Merge pull request #67 from EricLBuehler/develop

LLaMa3.1 chat completion

README.md

@@ -19,7 +19,7 @@ Currently, candle-vllm supports chat serving for the following models.
 
 | Model ID | Model Type | Supported | Speed (A100, BF16)
 |--|--|--|--|
-| #1 | **LLAMA/LLAMA2/LLaMa3** |✅|74 tks/s (7B)|
+| #1 | **LLAMA/LLAMA2/LLaMa3/LLaMa3.1** |✅|74 tks/s (7B), 65 tks/s (LLaMa3.1 8B)|
 | #2 | **Mistral** |✅|70 tks/s (7B)|
 | #3 | **Phi (v1, v1.5, v2)** |✅|97 tks/s (2.7B, F32+BF16)|
 | #4 | **Phi-3 （3.8B, 7B）** |✅|107 tks/s (3.8B)|
@@ -55,6 +55,11 @@ You may also run specific model using huggingface model-id, e.g.,
 cargo run --release -- --port 2000 --model-id meta-llama/Llama-2-7b-chat-hf llama --repeat-last-n 64
 ```
 
+Run latest LLaMa3.1 using local weights
+
+```
+cargo run --release -- --port 2000 --weight-path /home/Meta-Llama-3.1-8B-Instruct/ llama3 --repeat-last-n 64
+```
 ### Step 2:
 
 #### Option 1: Chat with ChatUI (recommended)

src/lib.rs

@@ -25,6 +25,22 @@ pub enum ModelSelected {
         max_gen_tokens: Option<usize>,
     },
 
+    /// Select the llama3 model (default llama3.1-8b).
+    Llama3 {
+        /// Control the application of repeat penalty for the last n tokens
+        #[arg(long)]
+        repeat_last_n: Option<usize>,
+
+        #[arg(long)]
+        temperature: Option<f32>,
+
+        #[arg(long)]
+        penalty: Option<f32>,
+
+        #[arg(long)]
+        max_gen_tokens: Option<usize>,
+    },
+
     /// Select the phi2 model (default 2.7b).
     Phi2 {
         /// Control the application of repeat penalty for the last n tokens
@@ -159,6 +175,12 @@ impl ToString for ModelSelected {
                 penalty: _,
                 max_gen_tokens: _,
             } => "llama".to_string(),
+            ModelSelected::Llama3 {
+                repeat_last_n: _,
+                temperature: _,
+                penalty: _,
+                max_gen_tokens: _,
+            } => "llama3".to_string(),
             ModelSelected::Phi2 {
                 repeat_last_n: _,
                 temperature: _,
@@ -237,6 +259,29 @@ pub fn get_model_loader(
                 "meta-llama/Llama-2-7b-chat-hf".to_string()
             },
         ),
+        ModelSelected::Llama3 {
+            repeat_last_n,
+            temperature,
+            penalty,
+            max_gen_tokens,
+        } => (
+            Box::new(DefaultLoader::new(
+                SpecificConfig::new(
+                    repeat_last_n,
+                    temperature,
+                    None,
+                    None,
+                    penalty,
+                    max_gen_tokens,
+                ),
+                "llama3".to_string(),
+            )),
+            if model_id.is_some() {
+                model_id.unwrap()
+            } else {
+                "meta-llama/Meta-Llama-3.1-8B-Instruct".to_string()
+            },
+        ),
         ModelSelected::Phi2 {
             repeat_last_n,
             temperature,

src/openai/conversation/default_conversation.rs

@@ -17,6 +17,7 @@ pub enum SeparatorStyle {
     NoColonTwo,
     AddNewLineSingle,
     Llama,
+    Llama3,
     Phi,
     Qwen2,
     Gemma,
@@ -248,6 +249,32 @@ impl Conversation for DefaultConversation {
                 accum
             }
 
+            SeparatorStyle::Llama3 => {
+                let mut accum = "<|begin_of_text|>".to_string();
+                for (i, message) in self.messages.iter().enumerate() {
+                    let Message((_role, message)) = message;
+                    if _role.clone() == self.roles.0 {
+                        //user message
+                        if let Some(message) = message {
+                            accum += &format!(
+                                "<|start_header_id|>user<|end_header_id|>\n\n {message} <|eot_id|>"
+                            );
+                        } else {
+                            accum +=
+                                &format!("<|start_header_id|>user<|end_header_id|>\n\n <|eot_id|>");
