添加intel_mkl

candle_demo/Cargo.toml

@@ -25,6 +25,8 @@ candle-transformers = "0.6.0"
 candle-examples = "0.6.0"
 candle-nn = "0.6.0"
 safetensors = "0.4.3"
+accelerate-src = { version = "0.3.2", optional = true}
+intel-mkl-src = { version = "0.8.1", features = ["mkl-static-lp64-iomp"] ,optional = true}
 #safetensors = {path ="../safetensors/safetensors"}
 [build-dependencies]
 bindgen_cuda = { version = "0.1.1", optional = true }
@@ -33,4 +35,5 @@ bindgen_cuda = { version = "0.1.1", optional = true }
 [features]
 default = []
 cuda = ["candle-core/cuda", "candle-nn/cuda", "candle-transformers/cuda", "dep:bindgen_cuda"]
-
+accelerate = ["dep:accelerate-src", "candle-core/accelerate", "candle-nn/accelerate", "candle-transformers/accelerate"]
+mkl = ["dep:intel-mkl-src", "candle-core/mkl", "candle-nn/mkl", "candle-transformers/mkl"]

candle_demo/src/codegeex4.rs

@@ -1,7 +1,7 @@
-use candle_transformers::models::with_tracing::{linear_b as linear, Linear};
-use candle_core::{DType, Device, IndexOp, Module, Result, Tensor, D};
 use candle_core as candle;
+use candle_core::{DType, Device, IndexOp, Module, Result, Tensor, D};
 use candle_nn::VarBuilder;
+use candle_transformers::models::with_tracing::{linear_b as linear, Linear};
 
 #[derive(Debug, Clone)]
 pub struct Config {
@@ -30,7 +30,7 @@ impl Config {
     pub fn codegeex4() -> Self {
         Self {
             num_layers: 40,
-	    padded_vocab_size: 151552,
+            padded_vocab_size: 151552,
             hidden_size: 4096,
             ffn_hidden_size: 13696,
             kv_channels: 128,
@@ -68,7 +68,8 @@ impl RotaryEmbedding {
         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, cfg.seq_length as u32, dev)?
-            .to_dtype(dtype).expect("unalbe to dytpe in Rotray Embedding new")
+            .to_dtype(dtype)
+            .expect("unalbe to dytpe in Rotray Embedding new")
             .reshape((cfg.seq_length, 1))?;
         let freqs = t.matmul(&inv_freq)?;
         let cache = Tensor::stack(&[&freqs.cos()?, &freqs.sin()?], D::Minus1)?;
@@ -455,11 +456,11 @@ impl Transformer {
     fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
         let vb_l = vb.pp("layers");
         let mut layers = Vec::with_capacity(cfg.num_layers);
-	println!("transofrmer layers create");
-	let mut count = 0;
+        println!("transofrmer layers create");
+        let mut count = 0;
         for layer_index in 0..cfg.num_layers {
-	    count += 1;
-	    println!("for layer index in {} total is {} ",count, cfg.num_layers);
+            count += 1;
+            println!("for layer index in {} total is {} ", count, cfg.num_layers);
             let block = Block::new(layer_index + 1, cfg, vb_l.pp(layer_index))?;
             layers.push(block)
         }
@@ -564,8 +565,7 @@ impl Model {
             false,
             vb.pp("output_layer"),
         )?;
-	
-	
+
         Ok(Self {
             embedding,
             encoder,

candle_demo/src/main.rs

@@ -1,3 +1,9 @@
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+
 use clap::Parser;
 use codegeex4_candle::codegeex4::*;