Add the phi-3 model. (#2120)

* Add the phi-3 model.

* Faster rope.

* Bugfix.

* Fix the detokenization.

candle-examples/examples/phi/main.rs

@@ -7,11 +7,13 @@ extern crate accelerate_src;
 use anyhow::{Error as E, Result};
 use clap::{Parser, ValueEnum};
 
+use candle_examples::token_output_stream::TokenOutputStream;
 use candle_transformers::models::mixformer::{Config, MixFormerSequentialForCausalLM as MixFormer};
 use candle_transformers::models::phi::{Config as PhiConfig, Model as Phi};
+use candle_transformers::models::phi3::{Config as Phi3Config, Model as Phi3};
 use candle_transformers::models::quantized_mixformer::MixFormerSequentialForCausalLM as QMixFormer;
 
-use candle::{DType, Device, Tensor};
+use candle::{DType, Device, IndexOp, Tensor};
 use candle_nn::VarBuilder;
 use candle_transformers::generation::LogitsProcessor;
 use hf_hub::{api::sync::Api, Repo, RepoType};
@@ -20,13 +22,14 @@ use tokenizers::Tokenizer;
 enum Model {
     MixFormer(MixFormer),
     Phi(Phi),
+    Phi3(Phi3),
     Quantized(QMixFormer),
 }
 