+                        }
+                    } else if _role.clone() == self.roles.1 {
+                        //assistant message
+                        if let Some(message) = message {
+                            accum += &format!("<|start_header_id|>assistant<|end_header_id|>\n\n {message} <|eot_id|>");
+                        }
+                    } else if i == 0 && !system_prompt.is_empty() {
+                        accum += &system_prompt;
+                    }
+                }
+                accum
+            }
+
             SeparatorStyle::Phi => {
                 let mut accum = "".to_string();
                 for (i, message) in self.messages.iter().enumerate() {

src/openai/models/gemma.rs

@@ -5,9 +5,9 @@ use candle::{DType, Device, Module, Result, Tensor, D};
 use candle_core as candle;
 use candle_nn::Activation;
 use candle_nn::{linear_b, linear_no_bias as linear, Linear, RmsNorm, VarBuilder};
+use either::Either;
 use std::iter::zip;
 use std::sync::Arc;
-
 #[derive(serde::Deserialize, Debug, Clone)]
 pub struct GemmaConfig {
     pub attention_bias: bool,
@@ -45,8 +45,8 @@ impl GemmaConfig {
             rms_norm_eps: self.rms_norm_eps,
             rope_theta: self.rope_theta,
             use_flash_attn,
-            bos_token_id: Some(self.bos_token_id as u32),
-            eos_token_id: Some(self.eos_token_id as u32),
+            bos_token_id: super::TokenID(Either::Left(Some(self.bos_token_id as u32))),
+            eos_token_id: super::TokenID(Either::Left(Some(self.eos_token_id as u32))),
             max_seq_len: self.max_position_embeddings.unwrap_or(4096),
             sliding_window: None,
             hidden_act: hidden_act,

src/openai/models/llama.rs

@@ -6,6 +6,7 @@ use candle_core as candle;
 use candle_nn::{embedding, Embedding, Module, VarBuilder};
 use candle_transformers::models::with_tracing::{linear_no_bias as linear, Linear, RmsNorm};
 pub const MAX_SEQ_LEN: usize = 4096;
+use crate::openai::models::TokenID;
 use std::iter::zip;
 #[derive(Debug, Clone, serde::Deserialize)]
 pub struct LlamaConfig {
@@ -18,8 +19,9 @@ pub struct LlamaConfig {
     pub rms_norm_eps: f64,
     #[serde(default = "default_rope")]
     pub rope_theta: f32,
-    pub bos_token_id: Option<u32>,
-    pub eos_token_id: Option<u32>,
+    pub bos_token_id: TokenID,
+    pub eos_token_id: TokenID,
+    pub max_position_embeddings: Option<usize>,
 }
 
 fn default_rope() -> f32 {
@@ -40,7 +42,7 @@ impl LlamaConfig {
             use_flash_attn,
             bos_token_id: self.bos_token_id,
             eos_token_id: self.eos_token_id,
-            max_seq_len: MAX_SEQ_LEN,
+            max_seq_len: self.max_position_embeddings.unwrap_or(MAX_SEQ_LEN),
             sliding_window: None,
             hidden_act: None,
             tie_word_embeddings: false,

src/openai/models/mistral.rs

@@ -4,6 +4,7 @@ use crate::paged_attention::PagedAttention;
 use candle_core::{DType, Device, Module, Result, Tensor, D};
 use candle_nn::{Activation, VarBuilder};
 use candle_transformers::models::with_tracing::{linear_no_bias, Linear, RmsNorm};
+use either::Either;
 use std::iter::zip;
 use std::sync::Arc;
 
@@ -37,8 +38,8 @@ impl MistralConfig {
             rms_norm_eps: self.rms_norm_eps,
             rope_theta: self.rope_theta,
             use_flash_attn,
-            bos_token_id: Some(self.bos_token_id as u32),
-            eos_token_id: Some(self.eos_token_id as u32),
+            bos_token_id: super::TokenID(Either::Left(Some(self.bos_token_id as u32))),
+            eos_token_id: super::TokenID(Either::Left(Some(self.eos_token_id as u32))),
             max_seq_len: self.max_position_embeddings,
             sliding_window: self.sliding_window,
             hidden_act: Some(self.hidden_act),

src/openai/models/mod.rs

@@ -14,6 +14,11 @@ use std::collections::HashMap;
 #[derive(Deserialize, Debug, Clone)]
