatTemplate<'a>,
|
|
|
|
|
tokenizer: TokenizerModel,
|
|
|
|
|
minicpm: MiniCPMModel,
|
|
|
|
|
device: Device,
|
2025-12-03 17:21:01 +08:00
|
|
|
model_name: String,
|
2025-09-25 12:09:25 +08:00
|
|
|
}
|
|
|
|
|
|
2025-10-15 21:03:49 +08:00
|
|
|
impl<'a> MiniCPMGenerateModel<'a> {
|
2025-09-25 12:44:17 +08:00
|
|
|
pub fn init(path: &str, device: Option<&Device>, dtype: Option<DType>) -> Result<Self> {
|
2025-09-25 12:09:25 +08:00
|
|
|
let chat_template = ChatTemplate::init(path)?;
|
|
|
|
|
let tokenizer = TokenizerModel::init(path)?;
|
|
|
|
|
let config_path = path.to_string() + "/config.json";
|
|
|
|
|
let cfg: MiniCPM4Config = serde_json::from_slice(&std::fs::read(config_path)?)?;
|
|
|
|
|
let device = &get_device(device);
|
|
|
|
|
let cfg_dtype = cfg.torch_dtype.as_str();
|
|
|
|
|
let dtype = get_dtype(dtype, cfg_dtype);
|
|
|
|
|
let endoftext_id = cfg.eos_token_id[0];
|
|
|
|
|
let im_end_id = cfg.eos_token_id[1];
|
2025-10-15 21:03:49 +08:00
|
|
|
let model_list = find_type_files(path, "safetensors")?;
|
2025-09-25 12:09:25 +08:00
|
|
|
let vb = unsafe { VarBuilder::from_mmaped_safetensors(&model_list, dtype, device)? };
|
2026-04-02 22:22:52 +08:00
|
|
|
let eos_ids = vec![endoftext_id, im_end_id];
|
|
|
|
|
let minicpm = MiniCPMModel::new(vb, cfg, eos_ids)?;
|
2026-03-30 20:33:48 +08:00
|
|
|
let model_name = std::path::Path::new(path)
|
2026-03-30 20:36:57 +08:00
|
|
|
.file_name()
|
2026-03-30 20:33:48 +08:00
|
|
|
.and_then(|s| s.to_str())
|
|
|
|
|
.unwrap_or("minicpm4")
|
|
|
|
|
.to_string();
|
2025-09-25 12:09:25 +08:00
|
|
|
Ok(MiniCPMGenerateModel {
|
|
|
|
|
chat_template,
|
|
|
|
|
tokenizer,
|
|
|
|
|
minicpm,
|
|
|
|
|
device: device.clone(),
|
2026-03-30 20:33:48 +08:00
|
|
|
model_name,
|
2025-09-25 12:09:25 +08:00
|
|
|
})
|
|
|
|
|
}
|
2025-09-25 12:44:17 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
impl<'a> GenerateModel for MiniCPMGenerateModel<'a> {
|
2025-09-25 12:09:25 +08:00
|
|
|
fn generate(&mut self, mes: ChatCompletionParameters) -> Result<ChatCompletionResponse> {
|
|
|
|
|
let mes_render = self.chat_template.apply_chat_template(&mes)?;
|
2026-04-02 22:22:52 +08:00
|
|
|
let input_ids = self.tokenizer.text_encode(mes_render, &self.device)?;
|
|
|
|
|
let seed = mes.seed.unwrap_or(34562) as u64;
|
2025-10-15 21:03:49 +08:00
|
|
|
let sample_len = mes.max_tokens.unwrap_or(2048);
|
2026-04-02 22:22:52 +08:00
|
|
|
let mut ctx = GenerationContext::new(
|
|
|
|
|
mes.temperature,
|
|
|
|
|
mes.top_p,
|
|
|
|
|
None,
|
2026-04-03 14:15:10 +08:00
|
|
|
mes.repeat_penalty,
|
|
|
|
|
mes.repeat_last_n,
|
2026-04-02 22:22:52 +08:00
|
|
|
seed,
|
|
|
|
|
input_ids.dim(1)?,
|
|
|
|
|
sample_len,
|
|
|
|
|
self.device.clone(),
|
|
|
|
|
);
|
|
|
|
|
|
|
|
|
|
let data = MultiModalData::new(vec![]);
|
|
|
|
|
generate_generic(
|
|
|
|
|
&mut self.minicpm,
|
|
|
|
|
&self.tokenizer,
|
|
|
|
|
input_ids,
|
|
|
|
|
data,
|
|
|
|
|
&mut ctx,
|
|
|
|
|
&self.model_name,
|
|
|
|
|
)
|
2025-09-25 12:09:25 +08:00
|
|
|
}
|
|
|
|
|
fn generate_stream(
|
|
|
|
|
&mut self,
|
|
|
|
|
mes: ChatCompletionParameters,
|
2025-11-05 14:46:03 +08:00
|
|
|
) -> Result<
|
|
|
|
|
Box<
|
|
|
|
|
dyn Stream<Item = Result<ChatCompletionChunkResponse, anyhow::Error>>
|
|
|
|
|
+ Send
|
|
|
|
|
+ Unpin
|
|
|
|
|
+ '_,
|
|
|
|
|
>,
|
|
|
|
|
> {
|
2026-03-06 19:08:58 +08:00
|
|
|
let seed = mes.seed.unwrap_or(34562) as u64;
|
2025-09-25 12:09:25 +08:00
|
|
|
let mes_render = self.chat_template.apply_chat_template(&mes)?;
|
2026-04-02 22:22:52 +08:00
|
|
|
let input_ids = self.tokenizer.text_encode(mes_render, &self.device)?;
|
|
|
|
|
let data = MultiModalData::new(vec![]);
|
2025-10-15 21:03:49 +08:00
|
|
|
let sample_len = mes.max_tokens.unwrap_or(512);
|
2026-04-02 22:22:52 +08:00
|
|
|
let stream = generate_stream_generic(
|
|
|
|
|
&mut self.minicpm,
|
|
|
|
|
&self.tokenizer,
|
|
|
|
|
input_ids,
|
|
|
|
|
data,
|
|
|
|
|
mes.temperature,
|
|
|
|
|
mes.top_p,
|
|
|
|
|
None,
|
2026-04-03 14:15:10 +08:00
|
|
|
mes.repeat_penalty,
|
|
|
|
|
mes.repeat_last_n,
|
2026-04-02 22:22:52 +08:00
|
|
|
seed,
|
|
|
|
|
sample_len,
|
|
|
|
|
false,
|
|
|
|
|
&self.device,
|
|
|
|
|
&self.model_name,
|
|
|
|
|
)?;
|
2025-11-05 14:46:03 +08:00
|
|
|
Ok(Box::new(Box::pin(stream)))
|
2025-09-25 12:09:25 +08:00
|
|
|
}
|
|
|
|
|
}
|