Local LLM Backend

Cargo.toml

@@ -4,6 +4,7 @@ members = [
     "llm-chain-openai",
     "llm-chain-llama",
     "llm-chain-llama/sys",
+    "llm-chain-local",
     "llm-chain-qdrant",
 ]
 

llm-chain-local/Cargo.toml

@@ -0,0 +1,23 @@
+[package]
+name = "llm-chain-local"
+version = "0.9.0"
+edition = "2021"
+description = "Use `llm-chain` with a local [`llm`](https://github.com/rustformers/llm) backend."
+license = "MIT"
+keywords = ["llm", "langchain", "ggml", "chain"]
+categories = ["science"]
+authors = ["Dan Forbes <dan@danforbes.dev>"]
+repository = "https://github.com/sobelio/llm-chain/"
+
+# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
+
+[dependencies]
+async-trait = "0.1.68"
+llm = { git = "https://github.com/danforbes/llama-rs", branch = "dfo/chore/llm-chain" }
+llm-chain = { path = "../llm-chain", version = "0.9.0", default-features = false }
+rand = "0.8.5"
+serde = { version = "1.0.160", features = ["derive"] }
+thiserror = "1.0.40"
+
+[dev-dependencies]
+tokio = { version = "1.28.0", features = ["macros", "rt"] }

llm-chain-local/examples/simple.rs

@@ -0,0 +1,37 @@
+use std::env::args;
+
+use llm_chain::traits::Executor;
+
+extern crate llm_chain_local;
+
+/// This example demonstrates how to use the llm-chain-local crate to generate text using a model.
+///
+/// Usage: cargo run --release --package llm-chain-local --example simple model_type path/to/model
+///
+/// For example, if the model is a LLaMA-type model located at "/models/llama"
+/// cargo run --release --package llm-chain-local --example simple llama /models/llama
+/// 
+/// An optional third argument can be used to customize the prompt passed to the model.
+#[tokio::main(flavor = "current_thread")]
+async fn main() -> Result<(), Box<dyn std::error::Error>> {
+    let raw_args: Vec<String> = args().collect();
+    let args = match &raw_args.len() {
+      3 => (raw_args[1].as_str(), raw_args[2].as_str(), "Rust is a cool programming language because"),
+      4 => (raw_args[1].as_str(), raw_args[2].as_str(), raw_args[3].as_str()),
+      _ => panic!("Usage: cargo run --release --example inference <model type> <path to model> <optional prompt>")
+    };
+
+    let model_type = args.0;
+    let model_path = args.1;
+    let prompt = args.2;
+
+    let exec_opts = llm_chain_local::options::PerExecutor {
+        model_path: Some(String::from(model_path)),
+        model_type: Some(String::from(model_type)),
+    };
+    let exec = llm_chain_local::Executor::new_with_options(Some(exec_opts), None)?;
+
+    let res = exec.execute(None, &llm_chain::prompt::Data::Text(String::from(prompt))).await?;
+    println!("{}", res);
+    Ok(())
+}

