Example quantized with custom GGUF model error: cannot find llama.attention.head_count in metadata

[evgenyigumnov]

C:\Users\igumn\candle\candle-examples\examples\quantized>cargo run --features=cuda --example quantized  --release -- --model=gemma-2-2b-it.q4_k_m.gguf --prompt "def fibonacci(n): "
    Finished `release` profile [optimized] target(s) in 0.48s
     Running `C:\Users\igumn\candle\target\release\examples\quantized.exe --model=gemma-2-2b-it.q4_k_m.gguf --prompt "def fibonacci(n): "`
avx: true, neon: false, simd128: false, f16c: true
temp: 0.80 repeat-penalty: 1.10 repeat-last-n: 64
loaded 288 tensors (1.70GB) in 1.70s
Error: cannot find llama.attention.head_count in metadata
error: process didn't exit successfully: `C:\Users\igumn\candle\target\release\examples\quantized.exe --model=gemma-2-2b-it.q4_k_m.gguf --prompt "def fibonacci(n): "` (exit code: 1)

https://huggingface.co/unsloth/gemma-2-it-GGUF/blob/main/gemma-2-2b-it.q4_k_m.gguf

Probably you could add "quantized" support for "gemma" example. For gemma-2-2b-it.q4_k_m.gguf model. I try adopting "recurrent-gemma" example but take error: "Error: missing field conv1d_width at line 36 column 1"

my changed "recurrent-gemma" example code:

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

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

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

use candle_transformers::models::quantized_recurrent_gemma::Model as QModel;
use candle_transformers::models::recurrent_gemma::{Config, Model as BModel};

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

enum Model {
    B(BModel),
    Q(QModel),
}

impl Model {
    fn forward(&mut self, xs: &Tensor, pos: usize) -> candle::Result<Tensor> {
        match self {
            Self::B(m) => m.forward(xs, pos),
            Self::Q(m) => m.forward(xs, pos),
        }
    }
}

#[derive(Clone, Debug, Copy, PartialEq, Eq, clap::ValueEnum)]
enum Which {
    #[value(name = "2b")]
    Base2B,
    #[value(name = "2b-it")]
    Instruct2B,
}

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

impl TextGeneration {
    #[allow(clippy::too_many_arguments)]
    fn new(
        model: Model,
        tokenizer: Tokenizer,
        seed: u64,
        temp: Option<f64>,
        top_p: Option<f64>,
        top_k: usize,
        repeat_penalty: f32,
        repeat_last_n: usize,
        device: &Device,
    ) -> Self {
        let sampling = match temp {
            None => candle_transformers::generation::Sampling::ArgMax,
            Some(temperature) => match top_p {
                None => candle_transformers::generation::Sampling::TopK {
                    temperature,
                    k: top_k,
                },
                Some(top_p) => candle_transformers::generation::Sampling::TopKThenTopP {
                    temperature,
                    k: top_k,
                    p: top_p,
                },
            },
        };
        let logits_processor = LogitsProcessor::from_sampling(seed, sampling);
        Self {
            model,
            tokenizer: TokenOutputStream::new(tokenizer),
            logits_processor,
            repeat_penalty,
            repeat_last_n,
            device: device.clone(),
        }
    }

    fn run(&mut self, prompt: &str, sample_len: usize) -> Result<()> {
        use std::io::Write;
        self.tokenizer.clear();
        let mut tokens = self
            .tokenizer
            .tokenizer()
            .encode(prompt, true)
            .map_err(E::msg)?
            .get_ids()
            .to_vec();
        for &t in tokens.iter() {
            if let Some(t) = self.tokenizer.next_token(t)? {
                print!("{t}")
            }
        }
        std::io::stdout().flush()?;

