[RFC] Start removal of VarBuilder.

- Uses `Initializer` trait instead.
- Allows more decoupling between init and load, which are very different
  ops.
- Allows more decoupling between backends (safetensors, npy, ggml,
  etc...)

This is a minimum viable change.

There are 3 kind of objects with various relations.

The `Model`:
    This is `Llama`, `Linear`, `Rms` ...
    They contain tensors (and possibly other things).  and are used to
    call `forward` basically.
    They should have no ownership of any internals like Rng state or
    actual shapes of the tensors (the tensors already own those)

The `Initializer`:
    This is a struct containing necessary information to generate new
    random tensors. Typically they should own a random generator, and
    generate different kind of random tensors based on what kind of
    `Model` they are initializing.
    This do not own any information about the `Model` itself.
    Default init stores the `Vec<Var>` for now, in order to send to the
    optimizer.

Ths `Config`:
    This is the necessary information to link between the `Model` and
    the `Initializer`. This is another struct which is a companion of
    the implementation of the initalization.
    Typical information is the shape of the tensors for simple `Model`,
    the `eps` for RMS, the `use_bias` boolean to know whether we should
    have a bias in the linear layer.

This should remove all need for `VarBuilder` during intialization, and
allow removing every initialization bit within `VarBuilder`.

Modifying `llama2-c` to follow that initialization is left on purpose
for a follow-up to keep the current PR rather small.

candle-nn/examples/basic_optimizer.rs

@@ -1,5 +1,6 @@
 use candle::{DType, Device, Result, Tensor};
-use candle_nn::{linear, AdamW, Linear, ParamsAdamW, VarBuilder, VarMap};
+use candle_nn::init::{DefaultInit, ModelInitializer};
+use candle_nn::{AdamW, Linear, ParamsAdamW};
 
 fn gen_data() -> Result<(Tensor, Tensor)> {
     // Generate some sample linear data.
@@ -15,14 +16,15 @@ fn main() -> Result<()> {
     let (sample_xs, sample_ys) = gen_data()?;
 
     // Use backprop to run a linear regression between samples and get the coefficients back.
-    let varmap = VarMap::new();
+    // let varmap = VarMap::new();
-    let vb = VarBuilder::from_varmap(&varmap, DType::F32, &Device::Cpu);
+    // let vb = VarBuilder::from_varmap(&varmap, DType::F32, &Device::Cpu);
-    let model = linear(2, 1, vb.pp("linear"))?;
+    let mut initializer = DefaultInit::new(DType::F32, Device::Cpu);
+    let model = Linear::init(&mut initializer, ((2, 1), true))?;
     let params = ParamsAdamW {
         lr: 0.1,
         ..Default::default()
     };
-    let mut opt = AdamW::new(varmap.all_vars(), params)?;
+    let mut opt = AdamW::new(initializer.vars().to_vec(), params)?;
     for step in 0..10000 {
         let ys = model.forward(&sample_xs)?;
         let loss = ys.sub(&sample_ys)?.sqr()?.sum_all()?;

candle-nn/src/init.rs

@@ -3,6 +3,50 @@
 // https://github.com/pytorch/pytorch/blob/07107919297db3f8ab37f11c12666b6d6d5f692e/torch/nn/init.py#
 use candle::{DType, Device, Result, Shape, Tensor, Var};
 
+pub trait Initializer<M> {
+    type Config;
+    fn init(&mut self, config: Self::Config) -> Result<M>;
+}
+
+pub trait ModelInitializer: Sized {
+    fn init<INIT: Initializer<Self>>(initializer: &mut INIT, config: INIT::Config) -> Result<Self> {
+        initializer.init(config)
+    }
+}
+
+pub struct DefaultInit {
+    vars: Vec<Var>,
+    dtype: DType,
+    device: Device,
+}
+
+impl DefaultInit {
+    pub fn new(dtype: DType, device: Device) -> Self {
+        let vars = vec![];
+        Self {
+            dtype,
+            device,
+            vars,
+        }
+    }
+
+    pub fn dtype(&self) -> DType {
+        self.dtype
+    }
+
+    pub fn device(&self) -> &Device {
+        &self.device
+    }
+
+    pub fn vars(&self) -> &[Var] {
+        &self.vars
+    }
+
+    pub fn push_var(&mut self, var: Var) {
+        self.vars.push(var)
+    }
+}
+
 /// Number of features as input or output of a layer.
 /// In Kaiming initialization, choosing `FanIn` preserves
 /// the magnitude of the variance of the weights in the

candle-nn/src/linear.rs

@@ -17,6 +17,7 @@
 //! assert_eq!(ys.to_vec2::<f32>()?, &[[210.0, 430.0, 650.0]]);
 //! # Ok(()) }
 //! ```
+use crate::init::{DefaultInit, Initializer, ModelInitializer};
 use candle::{Result, Tensor};
 
 #[derive(Debug)]
@@ -43,23 +44,45 @@ impl Linear {
     }
 }
 
-/// Create or initialize a new linear layer.
+impl ModelInitializer for Linear {}
+
+impl Initializer<Linear> for DefaultInit {
+    type Config = ((usize, usize), bool);
+
+    fn init(&mut self, (shape, has_bias): Self::Config) -> Result<Linear> {
+        let dtype = self.dtype();
+        let device = self.device().clone();
+        let (out_dim, in_dim) = shape;
+        let init_ws = crate::init::DEFAULT_KAIMING_NORMAL;
+        let ws = init_ws.var(shape, dtype, &device)?;
+        self.push_var(ws.clone());
+        let ws = ws.as_tensor().clone();
+        if has_bias {
+            let bound = 1. / (in_dim as f64).sqrt();
+            let init_bs = crate::Init::Uniform {
+                lo: -bound,
+                up: bound,
+            };
+            let bs = init_bs.var(out_dim, dtype, &device)?;
+            self.push_var(bs.clone());
+            let bs = bs.as_tensor().clone();
+            Ok(Linear::new(ws, Some(bs)))
+        } else {
+            Ok(Linear::new(ws, None))
+        }
+    }
+}
+
+/// Loads a linear layer.
 ///
 /// This uses some default names for weight and biases, namely `"weight"` and `"bias"`.
 pub fn linear(in_dim: usize, out_dim: usize, vs: crate::VarBuilder) -> Result<Linear> {
-    let init_ws = crate::init::DEFAULT_KAIMING_NORMAL;
+    let ws = vs.get((out_dim, in_dim), "weight")?;
-    let ws = vs.get_or_init((out_dim, in_dim), "weight", init_ws)?;
+    let bs = vs.get(out_dim, "bias")?;
-    let bound = 1. / (in_dim as f64).sqrt();
-    let init_bs = crate::Init::Uniform {
-        lo: -bound,
-        up: bound,
-    };
-    let bs = vs.get_or_init(out_dim, "bias", init_bs)?;
     Ok(Linear::new(ws, Some(bs)))
 }
 
 pub fn linear_no_bias(in_dim: usize, out_dim: usize, vs: crate::VarBuilder) -> Result<Linear> {
-    let init_ws = crate::init::DEFAULT_KAIMING_NORMAL;
+    let ws = vs.get((out_dim, in_dim), "weight")?;
-    let ws = vs.get_or_init((out_dim, in_dim), "weight", init_ws)?;
     Ok(Linear::new(ws, None))
 }