llm-chain-local/src/executor.rs

@@ -0,0 +1,195 @@
+use std::path::Path;
+
+use crate::options::{PerExecutor, PerInvocation};
+use crate::output::Output;
+use crate::LocalLlmTextSplitter;
+
+use async_trait::async_trait;
+use llm_chain::prompt::Prompt;
+use llm_chain::tokens::{PromptTokensError, Tokenizer, TokenizerError};
+use llm_chain::traits::ExecutorCreationError;
+
+/// Executor is responsible for running the LLM and managing its context.
+pub struct Executor {
+    llm: Box<dyn llm::Model>,
+    callback: Option<fn(&Output)>,
+}
+
+impl Executor {
+    pub fn with_callback(mut self, callback: fn(&Output)) -> Self {
+        self.callback = Some(callback);
+        self
+    }
+
+    pub(crate) fn get_llm(&self) -> &dyn llm::Model {
+        self.llm.as_ref()
+    }
+}
+
+#[derive(Debug, thiserror::Error)]
+pub enum Error {
+    #[error("unable to tokenize prompt")]
+    PromptTokensError(PromptTokensError),
+    #[error("unable to create executor: {0}")]
+    InnerError(#[from] Box<dyn std::error::Error>),
+}
+
+impl llm_chain::traits::ExecutorError for Error {}
+
+#[async_trait]
+impl llm_chain::traits::Executor for Executor {
+    type PerInvocationOptions = PerInvocation;
+    type PerExecutorOptions = PerExecutor;
+    type Output = Output;
+    type Error = Error;
+    type Token = i32;
+    type StepTokenizer<'a> = LocalLlmTokenizer<'a>;
+    type TextSplitter<'a> = LocalLlmTextSplitter<'a>;
+
+    fn new_with_options(
+        options: Option<Self::PerExecutorOptions>,
+        invocation_options: Option<Self::PerInvocationOptions>,
+    ) -> Result<Self, llm_chain::traits::ExecutorCreationError> {
+        let model_type = options
+            .as_ref()
+            .and_then(|x| x.model_type.clone())
+            .or_else(|| std::env::var("LLM_MODEL_TYPE").ok())
+            .ok_or(ExecutorCreationError::FieldRequiredError(
+                "model_type, ensure to provide the parameter or set `LLM_MODEL_TYPE` environment variable ".to_string(),
+            ))?;
+        let model_path = options
+                .as_ref()
+                .and_then(|x| x.model_path.clone())
+                .or_else(|| std::env::var("LLM_MODEL_PATH").ok())
+                .ok_or(ExecutorCreationError::FieldRequiredError(
+                    "model_path, ensure to provide the parameter or set `LLM_MODEL_PATH` environment variable ".to_string(),
+                ))?;
+
+        let params = invocation_options.unwrap_or_default().into();
+        let llm: Box<dyn llm::Model> = match model_type.as_str() {
+            "bloom" => load::<llm::models::Bloom>(&model_path, params),
+            "gpt2" => load::<llm::models::Gpt2>(&model_path, params),
+            "gptj" => load::<llm::models::GptJ>(&model_path, params),
+            "llama" => load::<llm::models::Llama>(&model_path, params),
+            "neox" => load::<llm::models::NeoX>(&model_path, params),
+            m => Err(llm::LoadError::InvariantBroken {
+                path: None,
+                invariant: format!("Unsupported model type {m}"),
+            }),
+        }
+        .map_err(|e| ExecutorCreationError::InnerError(Box::new(e)))?;
+
+        Ok(Executor {
+            llm,
+            callback: None,
+        })
+    }
+
+    async fn execute(
+        &self,
+        _: Option<&Self::PerInvocationOptions>,
+        prompt: &Prompt,
+    ) -> Result<Self::Output, Self::Error> {
+        let session = &mut self.llm.start_session(Default::default());
+        let mut output = String::new();
+        session
+            .infer::<std::convert::Infallible>(
+                self.llm.as_ref(),
+                prompt.to_text().as_str(),
+                // EvaluateOutputRequest
+                &mut Default::default(),
+                &mut rand::thread_rng(),
+                |t| {
+                    output.push_str(t);
+
+                    Ok(())
+                },
+            )
+            .map_err(|e| Error::InnerError(Box::new(e)))?;
+
+        Ok(output.into())
+    }
+
+    fn tokens_used(
+        &self,
+        options: Option<&Self::PerInvocationOptions>,
+        prompt: &llm_chain::prompt::Prompt,
+    ) -> Result<llm_chain::tokens::TokenCount, llm_chain::tokens::PromptTokensError> {
+        let tokenizer = self.get_tokenizer(options)?;
+        let input = prompt.to_text();
+
+        let tokens_used = tokenizer
+            .tokenize_str(&input)
+            .map_err(|_e| PromptTokensError::UnableToCompute)?
+            .len() as i32;
+        let max_tokens = self.max_tokens_allowed(options);
+        Ok(llm_chain::tokens::TokenCount::new(max_tokens, tokens_used))
+    }
+
+    fn max_tokens_allowed(&self, _: Option<&Self::PerInvocationOptions>) -> i32 {
+        self.llm.n_context_tokens().try_into().unwrap_or(2048)
+    }
+
+    fn get_tokenizer(
+        &self,
+        _: Option<&Self::PerInvocationOptions>,
+    ) -> Result<Self::StepTokenizer<'_>, llm_chain::tokens::TokenizerError> {
+        Ok(LocalLlmTokenizer::new(self))
+    }
+
+    fn get_text_splitter(
+        &self,
+        _: Option<&Self::PerInvocationOptions>,
+    ) -> Result<Self::TextSplitter<'_>, Self::Error> {
+        Ok(LocalLlmTextSplitter::new(self))
+    }
+}
+
+pub struct LocalLlmTokenizer<'a> {
+    llm: &'a dyn llm::Model,
+}
+
+impl<'a> LocalLlmTokenizer<'a> {
+    pub fn new(executor: &'a Executor) -> Self {
+        LocalLlmTokenizer {
+            llm: executor.llm.as_ref(),
+        }
+    }
+}
+
+impl Tokenizer<llm::TokenId> for LocalLlmTokenizer<'_> {
+    fn tokenize_str(&self, doc: &str) -> Result<Vec<llm::TokenId>, TokenizerError> {
+        match &self.llm.vocabulary().tokenize(doc, false) {
+            Ok(tokens) => Ok(tokens.into_iter().map(|t| t.1).collect()),
+            Err(_) => Err(TokenizerError::TokenizationError),
+        }
+    }
+
+    fn to_string(&self, tokens: Vec<llm::TokenId>) -> Result<String, TokenizerError> {
+        let mut res = String::new();
+        let mut token_utf8_buf = llm::TokenUtf8Buffer::new();
+        for token_id in tokens {
+            // Buffer the token until it's valid UTF-8, then call the callback.
+            if let Some(tokens) =
+                token_utf8_buf.push(self.llm.vocabulary().token(token_id as usize))
+            {
+                res.push_str(&tokens)
+            }
+        }
+
+        Ok(res)
+    }
+}
+
+pub fn load<M: llm::KnownModel + 'static>(
+    model_path: &str,
+    params: llm::ModelParameters,
+) -> Result<Box<dyn llm::Model>, llm::LoadError> {
+    let model = llm::load::<M>(
+        Path::new(model_path),
+        params,
+        llm::load_progress_callback_stdout,
+    )?;
+
+    Ok(Box::new(model))
+}