 struct TextGeneration {
     model: Model,
     device: Device,
-    tokenizer: Tokenizer,
+    tokenizer: TokenOutputStream,
     logits_processor: LogitsProcessor,
     repeat_penalty: f32,
     repeat_last_n: usize,
@@ -49,7 +52,7 @@ impl TextGeneration {
         let logits_processor = LogitsProcessor::new(seed, temp, top_p);
         Self {
             model,
-            tokenizer,
+            tokenizer: TokenOutputStream::new(tokenizer),
             logits_processor,
             repeat_penalty,
             repeat_last_n,
@@ -61,7 +64,11 @@ impl TextGeneration {
     fn run(&mut self, prompt: &str, sample_len: usize) -> Result<()> {
         use std::io::Write;
         println!("starting the inference loop");
-        let tokens = self.tokenizer.encode(prompt, true).map_err(E::msg)?;
+        let tokens = self
+            .tokenizer
+            .tokenizer()
+            .encode(prompt, true)
+            .map_err(E::msg)?;
         if tokens.is_empty() {
             anyhow::bail!("Empty prompts are not supported in the phi model.")
         }
@@ -73,13 +80,14 @@ impl TextGeneration {
         }
         let mut tokens = tokens.get_ids().to_vec();
         let mut generated_tokens = 0usize;
-        let eos_token = match self.tokenizer.get_vocab(true).get("<|endoftext|>") {
+        let eos_token = match self.tokenizer.get_token("<|endoftext|>") {
-            Some(token) => *token,
+            Some(token) => token,
             None => anyhow::bail!("cannot find the endoftext token"),
         };
         print!("{prompt}");
         std::io::stdout().flush()?;
         let start_gen = std::time::Instant::now();
+        let mut pos = 0;
         for index in 0..sample_len {
             let context_size = if index > 0 { 1 } else { tokens.len() };
             let ctxt = &tokens[tokens.len().saturating_sub(context_size)..];
@@ -88,6 +96,7 @@ impl TextGeneration {
                 Model::MixFormer(m) => m.forward(&input)?,
                 Model::Phi(m) => m.forward(&input)?,
                 Model::Quantized(m) => m.forward(&input)?,
+                Model::Phi3(m) => m.forward(&input, pos)?.i((.., 0, ..))?,
             };
             let logits = logits.squeeze(0)?.to_dtype(DType::F32)?;
             let logits = if self.repeat_penalty == 1. {
@@ -107,10 +116,12 @@ impl TextGeneration {
             if next_token == eos_token {
                 break;
             }
-            let token = self.tokenizer.decode(&[next_token], true).map_err(E::msg)?;
+            if let Some(t) = self.tokenizer.next_token(next_token)? {
-            print!("{token}");
+                print!("{t}");
                 std::io::stdout().flush()?;
             }
+            pos += context_size;
+        }
         let dt = start_gen.elapsed();
         println!(
             "\n{generated_tokens} tokens generated ({:.2} token/s)",
@@ -128,6 +139,8 @@ enum WhichModel {
     V1_5,
     #[value(name = "2")]
     V2,
+    #[value(name = "3")]
+    V3,
     #[value(name = "2-old")]
     V2Old,
     PuffinPhiV2,
@@ -236,6 +249,7 @@ fn main() -> Result<()> {
                     WhichModel::V1 => "microsoft/phi-1".to_string(),
                     WhichModel::V1_5 => "microsoft/phi-1_5".to_string(),
                     WhichModel::V2 | WhichModel::V2Old => "microsoft/phi-2".to_string(),
+                    WhichModel::V3 => "microsoft/Phi-3-mini-4k-instruct".to_string(),
                     WhichModel::PuffinPhiV2 | WhichModel::PhiHermes => {
                         "lmz/candle-quantized-phi".to_string()
                     }
@@ -253,9 +267,10 @@ fn main() -> Result<()> {
                     WhichModel::V1 => "refs/pr/8".to_string(),
                     WhichModel::V1_5 => "refs/pr/73".to_string(),
                     WhichModel::V2Old => "834565c23f9b28b96ccbeabe614dd906b6db551a".to_string(),
-                    WhichModel::V2 | WhichModel::PuffinPhiV2 | WhichModel::PhiHermes => {
+                    WhichModel::V2
-                        "main".to_string()
+                    | WhichModel::V3
-                    }
+                    | WhichModel::PuffinPhiV2
+                    | WhichModel::PhiHermes => "main".to_string(),
