main.rs

#[cfg(feature = "mkl")]
extern crate intel_mkl_src;

#[cfg(feature = "accelerate")]
extern crate accelerate_src;

use anyhow::{Error as E, Result};
use clap::{Parser, ValueEnum};

use candle_transformers::models::mixformer::{Config, MixFormerSequentialForCausalLM as MixFormer};
use candle_transformers::models::quantized_mixformer::MixFormerSequentialForCausalLM as QMixFormer;

use candle::{DType, Device, Tensor};
use candle_nn::VarBuilder;
use candle_transformers::generation::LogitsProcessor;
use hf_hub::{api::sync::Api, Repo, RepoType};
use tokenizers::Tokenizer;

enum Model {
    MixFormer(MixFormer),
    Quantized(QMixFormer),
}

struct TextGeneration {
    model: Model,
    device: Device,
    tokenizer: Tokenizer,
    logits_processor: LogitsProcessor,
    repeat_penalty: f32,
    repeat_last_n: usize,
    verbose_prompt: bool,
}

impl TextGeneration {
    #[allow(clippy::too_many_arguments)]
    fn new(
        model: Model,
        tokenizer: Tokenizer,
        seed: u64,
        temp: Option<f64>,
        top_p: Option<f64>,
        repeat_penalty: f32,
        repeat_last_n: usize,
        verbose_prompt: bool,
        device: &Device,
    ) -> Self {
        let logits_processor = LogitsProcessor::new(seed, temp, top_p);
        Self {
            model,
            tokenizer,
            logits_processor,
            repeat_penalty,
            repeat_last_n,
            verbose_prompt,
            device: device.clone(),
        }
    }

    fn run(&mut self, prompt: &str, sample_len: usize) -> Result<()> {
        use std::io::Write;
        println!("starting the inference loop");
        let tokens = self.tokenizer.encode(prompt, true).map_err(E::msg)?;
        if tokens.is_empty() {
            anyhow::bail!("Empty prompts are not supported in the phi model.")
        }
        if self.verbose_prompt {
            for (token, id) in tokens.get_tokens().iter().zip(tokens.get_ids().iter()) {
                let token = token.replace('▁', " ").replace("<0x0A>", "\n");
                println!("{id:7} -> '{token}'");
            }
        }
        let mut tokens = tokens.get_ids().to_vec();
        let mut generated_tokens = 0usize;
        let eos_token = match self.tokenizer.get_vocab(true).get("<|endoftext|>") {
            Some(token) => *token,
            None => anyhow::bail!("cannot find the endoftext token"),
        };
        print!("{prompt}");
        std::io::stdout().flush()?;
        let start_gen = std::time::Instant::now();
        for index in 0..sample_len {
            let context_size = if index > 0 { 1 } else { tokens.len() };
            let ctxt = &tokens[tokens.len().saturating_sub(context_size)..];
            let input = Tensor::new(ctxt, &self.device)?.unsqueeze(0)?;
            let logits = match &mut self.model {
                Model::MixFormer(m) => m.forward(&input)?,
                Model::Quantized(m) => m.forward(&input)?,
            };
            let logits = logits.squeeze(0)?.to_dtype(DType::F32)?;
            let logits = if self.repeat_penalty == 1. {
                logits
            } else {
                let start_at = tokens.len().saturating_sub(self.repeat_last_n);
                candle_transformers::utils::apply_repeat_penalty(
                    &logits,
                    self.repeat_penalty,
                    &tokens[start_at..],
                )?
            };

            let next_token = self.logits_processor.sample(&logits)?;
            tokens.push(next_token);
            generated_tokens += 1;
            if next_token == eos_token {
                break;
            }
            let token = self.tokenizer.decode(&[next_token], true).map_err(E::msg)?;
            print!("{token}");
            std::io::stdout().flush()?;
        }
        let dt = start_gen.elapsed();
        println!(
            "\n{generated_tokens} tokens generated ({:.2} token/s)",
            generated_tokens as f64 / dt.as_secs_f64(),
        );
        Ok(())
    }
}

#[derive(Clone, Copy, Debug, ValueEnum)]
enum WhichModel {
    #[value(name = "1")]
    V1,
    #[value(name = "1.5")]
    V1_5,
    PuffinPhiV2,
    PhiHermes,
}

#[derive(Parser, Debug)]
#[command(author, version, about, long_about = None)]
struct Args {
    /// Run on CPU rather than on GPU.
    #[arg(long)]
    cpu: bool,

    /// Enable tracing (generates a trace-timestamp.json file).
    #[arg(long)]
    tracing: bool,

    /// Display the token for the specified prompt.
    #[arg(long)]
    verbose_prompt: bool,

    #[arg(long)]
    prompt: String,

    /// The temperature used to generate samples.
    #[arg(long)]
    temperature: Option<f64>,

    /// Nucleus sampling probability cutoff.
    #[arg(long)]
    top_p: Option<f64>,

    /// The seed to use when generating random samples.
    #[arg(long, default_value_t = 299792458)]
    seed: u64,

    /// The length of the sample to generate (in tokens).
    #[arg(long, short = 'n', default_value_t = 100)]
    sample_len: usize,

