mirror of
https://github.com/huggingface/candle.git
synced 2025-06-16 10:38:54 +00:00
Quantized version of mistral. (#1009)
* Quantized version of mistral. * Integrate the quantized mistral variant. * Use the quantized weight files. * Tweak the quantization command. * Fix the dtype when computing the rotary embeddings. * Update the readme with the quantized version. * Fix the decoding of the remaining tokens.
This commit is contained in:
Laurent Mazare
@ -6,18 +6,18 @@ use std::sync::Arc;
|
||||
|
||||
#[derive(Debug, Clone, PartialEq)]
|
||||
pub struct Config {
|
||||
vocab_size: usize,
|
||||
hidden_size: usize,
|
||||
intermediate_size: usize,
|
||||
num_hidden_layers: usize,
|
||||
num_attention_heads: usize,
|
||||
num_key_value_heads: usize,
|
||||
hidden_act: Activation,
|
||||
max_position_embeddings: usize,
|
||||
rms_norm_eps: f64,
|
||||
rope_theta: f64,
|
||||
sliding_window: usize,
|
||||
use_flash_attn: bool,
|
||||
pub(crate) vocab_size: usize,
|
||||
pub(crate) hidden_size: usize,
|
||||
pub(crate) intermediate_size: usize,
|
||||
pub(crate) num_hidden_layers: usize,
|
||||
pub(crate) num_attention_heads: usize,
|
||||
pub(crate) num_key_value_heads: usize,
|
||||
pub(crate) hidden_act: Activation,
|
||||
pub(crate) max_position_embeddings: usize,
|
||||
pub(crate) rms_norm_eps: f64,
|
||||
pub(crate) rope_theta: f64,
|
||||
pub(crate) sliding_window: usize,
|
||||
pub(crate) use_flash_attn: bool,
|
||||
}
|
||||
|
||||
impl Config {
|
||||
|
||||
@ -7,6 +7,7 @@ pub mod llama;
|
||||
pub mod mistral;
|
||||
pub mod mixformer;
|
||||
pub mod quantized_llama;
|
||||
pub mod quantized_mistral;
|
||||
pub mod quantized_mixformer;
|
||||
pub mod quantized_t5;
|
||||
pub mod segment_anything;
|
||||
|
||||
364
candle-transformers/src/models/quantized_mistral.rs
Normal file
364
candle-transformers/src/models/quantized_mistral.rs
Normal file
@ -0,0 +1,364 @@
|
||||
use crate::models::quantized_t5::Embedding;
|
||||
use crate::models::with_tracing::QMatMul;
|
||||
pub use crate::quantized_var_builder::VarBuilder;
|
||||
use candle::{DType, Device, Module, Result, Tensor, D};
|
||||
use candle_nn::Activation;
|
||||
use std::sync::Arc;
|
||||
|
||||
pub use crate::models::mistral::Config;
|
||||
|
||||
#[derive(Debug)]
|
||||
struct Linear {
|
||||
weight: QMatMul,
|
||||
}
|
||||
|
||||
impl Module for Linear {
|
||||
fn forward(&self, x: &Tensor) -> candle::Result<Tensor> {
|
||||
x.apply(&self.weight)
|
||||
}
|
||||
}
|
||||
|
||||
fn linear_no_bias(in_dim: usize, out_dim: usize, vb: VarBuilder) -> Result<Linear> {
|
||||
let weight = QMatMul::new(in_dim, out_dim, vb)?;
|
||||
Ok(Linear { weight })
|
||||
}
|
||||
|
||||
#[derive(Debug)]
|
||||
struct RmsNorm {
|
||||
inner: candle_nn::RmsNorm,
|
||||
span: tracing::Span,
|
||||
}
|
||||
|
||||
impl RmsNorm {
|
||||
fn new(size: usize, eps: f64, vb: VarBuilder) -> Result<Self> {
|
||||
let span = tracing::span!(tracing::Level::TRACE, "rms-norm");
|
||||
let weight = vb.get(size, "weight")?.dequantize(vb.device())?;
|
||||
let inner = candle_nn::RmsNorm::new(weight, eps);
|
||||
Ok(Self { inner, span })
|
||||