                 }
             }
         }
@@ -264,9 +279,11 @@ fn main() -> Result<()> {
     let tokenizer_filename = match args.tokenizer {
         Some(file) => std::path::PathBuf::from(file),
         None => match args.model {
-            WhichModel::V1 | WhichModel::V1_5 | WhichModel::V2 | WhichModel::V2Old => {
+            WhichModel::V1
-                repo.get("tokenizer.json")?
+            | WhichModel::V1_5
-            }
+            | WhichModel::V2
+            | WhichModel::V2Old
+            | WhichModel::V3 => repo.get("tokenizer.json")?,
             WhichModel::PuffinPhiV2 | WhichModel::PhiHermes => {
                 repo.get("tokenizer-puffin-phi-v2.json")?
             }
@@ -282,14 +299,19 @@ fn main() -> Result<()> {
                     WhichModel::V2 | WhichModel::V2Old => vec![repo.get("model-v2-q4k.gguf")?],
                     WhichModel::PuffinPhiV2 => vec![repo.get("model-puffin-phi-v2-q4k.gguf")?],
                     WhichModel::PhiHermes => vec![repo.get("model-phi-hermes-1_3B-q4k.gguf")?],
+                    WhichModel::V3 => anyhow::bail!(
+                        "use the quantized or quantized-phi examples for quantized phi-v3"
+                    ),
                 }
             } else {
                 match args.model {
                     WhichModel::V1 | WhichModel::V1_5 => vec![repo.get("model.safetensors")?],
-                    WhichModel::V2 | WhichModel::V2Old => candle_examples::hub_load_safetensors(
+                    WhichModel::V2 | WhichModel::V2Old | WhichModel::V3 => {
+                        candle_examples::hub_load_safetensors(
                             &repo,
                             "model.safetensors.index.json",
-                    )?,
+                        )?
+                    }
                     WhichModel::PuffinPhiV2 => vec![repo.get("model-puffin-phi-v2.safetensors")?],
                     WhichModel::PhiHermes => vec![repo.get("model-phi-hermes-1_3B.safetensors")?],
                 }
@@ -306,6 +328,9 @@ fn main() -> Result<()> {
         WhichModel::V2 | WhichModel::V2Old => Config::v2(),
         WhichModel::PuffinPhiV2 => Config::puffin_phi_v2(),
         WhichModel::PhiHermes => Config::phi_hermes_1_3b(),
+        WhichModel::V3 => {
+            panic!("use the quantized or quantized-phi examples for quantized phi-v3")
+        }
     };
     let device = candle_examples::device(args.cpu)?;
     let model = if args.quantized {
@@ -320,7 +345,12 @@ fn main() -> Result<()> {
         };
         Model::Quantized(model)
     } else {
-        let vb = unsafe { VarBuilder::from_mmaped_safetensors(&filenames, DType::F32, &device)? };
+        let dtype = if args.model == WhichModel::V3 && device.is_cuda() {
+            DType::BF16
+        } else {
+            DType::F32
+        };
+        let vb = unsafe { VarBuilder::from_mmaped_safetensors(&filenames, dtype, &device)? };
         match args.model {
             WhichModel::V1 | WhichModel::V1_5 | WhichModel::V2 => {
                 let config_filename = repo.get("config.json")?;
@@ -329,6 +359,13 @@ fn main() -> Result<()> {
                 let phi = Phi::new(&config, vb)?;
                 Model::Phi(phi)
             }
+            WhichModel::V3 => {
+                let config_filename = repo.get("config.json")?;
+                let config = std::fs::read_to_string(config_filename)?;
+                let config: Phi3Config = serde_json::from_str(&config)?;
+                let phi3 = Phi3::new(&config, vb)?;
+                Model::Phi3(phi3)
+            }
             WhichModel::V2Old => {
                 let config = config();
                 Model::MixFormer(MixFormer::new_v2(&config, vb)?)
@@ -421,6 +458,10 @@ fn mmlu<P: AsRef<std::path::Path>>(
                     m.clear_kv_cache();
                     m.forward(&input)?
                 }
+                Model::Phi3(m) => {
+                    m.clear_kv_cache();
+                    m.forward(&input, 0)?
+                }
                 Model::Quantized(m) => {
                     m.clear_kv_cache();
                     m.forward(&input)?