    #[arg(long)]
    model_id: Option<String>,

    #[arg(long, default_value = "1.5")]
    model: WhichModel,

    #[arg(long)]
    revision: Option<String>,

    #[arg(long)]
    weight_file: Option<String>,

    #[arg(long)]
    tokenizer: Option<String>,

    #[arg(long)]
    quantized: bool,

    /// Penalty to be applied for repeating tokens, 1. means no penalty.
    #[arg(long, default_value_t = 1.1)]
    repeat_penalty: f32,

    /// The context size to consider for the repeat penalty.
    #[arg(long, default_value_t = 64)]
    repeat_last_n: usize,
}

fn main() -> Result<()> {
    use tracing_chrome::ChromeLayerBuilder;
    use tracing_subscriber::prelude::*;

    let args = Args::parse();
    let _guard = if args.tracing {
        let (chrome_layer, guard) = ChromeLayerBuilder::new().build();
        tracing_subscriber::registry().with(chrome_layer).init();
        Some(guard)
    } else {
        None
    };
    println!(
        "avx: {}, neon: {}, simd128: {}, f16c: {}",
        candle::utils::with_avx(),
        candle::utils::with_neon(),
        candle::utils::with_simd128(),
        candle::utils::with_f16c()
    );
    println!(
        "temp: {:.2} repeat-penalty: {:.2} repeat-last-n: {}",
        args.temperature.unwrap_or(0.),
        args.repeat_penalty,
        args.repeat_last_n
    );

    let start = std::time::Instant::now();
    let api = Api::new()?;
    let model_id = match args.model_id {
        Some(model_id) => model_id.to_string(),
        None => {
            if args.quantized {
                "lmz/candle-quantized-phi".to_string()
            } else {
                match args.model {
                    WhichModel::V1 => "microsoft/phi-1".to_string(),
                    WhichModel::V1_5 => "microsoft/phi-1_5".to_string(),
                    WhichModel::PuffinPhiV2 | WhichModel::PhiHermes => {
                        "lmz/candle-quantized-phi".to_string()
                    }
                }
            }
        }
    };
    let revision = match args.revision {
        Some(rev) => rev.to_string(),
        None => {
            if args.quantized {
                "main".to_string()
            } else {
                match args.model {
                    WhichModel::V1 => "refs/pr/2".to_string(),
                    WhichModel::V1_5 => "refs/pr/18".to_string(),
                    WhichModel::PuffinPhiV2 | WhichModel::PhiHermes => "main".to_string(),
                }
            }
        }
    };
    let repo = api.repo(Repo::with_revision(model_id, RepoType::Model, revision));
    let tokenizer_filename = match args.tokenizer {
        Some(file) => std::path::PathBuf::from(file),
        None => match args.model {
            WhichModel::V1 | WhichModel::V1_5 => repo.get("tokenizer.json")?,
            WhichModel::PuffinPhiV2 | WhichModel::PhiHermes => {
                repo.get("tokenizer-puffin-phi-v2.json")?
            }
        },
    };
    let filename = match args.weight_file {
        Some(weight_file) => std::path::PathBuf::from(weight_file),
        None => {
            if args.quantized {
                match args.model {
                    WhichModel::V1 => repo.get("model-v1-q4k.gguf")?,
                    WhichModel::V1_5 => repo.get("model-q4k.gguf")?,
                    WhichModel::PuffinPhiV2 => repo.get("model-puffin-phi-v2-q4k.gguf")?,
                    WhichModel::PhiHermes => repo.get("model-phi-hermes-1_3B-q4k.gguf")?,
                }
            } else {
                match args.model {
                    WhichModel::V1 | WhichModel::V1_5 => repo.get("model.safetensors")?,
                    WhichModel::PuffinPhiV2 => repo.get("model-puffin-phi-v2.safetensors")?,
                    WhichModel::PhiHermes => repo.get("model-phi-hermes-1_3B.safetensors")?,
                }
            }
        }
    };
    println!("retrieved the files in {:?}", start.elapsed());
    let tokenizer = Tokenizer::from_file(tokenizer_filename).map_err(E::msg)?;

    let start = std::time::Instant::now();
    let config = match args.model {
        WhichModel::V1 => Config::v1(),
        WhichModel::V1_5 => Config::v1_5(),
        WhichModel::PuffinPhiV2 => Config::puffin_phi_v2(),
        WhichModel::PhiHermes => Config::phi_hermes_1_3b(),
    };
    let (model, device) = if args.quantized {
        let vb = candle_transformers::quantized_var_builder::VarBuilder::from_gguf(&filename)?;
        let model = QMixFormer::new(&config, vb)?;
        (Model::Quantized(model), Device::Cpu)
    } else {
        let device = candle_examples::device(args.cpu)?;
        let vb = unsafe { VarBuilder::from_mmaped_safetensors(&[filename], DType::F32, &device)? };
        let model = MixFormer::new(&config, vb)?;
        (Model::MixFormer(model), device)
    };
    println!("loaded the model in {:?}", start.elapsed());

    let mut pipeline = TextGeneration::new(
        model,
        tokenizer,
        args.seed,
        args.temperature,
        args.top_p,
        args.repeat_penalty,
        args.repeat_last_n,
        args.verbose_prompt,
        &device,
    );
    pipeline.run(&args.prompt, args.sample_len)?;
    Ok(())
}