 pub struct RopeScaling(#[serde(with = "either::serde_untagged")] pub Either<Vec<f64>, String>);
 
+#[derive(Deserialize, Debug, Clone)]
+pub struct TokenID(
+    #[serde(with = "either::serde_untagged")] pub Either<Option<u32>, Option<Vec<u32>>>,
+);
+
 #[derive(Debug, Clone)]
 pub struct Config {
     pub hidden_size: usize,
@@ -25,8 +30,8 @@ pub struct Config {
     pub use_flash_attn: bool,
     pub rms_norm_eps: f64,
     pub rope_theta: f64,
-    pub bos_token_id: Option<u32>,
-    pub eos_token_id: Option<u32>,
+    pub bos_token_id: TokenID,
+    pub eos_token_id: TokenID,
     pub max_seq_len: usize,
     pub sliding_window: Option<usize>,
     pub hidden_act: Option<candle_nn::Activation>,

src/openai/models/phi2.rs

@@ -6,6 +6,7 @@ use candle_nn::{Activation, VarBuilder};
 use candle_transformers::models::with_tracing::{
     layer_norm, linear_no_bias as linear, Embedding, LayerNorm, Linear,
 };
+use either::Either;
 use serde::Deserialize;
 use std::iter::zip;
 
@@ -41,8 +42,8 @@ impl Phi2Config {
             num_key_value_heads: self.num_key_value_heads.unwrap_or(self.num_attention_heads),
             rms_norm_eps: self.layer_norm_eps,
             rope_theta: self.rope_theta,
-            bos_token_id: self.bos_token_id,
-            eos_token_id: self.eos_token_id,
+            bos_token_id: super::TokenID(Either::Left(self.bos_token_id)),
+            eos_token_id: super::TokenID(Either::Left(self.eos_token_id)),
             max_seq_len: self.max_position_embeddings,
             sliding_window: self.sliding_window,
             hidden_act: Some(self.hidden_act),

src/openai/models/phi3.rs

@@ -43,8 +43,8 @@ impl PhiConfig {
             rms_norm_eps: self.rms_norm_eps,
             rope_theta: self.rope_theta,
             use_flash_attn,
-            bos_token_id: self.bos_token_id,
-            eos_token_id: self.eos_token_id,
+            bos_token_id: super::TokenID(Either::Left(self.bos_token_id)),
+            eos_token_id: super::TokenID(Either::Left(self.eos_token_id)),
             max_seq_len: self.max_position_embeddings,
             sliding_window: self.sliding_window,
             hidden_act: Some(self.hidden_act),

src/openai/models/qwen2.rs

@@ -5,6 +5,7 @@ use candle::{DType, Device, Module, Result, Tensor, D};
 use candle_core as candle;
 use candle_nn::VarBuilder;
 use candle_transformers::models::with_tracing::{linear, linear_no_bias, Linear, RmsNorm};
+use either::Either;
 use std::iter::zip;
 use std::sync::Arc;
 
@@ -40,8 +41,8 @@ impl QwenConfig {
             rms_norm_eps: self.rms_norm_eps,
             rope_theta: self.rope_theta,
             use_flash_attn,
-            bos_token_id: Some(self.bos_token_id as u32),
-            eos_token_id: Some(self.eos_token_id as u32),
+            bos_token_id: super::TokenID(Either::Left(Some(self.bos_token_id as u32))),
+            eos_token_id: super::TokenID(Either::Left(Some(self.bos_token_id as u32))),
             max_seq_len: self.max_position_embeddings,
             sliding_window: Some(self.sliding_window),
             hidden_act: Some(self.hidden_act),

src/openai/models/stable_lm.rs

@@ -4,6 +4,7 @@ use crate::paged_attention::PagedAttention;
 use candle_core::{DType, Device, Module, Result, Tensor, D};
 use candle_nn::{Activation, LayerNorm, VarBuilder};
 use candle_transformers::models::with_tracing::{linear, linear_no_bias, Linear};
+use either::Either;
 use std::iter::zip;
 use std::sync::Arc;
 
@@ -41,8 +42,8 @@ impl StableLMConfig {
             rms_norm_eps: self.norm_eps,
             rope_theta: self.rope_theta,
             use_flash_attn,
-            bos_token_id: Some(self.bos_token_id as u32),
-            eos_token_id: Some(self.eos_token_id as u32),
+            bos_token_id: super::TokenID(Either::Left(Some(self.bos_token_id as u32))),
+            eos_token_id: super::TokenID(Either::Left(Some(self.bos_token_id as u32))),
             max_seq_len: self.max_position_embeddings,
             sliding_window: self.sliding_window,