candle-transformers/src/models/mod.rs

@@ -28,6 +28,7 @@ pub mod moondream;
 pub mod mpt;
 pub mod persimmon;
 pub mod phi;
+pub mod phi3;
 pub mod quantized_blip;
 pub mod quantized_blip_text;
 pub mod quantized_llama;

candle-transformers/src/models/phi3.rs

@@ -0,0 +1,329 @@
+// This implementation is based on:
+// https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/blob/main/modeling_phi3.py
+use crate::models::with_tracing::{linear_no_bias as linear, Linear, RmsNorm};
+use candle::{DType, Device, Module, Result, Tensor, D};
+use candle_nn::VarBuilder;
+use std::sync::Arc;
+
+// https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/blob/main/config.json
+#[derive(Debug, Clone, serde::Deserialize)]
+pub struct Config {
+    pub vocab_size: usize,
+    pub hidden_act: candle_nn::Activation,
+    pub hidden_size: usize,
+    pub intermediate_size: usize,
+    pub num_hidden_layers: usize,
+    pub num_attention_heads: usize,
+    pub num_key_value_heads: usize,
+    pub rms_norm_eps: f64,
+    pub rope_theta: f64,
+    pub bos_token_id: Option<u32>,
+    pub eos_token_id: Option<u32>,
+    pub rope_scaling: Option<String>,
+    pub max_position_embeddings: usize,
+}
+
+impl Config {
+    fn head_dim(&self) -> usize {
+        self.hidden_size / self.num_attention_heads
+    }
+}
+
+#[derive(Debug, Clone)]
+struct RotaryEmbedding {
+    sin: Tensor,
+    cos: Tensor,
+}
+
+impl RotaryEmbedding {
+    fn new(dtype: DType, cfg: &Config, dev: &Device) -> Result<Self> {
+        let dim = cfg.head_dim();
+        let max_seq_len = cfg.max_position_embeddings;
+        let inv_freq: Vec<_> = (0..dim)
+            .step_by(2)
+            .map(|i| 1f32 / cfg.rope_theta.powf(i as f64 / dim as f64) as f32)
+            .collect();
+        let inv_freq_len = inv_freq.len();
+        let inv_freq = Tensor::from_vec(inv_freq, (1, inv_freq_len), dev)?.to_dtype(dtype)?;
+        let t = Tensor::arange(0u32, max_seq_len as u32, dev)?
+            .to_dtype(dtype)?
+            .reshape((max_seq_len, 1))?;
+        let freqs = t.matmul(&inv_freq)?;
+        Ok(Self {
+            sin: freqs.sin()?,
+            cos: freqs.cos()?,
+        })
+    }
+
+    fn apply_rotary_emb_qkv(
+        &self,
+        q: &Tensor,
+        k: &Tensor,
+        seqlen_offset: usize,
+    ) -> Result<(Tensor, Tensor)> {
+        let (_b_sz, _h, seq_len, _n_embd) = q.dims4()?;
+        let cos = self.cos.narrow(0, seqlen_offset, seq_len)?;
+        let sin = self.sin.narrow(0, seqlen_offset, seq_len)?;
+        let q_embed = candle_nn::rotary_emb::rope(&q.contiguous()?, &cos, &sin)?;
+        let k_embed = candle_nn::rotary_emb::rope(&k.contiguous()?, &cos, &sin)?;
+        Ok((q_embed, k_embed))
+    }
+}
+
+#[derive(Debug, Clone)]
+struct Attention {
+    qkv_proj: Linear,
+    o_proj: Linear,
+    num_heads: usize,
+    num_kv_heads: usize,
+    num_kv_groups: usize,
+    head_dim: usize,
+    rotary_emb: Arc<RotaryEmbedding>,
+    kv_cache: Option<(Tensor, Tensor)>,
+}
+
+impl Attention {
+    fn new(rotary_emb: Arc<RotaryEmbedding>, cfg: &Config, vb: VarBuilder) -> Result<Self> {
+        let num_heads = cfg.num_attention_heads;
+        let num_kv_heads = cfg.num_key_value_heads;
+        let head_dim = cfg.head_dim();
+        let op_size = num_heads * head_dim + 2 * num_kv_heads * head_dim;
+        let qkv_proj = linear(cfg.hidden_size, op_size, vb.pp("qkv_proj"))?;
+        let o_proj = linear(num_heads * head_dim, cfg.hidden_size, vb.pp("o_proj"))?;
+        Ok(Self {
+            qkv_proj,
+            o_proj,
+            rotary_emb,
+            kv_cache: None,
+            num_heads,
+            num_kv_heads,
+            num_kv_groups: num_heads / num_kv_heads,
+            head_dim,
+        })
+    }
+
+    fn forward(
+        &mut self,
+        xs: &Tensor,
+        attention_mask: Option<&Tensor>,