}
|
||||
}
|
||||
|
||||
impl Module for RmsNorm {
|
||||
fn forward(&self, x: &Tensor) -> Result<Tensor> {
|
||||
let _enter = self.span.enter();
|
||||
self.inner.forward(x)
|
||||
}
|
||||
}
|
||||
|
||||
#[derive(Debug)]
|
||||
struct RotaryEmbedding {
|
||||
sin: Tensor,
|
||||
cos: Tensor,
|
||||
}
|
||||
|
||||
fn rotate_half(xs: &Tensor) -> Result<Tensor> {
|
||||
let last_dim = xs.dim(D::Minus1)?;
|
||||
let xs1 = xs.narrow(D::Minus1, 0, last_dim / 2)?;
|
||||
let xs2 = xs.narrow(D::Minus1, last_dim / 2, last_dim - last_dim / 2)?;
|
||||
Tensor::cat(&[&xs2.neg()?, &xs1], D::Minus1)
|
||||
}
|
||||
|
||||
impl RotaryEmbedding {
|
||||
fn new(cfg: &Config, dev: &Device) -> Result<Self> {
|
||||
let dim = cfg.hidden_size / cfg.num_attention_heads;
|
||||
let max_seq_len = cfg.max_position_embeddings;
|
||||
let inv_freq: Vec<_> = (0..dim)
|
||||
.step_by(2)
|
||||
.map(|i| 1f32 / 10000f32.powf(i as f32 / dim as f32))
|
||||
.collect();
|
||||
let inv_freq_len = inv_freq.len();
|
||||
let inv_freq = Tensor::from_vec(inv_freq, (1, inv_freq_len), dev)?;
|
||||
let t = Tensor::arange(0u32, max_seq_len as u32, dev)?
|
||||
.to_dtype(DType::F32)?
|
||||
.reshape((max_seq_len, 1))?;
|
||||
let freqs = t.matmul(&inv_freq)?;
|
||||
let freqs = Tensor::cat(&[&freqs, &freqs], D::Minus1)?;
|
||||
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 cos = cos.unsqueeze(0)?.unsqueeze(0)?; // (1, 1, seq_len, dim)
|
||||
let sin = sin.unsqueeze(0)?.unsqueeze(0)?; // (1, 1, seq_len, dim)
|
||||
let q_embed = (q.broadcast_mul(&cos)? + rotate_half(q)?.broadcast_mul(&sin))?;
|
||||
let k_embed = (k.broadcast_mul(&cos)? + rotate_half(k)?.broadcast_mul(&sin))?;
|
||||
Ok((q_embed, k_embed))
|
||||
}
|
||||
}
|
||||
|
||||
#[derive(Debug)]
|
||||
#[allow(clippy::upper_case_acronyms)]
|
||||
struct MLP {
|
||||
gate_proj: Linear,
|
||||
up_proj: Linear,
|
||||
down_proj: Linear,
|
||||
act_fn: Activation,
|
||||
}
|
||||
|
||||
impl MLP {
|
||||
fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
|
||||
let hidden_sz = cfg.hidden_size;
|
||||
let intermediate_sz = cfg.intermediate_size;
|
||||
let gate_proj = linear_no_bias(hidden_sz, intermediate_sz, vb.pp("gate_proj"))?;
|
||||
let up_proj = linear_no_bias(hidden_sz, intermediate_sz, vb.pp("up_proj"))?;
|
||||
let down_proj = linear_no_bias(intermediate_sz, hidden_sz, vb.pp("down_proj"))?;
|
||||
Ok(Self {
|
||||
gate_proj,
|
||||
up_proj,
|
||||
down_proj,
|
||||
act_fn: cfg.hidden_act,
|
||||
})
|
||||
}
|
||||
}
|
||||
|
||||
impl Module for MLP {
|
||||
fn forward(&self, xs: &Tensor) -> Result<Tensor> {
|
||||
let lhs = xs.apply(&self.gate_proj)?.apply(&self.act_fn)?;
|
||||
let rhs = xs.apply(&self.up_proj)?;
|
||||
(lhs * rhs)?.apply(&self.down_proj)
|
||||
}
|
||||
}
|
||||
|
||||
#[derive(Debug)]
|
||||
struct Attention {
|
||||
q_proj: Linear,
|
||||
k_proj: Linear,
|
||||
v_proj: Linear,
|
||||
o_proj: Linear,
|
||||
num_heads: usize,
|
||||
num_kv_heads: usize,
|
||||
num_kv_groups: usize,
|
||||
head_dim: usize,
|
||||
hidden_size: 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 hidden_sz = cfg.hidden_size;
|
||||