llm-chain-local/src/lib.rs

@@ -0,0 +1,7 @@
+mod executor;
+pub mod options;
+mod output;
+mod text_splitter;
+
+pub use executor::Executor;
+pub use text_splitter::LocalLlmTextSplitter;

llm-chain-local/src/options.rs

@@ -0,0 +1,87 @@
+use llm::{InferenceParameters, InferenceWithPromptParameters, ModelParameters};
+use llm_chain::traits::Options;
+use serde::{Deserialize, Serialize};
+
+#[derive(Debug, Clone, Default, Serialize, Deserialize)]
+/// An overridable collection of configuration parameters for an LLM. It is combined with a prompt to create an invocation.
+pub struct PerInvocation {
+    pub n_threads: Option<usize>,
+    pub n_tok_predict: Option<usize>,
+    pub top_k: Option<usize>,
+    pub top_p: Option<f32>,
+    pub temp: Option<f32>,
+    pub repeat_penalty: Option<f32>,
+    pub stop_sequence: Option<String>,
+}
+
+impl Options for PerInvocation {}
+
+impl Into<ModelParameters> for PerInvocation {
+    fn into(self) -> ModelParameters {
+        let inference_params = InferenceParameters {
+            n_threads: self.n_threads.unwrap_or(4),
+            n_batch: 8,
+            top_k: self.top_k.unwrap_or(40),
+            top_p: self.top_p.unwrap_or(0.95),
+            repeat_penalty: self.temp.unwrap_or(1.3),
+            temperature: self.temp.unwrap_or(0.8),
+            bias_tokens: Default::default(),
+        };
+
+        let prompt_params = InferenceWithPromptParameters {
+            play_back_previous_tokens: false,
+            maximum_token_count: None,
+        };
+
+        ModelParameters {
+            prefer_mmap: true,
+            n_context_tokens: 2048,
+            inference_params,
+            inference_prompt_params: prompt_params,
+        }
+    }
+}
+
+/// `PerExecutor` represents a collection of configuration parameters for the executor of the LLM.
+/// It contains optional fields for the model path and context parameters.
+///
+/// # Examples
+///
+/// ```
+/// use llm_chain_local::PerExecutor;
+/// let executor = PerExecutor::new().with_model_path("path/to/model");
+/// ```
+#[derive(Debug, Clone, Serialize, Deserialize, Default)]
+pub struct PerExecutor {
+    /// Optional path to the LLM.
+    pub model_path: Option<String>,
+    /// Optional type (e.g. LLaMA, GPT-Neo-X) of the LLM.
+    pub model_type: Option<String>,
+}
+
+impl PerExecutor {
+    /// Creates a new `PerExecutor` instance with default values.
+    ///
+    /// # Returns
+    ///
+    /// A `PerExecutor` instance with default values for the model path and context parameters.
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    /// Sets the model path for the current `PerExecutor` instance.
+    ///
+    /// # Arguments
+    ///
+    /// * `model_path` - The path to the LLM.
+    ///
+    /// # Returns
+    ///
+    /// A new `PerExecutor` instance with the updated model path.
+    pub fn with_model_path(mut self, model_path: &str) -> Self {
+        self.model_path = Some(model_path.to_string());
+        self
+    }
+}
+
+impl Options for PerExecutor {}