+        seqlen_offset: usize,
+    ) -> Result<Tensor> {
+        let (b_sz, q_len, _) = xs.dims3()?;
+
+        let qkv = self.qkv_proj.forward(xs)?;
+        let query_pos = self.num_heads * self.head_dim;
+        let query_states = qkv.narrow(D::Minus1, 0, query_pos)?;
+        let key_states = qkv.narrow(D::Minus1, query_pos, self.num_kv_heads * self.head_dim)?;
+        let value_states = qkv.narrow(
+            D::Minus1,
+            query_pos + self.num_kv_heads * self.head_dim,
+            self.num_kv_heads * self.head_dim,
+        )?;
+
+        let query_states = query_states
+            .reshape((b_sz, q_len, self.num_heads, self.head_dim))?
+            .transpose(1, 2)?;
+        let key_states = key_states
+            .reshape((b_sz, q_len, self.num_kv_heads, self.head_dim))?
+            .transpose(1, 2)?;
+        let value_states = value_states
+            .reshape((b_sz, q_len, self.num_kv_heads, self.head_dim))?
+            .transpose(1, 2)?;
+
+        let (query_states, key_states) =
+            self.rotary_emb
+                .apply_rotary_emb_qkv(&query_states, &key_states, seqlen_offset)?;
+
+        let (key_states, value_states) = match &self.kv_cache {
+            None => (key_states, value_states),
+            Some((prev_k, prev_v)) => {
+                let key_states = Tensor::cat(&[prev_k, &key_states], 2)?;
+                let value_states = Tensor::cat(&[prev_v, &value_states], 2)?;
+                (key_states, value_states)
+            }
+        };
+        self.kv_cache = Some((key_states.clone(), value_states.clone()));
+
+        let key_states = crate::utils::repeat_kv(key_states, self.num_kv_groups)?.contiguous()?;
+        let value_states =
+            crate::utils::repeat_kv(value_states, self.num_kv_groups)?.contiguous()?;
+
+        let attn_output = {
+            let scale = 1f64 / f64::sqrt(self.head_dim as f64);
+            let attn_weights = (query_states.matmul(&key_states.transpose(2, 3)?)? * scale)?;
+
+            let attn_weights = match attention_mask {
+                None => attn_weights,
+                Some(mask) => attn_weights.broadcast_add(mask)?,
+            };
+            let attn_weights = candle_nn::ops::softmax_last_dim(&attn_weights)?;
+            attn_weights.matmul(&value_states)?
+        };
+        attn_output
+            .transpose(1, 2)?
+            .reshape((b_sz, q_len, ()))?
+            .apply(&self.o_proj)
+    }
+
+    fn clear_kv_cache(&mut self) {
+        self.kv_cache = None
+    }
+}
+
+#[derive(Debug, Clone)]
+struct Mlp {
+    gate_up_proj: Linear,
+    down_proj: Linear,
+    act_fn: candle_nn::Activation,
+    i_size: usize,
+}
+
+impl Mlp {
+    fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
+        let hidden_size = cfg.hidden_size;
+        let i_size = cfg.intermediate_size;
+        let gate_up_proj = linear(hidden_size, 2 * i_size, vb.pp("gate_up_proj"))?;
+        let down_proj = linear(i_size, hidden_size, vb.pp("down_proj"))?;
+        Ok(Self {
+            gate_up_proj,
+            down_proj,
+            act_fn: cfg.hidden_act,
+            i_size,
+        })
+    }
+}
+
+impl Module for Mlp {
+    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
+        let up_states = xs.apply(&self.gate_up_proj)?;
+        let gate = up_states.narrow(D::Minus1, 0, self.i_size)?;
+        let up_states = up_states.narrow(D::Minus1, self.i_size, self.i_size)?;
+        let up_states = (up_states * gate.apply(&self.act_fn))?;
+        up_states.apply(&self.down_proj)
+    }
+}
+
+#[derive(Debug, Clone)]
+struct DecoderLayer {
+    self_attn: Attention,
+    mlp: Mlp,
+    input_layernorm: RmsNorm,
+    post_attention_layernorm: RmsNorm,
+}
+
+impl DecoderLayer {
+    fn new(rotary_emb: Arc<RotaryEmbedding>, cfg: &Config, vb: VarBuilder) -> Result<Self> {
+        let self_attn = Attention::new(rotary_emb, cfg, vb.pp("self_attn"))?;
