Add the Mixtral model. (huggingface#1437)

* Add the Mixtral model.

* Add more of the mixtral layers.

* Add the final layers for mixtral.

* Sketch the expert selection.

* Add some expert routing logic.

* Hopefully finish the routing logic for mixtral.

* Add the mixtral example.

* Fix the weight filenames.

* Bugfix.

* Another fix.

* Yet another fix + remove the unused pragma.

* Shape fix.

* Add a readme.

candle-examples/examples/mixtral/README.md

@@ -0,0 +1,25 @@
+# candle-mixtral: 8x7b LLM using a sparse mixture of experts.
+
+Mixtral-8x7B-v0.1 is a pretrained generative LLM with 56 billion parameters. 
+
+- [Blog post](https://mistral.ai/news/mixtral-of-experts/) from Mistral announcing the model release.
+- [Model card](https://huggingface.co/mistralai/Mixtral-8x7B-v0.1) on the HuggingFace Hub.
+
+## Running the example
+
+```bash
+$ cargo run --example mixtral --release  -- --prompt "def print_prime(n): "
+def print_prime(n):  # n is the number of prime numbers to be printed
+    i = 2
+    count = 0
+    while (count < n):
+        if (isPrime(i)):
+            print(i)
+            count += 1
+        i += 1
+
+def isPrime(n):
+    for x in range(2, int(n**0.5)+1):
+        if (n % x == 0):
+            ...
+```

candle-examples/examples/mixtral/main.rs

@@ -0,0 +1,263 @@
+#[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::mixtral::{Config, Model};
+
+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;
+
+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>,
+        repeat_penalty: f32,
+        repeat_last_n: usize,
+        device: &Device,
+    ) -> Self {
+        let logits_processor = LogitsProcessor::new(seed, temp, top_p);
+        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("</s>") {
+            Some(token) => token,
+            None => anyhow::bail!("cannot find the </s> 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)]
+    use_flash_attn: 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, default_value = "mistralai/Mixtral-8x7B-v0.1")]
+    model_id: String,
+
+    #[arg(long, default_value = "main")]
+    revision: String,
+
+    #[arg(long)]
+    tokenizer_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,
+}
+
+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 repo = api.repo(Repo::with_revision(
+        args.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 filenames = match args.weight_files {
+        Some(files) => files
+            .split(',')
+            .map(std::path::PathBuf::from)
+            .collect::<Vec<_>>(),
+        None => {
+            vec![
+                repo.get("model-00001-of-00019.safetensors")?,
+                repo.get("model-00002-of-00019.safetensors")?,
+                repo.get("model-00003-of-00019.safetensors")?,
+                repo.get("model-00004-of-00019.safetensors")?,
+                repo.get("model-00005-of-00019.safetensors")?,
+                repo.get("model-00006-of-00019.safetensors")?,
+                repo.get("model-00007-of-00019.safetensors")?,
+                repo.get("model-00008-of-00019.safetensors")?,
+                repo.get("model-00009-of-00019.safetensors")?,
+                repo.get("model-00010-of-00019.safetensors")?,
+                repo.get("model-00011-of-00019.safetensors")?,
+                repo.get("model-00012-of-00019.safetensors")?,
+                repo.get("model-00013-of-00019.safetensors")?,
+                repo.get("model-00014-of-00019.safetensors")?,
+                repo.get("model-00015-of-00019.safetensors")?,
+                repo.get("model-00016-of-00019.safetensors")?,
+                repo.get("model-00017-of-00019.safetensors")?,
+                repo.get("model-00018-of-00019.safetensors")?,
+                repo.get("model-00019-of-00019.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 = Config::v0_1_8x7b(args.use_flash_attn);
+    let device = candle_examples::device(args.cpu)?;
+    let dtype = if device.is_cuda() {
+        DType::BF16
+    } else {
+        DType::F32
+    };
+    let vb = unsafe { VarBuilder::from_mmaped_safetensors(&filenames, dtype, &device)? };
+    let model = Model::new(&config, vb)?;
+    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,
+        &device,
+    );
+    pipeline.run(&args.prompt, args.sample_len)?;
+    Ok(())
+}