        let mut generated_tokens = 0usize;
        let eos_token = match self.tokenizer.get_token("<eos>") {
            Some(token) => token,
            None => anyhow::bail!("cannot find the <eos> token"),
        };
        let start_gen = std::time::Instant::now();
        for index in 0..sample_len {
            let context_size = if index > 0 { 1 } else { tokens.len() };
            let start_pos = tokens.len().saturating_sub(context_size);
            let ctxt = &tokens[start_pos..];
            let input = Tensor::new(ctxt, &self.device)?.unsqueeze(0)?;
            let logits = self.model.forward(&input, start_pos)?;
            let logits = logits.squeeze(0)?.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;
            }
            if let Some(t) = self.tokenizer.next_token(next_token)? {
                print!("{t}");
                std::io::stdout().flush()?;
            }
        }
        let dt = start_gen.elapsed();
        if let Some(rest) = self.tokenizer.decode_rest().map_err(E::msg)? {
            print!("{rest}");
        }
        std::io::stdout().flush()?;
        println!(
            "\n{generated_tokens} tokens generated ({:.2} token/s)",
            generated_tokens as f64 / dt.as_secs_f64(),
        );
        Ok(())
    }
}

#[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,

    #[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>,

    #[arg(long, default_value_t = 250)]
    top_k: usize,

    /// 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 = 8000)]
    sample_len: usize,

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

    #[arg(long, default_value = "main")]
    revision: String,

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

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

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

    /// 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,

    /// The model to use.
    #[arg(long, default_value = "2b")]
    which: Which,

    #[arg(long)]
    quantized: bool,
}

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 => match args.which {
            Which::Base2B => "google/recurrentgemma-2b".to_string(),
            Which::Instruct2B => "google/recurrentgemma-2b-it".to_string(),
        },
    };
    let repo = api.repo(Repo::with_revision(
        model_id,
        RepoType::Model,
        args.revision,
    ));
    // let tokenizer_filename = match args.tokenizer_file {
    //     Some(file) => std::path::PathBuf::from(file),
    //     None => repo.get("tokenizer.json")?,
    // };
    let tokenizer_filename = std::path::PathBuf::from("tokenizer.json");
    // let config_filename = match args.config_file {
    //     Some(file) => std::path::PathBuf::from(file),
    //     None => repo.get("config.json")?,
    // };
    let config_filename = std::path::PathBuf::from("config.json");
    // let filenames = match args.weight_files {
    //     Some(files) => files
    //         .split(',')
    //         .map(std::path::PathBuf::from)
    //         .collect::<Vec<_>>(),
    //     None => {
    //         if args.quantized {
    //             let filename = match args.which {
    //                 Which::Base2B => "recurrent-gemma-2b-q4k.gguf",
    //                 Which::Instruct2B => "recurrent-gemma-7b-q4k.gguf",
    //             };
    //             let filename = api.model("lmz/candle-gemma".to_string()).get(filename)?;
    //             vec![filename]
    //         } else {
    //             candle_examples::hub_load_safetensors(&repo, "model.safetensors.index.json")?
    //         }
    //     }
    // };
    let filenames = vec![std::path::PathBuf::from("gemma-2-2b-it.q4_k_m.gguf ")];
    println!("retrieved the files in {:?}", start.elapsed());
    let tokenizer = Tokenizer::from_file(tokenizer_filename).map_err(E::msg)?;
    let config: Config = serde_json::from_reader(std::fs::File::open(config_filename)?)?;

    let start = std::time::Instant::now();
    let device = candle_examples::device(args.cpu)?;
    let dtype = if device.is_cuda() {
        DType::BF16
    } else {
        DType::F32
    };
    let model = if args.quantized {
        let vb = candle_transformers::quantized_var_builder::VarBuilder::from_gguf(
            &filenames[0],
            &device,
        )?;
        Model::Q(QModel::new(&config, vb.pp("model"))?)
    } else {
        let vb = unsafe { VarBuilder::from_mmaped_safetensors(&filenames, dtype, &device)? };
        Model::B(BModel::new(&config, vb.pp("model"))?)
    };

    println!("loaded the model in {:?}", start.elapsed());

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

cargo run --features cuda -r --example recurrent-gemma -- --prompt "Write me a poem about Machine Learning." --quantized