+        let mlp = Mlp::new(cfg, vb.pp("mlp"))?;
+        let input_layernorm =
+            RmsNorm::new(cfg.hidden_size, cfg.rms_norm_eps, vb.pp("input_layernorm"))?;
+        let post_attention_layernorm = RmsNorm::new(
+            cfg.hidden_size,
+            cfg.rms_norm_eps,
+            vb.pp("post_attention_layernorm"),
+        )?;
+        Ok(Self {
+            self_attn,
+            mlp,
+            input_layernorm,
+            post_attention_layernorm,
+        })
+    }
+
+    fn forward(
+        &mut self,
+        xs: &Tensor,
+        attention_mask: Option<&Tensor>,
+        seqlen_offset: usize,
+    ) -> Result<Tensor> {
+        let residual = xs;
+        let xs = self.input_layernorm.forward(xs)?;
+        let xs = self.self_attn.forward(&xs, attention_mask, seqlen_offset)?;
+        let xs = (xs + residual)?;
+        let residual = &xs;
+        let xs = xs.apply(&self.post_attention_layernorm)?.apply(&self.mlp)?;
+        residual + xs
+    }
+
+    fn clear_kv_cache(&mut self) {
+        self.self_attn.clear_kv_cache()
+    }
+}
+
+#[derive(Debug, Clone)]
+pub struct Model {
+    embed_tokens: candle_nn::Embedding,
+    layers: Vec<DecoderLayer>,
+    norm: RmsNorm,
+    lm_head: Linear,
+    device: Device,
+    dtype: DType,
+}
+
+impl Model {
+    pub fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
+        let vb_m = vb.pp("model");
+        let embed_tokens =
+            candle_nn::embedding(cfg.vocab_size, cfg.hidden_size, vb_m.pp("embed_tokens"))?;
+        let rotary_emb = Arc::new(RotaryEmbedding::new(vb.dtype(), cfg, vb_m.device())?);
+        let mut layers = Vec::with_capacity(cfg.num_hidden_layers);
+        let vb_l = vb_m.pp("layers");
+        for layer_idx in 0..cfg.num_hidden_layers {
+            let layer = DecoderLayer::new(rotary_emb.clone(), cfg, vb_l.pp(layer_idx))?;
+            layers.push(layer)
+        }
+        let norm = RmsNorm::new(cfg.hidden_size, cfg.rms_norm_eps, vb_m.pp("norm"))?;
+        let lm_head = linear(cfg.hidden_size, cfg.vocab_size, vb.pp("lm_head"))?;
+        Ok(Self {
+            embed_tokens,
+            layers,
+            norm,
+            lm_head,
+            device: vb.device().clone(),
+            dtype: vb.dtype(),
+        })
+    }
+
+    fn prepare_decoder_attention_mask(
+        &self,
+        b_size: usize,
+        tgt_len: usize,
+        seqlen_offset: usize,
+    ) -> Result<Tensor> {
+        let mask: Vec<_> = (0..tgt_len)
+            .flat_map(|i| (0..tgt_len).map(move |j| if i < j { f32::NEG_INFINITY } else { 0. }))
+            .collect();
+        let mask = Tensor::from_slice(&mask, (tgt_len, tgt_len), &self.device)?;
+        let mask = if seqlen_offset > 0 {
+            let mask0 = Tensor::zeros((tgt_len, seqlen_offset), DType::F32, &self.device)?;
+            Tensor::cat(&[&mask0, &mask], D::Minus1)?
+        } else {
+            mask
+        };
+        mask.expand((b_size, 1, tgt_len, tgt_len + seqlen_offset))?
+            .to_dtype(self.dtype)
+    }
+
+    pub fn forward(&mut self, input_ids: &Tensor, seqlen_offset: usize) -> Result<Tensor> {
+        let (b_size, seq_len) = input_ids.dims2()?;
+        let attention_mask = if seq_len <= 1 {
+            None
+        } else {
+            let mask = self.prepare_decoder_attention_mask(b_size, seq_len, seqlen_offset)?;
+            Some(mask)
+        };
+        let mut xs = self.embed_tokens.forward(input_ids)?;
+        for layer in self.layers.iter_mut() {
+            xs = layer.forward(&xs, attention_mask.as_ref(), seqlen_offset)?
+        }
+        xs.narrow(1, seq_len - 1, 1)?
+            .apply(&self.norm)?
+            .apply(&self.lm_head)
+    }
+
+    pub fn clear_kv_cache(&mut self) {
+        for layer in self.layers.iter_mut() {
+            layer.clear_kv_cache()
+        }
+    }
+}