let num_heads = cfg.num_attention_heads;
|
||||
let num_kv_heads = cfg.num_key_value_heads;
|
||||
let num_kv_groups = num_heads / num_kv_heads;
|
||||
let head_dim = hidden_sz / num_heads;
|
||||
let q_proj = linear_no_bias(hidden_sz, num_heads * head_dim, vb.pp("q_proj"))?;
|
||||
let k_proj = linear_no_bias(hidden_sz, num_kv_heads * head_dim, vb.pp("k_proj"))?;
|
||||
let v_proj = linear_no_bias(hidden_sz, num_kv_heads * head_dim, vb.pp("v_proj"))?;
|
||||
let o_proj = linear_no_bias(num_heads * head_dim, hidden_sz, vb.pp("o_proj"))?;
|
||||
Ok(Self {
|
||||
q_proj,
|
||||
k_proj,
|
||||
v_proj,
|
||||
o_proj,
|
||||
num_heads,
|
||||
num_kv_heads,
|
||||
num_kv_groups,
|
||||
head_dim,
|
||||
hidden_size: hidden_sz,
|
||||
rotary_emb,
|
||||
kv_cache: None,
|
||||
})
|
||||
}
|
||||
|
||||
fn repeat_kv(&self, xs: Tensor) -> Result<Tensor> {
|
||||
let n_rep = self.num_kv_groups;
|
||||
if n_rep == 1 {
|
||||
Ok(xs)
|
||||
} else {
|
||||
let (b_sz, num_kv_heads, seq_len, head_dim) = xs.dims4()?;
|
||||
xs.unsqueeze(2)?
|
||||
.expand((b_sz, num_kv_heads, n_rep, seq_len, head_dim))?
|
||||
.reshape((b_sz, num_kv_heads * n_rep, seq_len, 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 query_states = self.q_proj.forward(xs)?;
|
||||
let key_states = self.k_proj.forward(xs)?;
|
||||
let value_states = self.v_proj.forward(xs)?;
|
||||
|
||||
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 = self.repeat_kv(key_states)?;
|
||||
let value_states = self.repeat_kv(value_states)?;
|
||||
|
||||
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, self.hidden_size))?
|
||||
.apply(&self.o_proj)
|
||||
}
|
||||
}
|
||||
|
||||
#[derive(Debug)]
|
||||
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
|
||||
}
|
||||
}
|
||||
|
||||
#[derive(Debug)]
|
||||
pub struct Model {
|
||||
embed_tokens: Embedding,
|
||||
layers: Vec<DecoderLayer>,
|
||||
norm: RmsNorm,
|
||||
lm_head: Linear,
|
||||
sliding_window: usize,
|
||||
device: Device,
|
||||
}
|
||||
|
||||
impl Model {
|
||||
pub fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
|
||||
let vb_m = vb.pp("model");
|
||||
let embed_tokens =
|
||||
Embedding::new(cfg.vocab_size, cfg.hidden_size, vb_m.pp("embed_tokens"))?;
|
||||
let rotary_emb = Arc::new(RotaryEmbedding::new(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_no_bias(cfg.hidden_size, cfg.vocab_size, vb.pp("lm_head"))?;
|
||||
Ok(Self {
|
||||
embed_tokens,
|
||||
layers,
|
||||
norm,
|
||||
lm_head,
|
||||
sliding_window: cfg.sliding_window,
|
||||
device: vb.device().clone(),
|
||||
})
|
||||
}
|
||||
|
||||
fn prepare_decoder_attention_mask(
|
||||
&self,
|
||||
b_size: usize,
|
||||
tgt_len: usize,
|
||||
seqlen_offset: usize,
|
||||
) -> Result<Tensor> {
|
||||
// Sliding window mask?
|
||||
let mask: Vec<_> = (0..tgt_len)
|
||||
.flat_map(|i| {
|
||||
(0..tgt_len).map(move |j| {
|
||||
if i < j || j + self.sliding_window < i {
|
||||
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(DType::F32)
|
||||
}
|
||||
|
||||
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)
|
||||
}
|
||||
}
|
||||
Reference in New Issue
Block a user