Improved mamba model optimized for inference (huggingface#1694)

* Sketch the mamba model for inference.

* Complete the forward pass.

* Add the mamba example.

* Optimize the selective-scan part.

* Fix a couple shape mismatches and get inference to work.

* Tweak the readmes.

* More readme tweaks.

README.md

@@ -67,7 +67,7 @@ We also provide a some command line based examples using state of the art models
 - [StableLM-3B-4E1T](./candle-examples/examples/stable-lm/): a 3b general LLM
   pre-trained on 1T tokens of English and code datasets. Also supports
   StableLM-2, a 1.6b LLM trained on 2T tokens, as well as the code variants.
-- [Minimal Mamba](./candle-examples/examples/mamba-minimal/): a minimal
+- [Mamba](./candle-examples/examples/mamba/): an inference only
   implementation of the Mamba state space model.
 - [Mistral7b-v0.1](./candle-examples/examples/mistral/): a 7b general LLM with
   better performance than all publicly available 13b models as of 2023-09-28.
@@ -186,7 +186,7 @@ If you have an addition to this list, please submit a pull request.
         - Falcon.
         - StarCoder.
         - Phi 1, 1.5, and 2.
-        - Minimal Mamba
+        - Mamba, Minimal Mamba
         - Mistral 7b v0.1.
         - Mixtral 8x7b v0.1.
         - StableLM-3B-4E1T, StableLM-2-1.6B, Stable-Code-3B.

candle-examples/examples/mamba-minimal/README.md

@@ -2,6 +2,9 @@
 
 This is based on [mamba-minimal](https://github.com/johnma2006/mamba-minimal).
 
+Compared to the mamba example, this version can handle training but is much
+slower.
+
 ## Running the example
 
 ```bash

candle-examples/examples/mamba/README.md

@@ -0,0 +1,17 @@
+# candle-mamba: Mamba implementation
+
+Candle implementation of *Mamba* [1] inference only. Mamba is an alternative to
+the transformer architecture. It leverages State Space Models (SSMs) with the
+goal of being computationally efficient on long sequences. The implementation is
+based on [mamba.rs](https://github.com/LaurentMazare/mamba.rs).
+
+- [1]. [Mamba: Linear-Time Sequence Modeling with Selective State Spaces](https://arxiv.org/abs/2312.00752).
+
+Compared to the mamba-minimal example, this version is far more efficient but
+would only work for inference.
+## Running the example
+
+```bash
+$ cargo run --example mamba-minimal --release -- --prompt "Mamba is the"
+```
+

candle-examples/examples/mamba/main.rs

@@ -0,0 +1,299 @@
+#[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::mamba::{Config, Model, State};
+
+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,
+    config: Config,
+    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,
+        config: Config,
+        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,
+            config,
+            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();
+        let mut generated_tokens = 0usize;
+        let eos_token = match self.tokenizer.get_token("<|endoftext|>") {
+            Some(token) => token,
+            None => anyhow::bail!("cannot find the </s> token"),
+        };
+        let mut state = State::new(1, &self.config, &self.device)?;
+        let mut next_logits = None;
+        for &t in tokens.iter() {
+            let input = Tensor::new(&[t], &self.device)?;
+            let logits = self.model.forward(&input, &mut state)?;
+            next_logits = Some(logits);
+            if let Some(t) = self.tokenizer.next_token(t)? {
+                print!("{t}")
+            }
+        }
+        std::io::stdout().flush()?;
+
+        let start_gen = std::time::Instant::now();
+        for _ in 0..sample_len {
+            let logits = match next_logits.as_ref() {
+                Some(logits) => logits,
+                None => anyhow::bail!("cannot work on an empty prompt"),
+            };
+            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;
+            }
+            if let Some(t) = self.tokenizer.next_token(next_token)? {
+                print!("{t}");
+                std::io::stdout().flush()?;
+            }
+
+            let input = Tensor::new(&[next_token], &self.device)?;
+            next_logits = Some(self.model.forward(&input, &mut state)?)
+        }
+        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, ValueEnum, Clone, Copy, PartialEq, Eq, Debug)]
+enum Which {
+    Mamba130m,
+    Mamba370m,
+    Mamba790m,
+    Mamba1_4b,
+    Mamba2_8b,
+    Mamba2_8bSlimPj,
+}
+
+impl std::fmt::Display for Which {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "{:?}", self)
+    }
+}
+
+impl Which {
+    fn model_id(&self) -> &'static str {
+        match self {
+            Self::Mamba130m => "state-spaces/mamba-130m",
+            Self::Mamba370m => "state-spaces/mamba-370m",
+            Self::Mamba790m => "state-spaces/mamba-790m",
+            Self::Mamba1_4b => "state-spaces/mamba-1.4b",
+            Self::Mamba2_8b => "state-spaces/mamba-2.8b",
+            Self::Mamba2_8bSlimPj => "state-spaces/mamba-2.8b-slimpj'",
+        }
+    }
+
+    fn revision(&self) -> &'static str {
+        match self {
+            Self::Mamba130m
+            | Self::Mamba370m
+            | Self::Mamba790m
+            | Self::Mamba1_4b
+            | Self::Mamba2_8bSlimPj => "refs/pr/1",
+            Self::Mamba2_8b => "refs/pr/4",
+        }
+    }
+}
+
+#[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>,
+
+    /// 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 = 5000)]
+    sample_len: usize,
+
+    #[arg(long, default_value = "mamba130m")]
+    which: Which,
+
+    #[arg(long)]
+    model_id: Option<String>,
+
+    #[arg(long)]
+    revision: Option<String>,
+
+    #[arg(long)]
+    tokenizer_file: Option<String>,
+
+    #[arg(long)]
+    weight_files: Option<String>,
+
+    #[arg(long)]
+    config_file: 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
+            .unwrap_or_else(|| args.which.model_id().to_string()),
+        RepoType::Model,
+        args.revision
+            .unwrap_or_else(|| args.which.revision().to_string()),
+    ));
+    let tokenizer_filename = match args.tokenizer_file {
+        Some(file) => std::path::PathBuf::from(file),
+        None => api
+            .model("EleutherAI/gpt-neox-20b".to_string())
+            .get("tokenizer.json")?,
+    };
+    let config_filename = match args.config_file {
+        Some(file) => std::path::PathBuf::from(file),
+        None => repo.get("config.json")?,
+    };
+    let filenames = match args.weight_files {
+        Some(files) => files
+            .split(',')
+            .map(std::path::PathBuf::from)
+            .collect::<Vec<_>>(),
+        None => {
+            vec![repo.get("model.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 = serde_json::from_slice(&std::fs::read(config_filename)?)?;
+    let device = candle_examples::device(args.cpu)?;
+    let vb = unsafe { VarBuilder::from_mmaped_safetensors(&filenames, DType::F32, &device)? };
+    let model = Model::new(&config, vb.pp("backbone"))?;
+    println!("loaded the model in {:?}", start.elapsed());
+
+    let mut pipeline = TextGeneration::new(
+        model,
+        config,
+        tokenizer,
+        args.seed,
+        args.temperature,
+        args.top_p,
+        args.repeat_penalty,
+        args.repeat_last_n,
+        &device,
+    );
+    pipeline.run(&args.prompt, args.sample_len)?;
+    Ok(())
+}