             hidden_act: Some(self.hidden_act),

src/openai/models/yi.rs

@@ -4,6 +4,7 @@ use crate::paged_attention::PagedAttention;
 use candle_core::{DType, Device, Module, Result, Tensor, D};
 use candle_nn::{Activation, VarBuilder};
 use candle_transformers::models::with_tracing::{linear_no_bias, Linear, RmsNorm};
+use either::Either;
 use std::iter::zip;
 use std::sync::Arc;
 
@@ -37,8 +38,8 @@ impl YiConfig {
             rms_norm_eps: self.rms_norm_eps,
             rope_theta: self.rope_theta,
             use_flash_attn,
-            bos_token_id: Some(self.bos_token_id as u32),
-            eos_token_id: Some(self.eos_token_id as u32),
+            bos_token_id: super::TokenID(Either::Left(Some(self.bos_token_id as u32))),
+            eos_token_id: super::TokenID(Either::Left(Some(self.bos_token_id as u32))),
             max_seq_len: self.max_position_embeddings,
             sliding_window: self.sliding_window,
             hidden_act: Some(self.hidden_act),

src/openai/pipelines/pipeline.rs

@@ -1,4 +1,5 @@
 use super::{get_token, ModelLoader, ModelPaths, ModulePipeline, TokenOrFinishReason};
+use crate::openai::models::TokenID;
 use crate::openai::sampling_params::{Logprobs, TopLogprob};
 use crate::scheduler::sequence::SequenceGroup;
 use crate::{
@@ -31,6 +32,7 @@ use candle_core::{DType, Device, Tensor};
 use candle_examples::token_output_stream::TokenOutputStream;
 use candle_nn::VarBuilder;
 use candle_transformers::generation::{LogitsProcessor, Sampling};
+use either::Either;
 use either::Either::{Left, Right};
 use hf_hub::{api::sync::ApiBuilder, Repo, RepoType};
 use std::{path::PathBuf, sync::Arc};
@@ -170,7 +172,7 @@ impl ModelLoader for DefaultLoader {
         let specific_args = self.config.clone();
 
         let config = match self.name.as_str() {
-            "llama" => {
+            "llama" | "llama3" => {
                 let config: LlamaConfig = try_api!(serde_json::from_slice(&try_api!(
                     std::fs::read(paths.get_config_filename())
                 ),));
@@ -238,6 +240,10 @@ impl ModelLoader for DefaultLoader {
                 LLMModel::LLAMA(try_api!(Llama::load(vb, &config, dtype, &device))),
                 SeparatorStyle::Llama,
             ),
+            "llama3" => (
+                LLMModel::LLAMA(try_api!(Llama::load(vb, &config, dtype, &device))),
+                SeparatorStyle::Llama3,
+            ),
             "phi2" => (
                 LLMModel::Phi2(try_api!(Phi2::new(vb, &config, dtype, &device))),
                 SeparatorStyle::Phi,
@@ -296,13 +302,32 @@ impl ModelLoader for DefaultLoader {
 
         println!("{:?}", pipeline_config);
 
-        let eos_token = match tokenizer.get_token("<|endoftext|>") {
-            Some(token) => token,
-            None => tokenizer.tokenizer().token_to_id(EOS_TOKEN).unwrap(),
-        };
         let mut stop_token_ids = Vec::<u32>::new();
-        stop_token_ids.push(eos_token);
 
+        match &config.eos_token_id {
+            //eos_token defined in the config
+            TokenID(Either::Left(eos_token)) => {
+                if let Some(tk) = eos_token {
+                    stop_token_ids.push(*tk);
+                }
+            }
+            TokenID(Either::Right(eos_token_list)) => {
+                if let Some(tks) = eos_token_list {
+                    stop_token_ids.extend(tks)
+                }
+            }
+        }
+
+        if stop_token_ids.len() == 0 {
+            //if no eos_token defined in the config, use default
+            let eos_token = match tokenizer.get_token("<|endoftext|>") {
+                Some(token) => token,
+                _ => tokenizer.tokenizer().token_to_id(EOS_TOKEN).unwrap_or(0),
+            };
+            stop_token_ids.push(eos_token);
+        }
+
+        //custome stop tokens
         if let Some(custom_stop) = &config.custom_stop_tokens {
             for stop in custom_stop {
                 match tokenizer.get_token(&stop) {