More model cloning. (#1126)

* More model cloning.

* More cloning on quantized models.

candle-transformers/src/models/quantized_mistral.rs

@@ -6,7 +6,7 @@ use std::sync::Arc;
 
 pub use crate::models::mistral::Config;
 
-#[derive(Debug)]
+#[derive(Debug, Clone)]
 struct RotaryEmbedding {
     sin: Tensor,
     cos: Tensor,
@@ -57,7 +57,7 @@ impl RotaryEmbedding {
     }
 }
 
-#[derive(Debug)]
+#[derive(Debug, Clone)]
 #[allow(clippy::upper_case_acronyms)]
 struct MLP {
     gate_proj: Linear,
@@ -90,7 +90,7 @@ impl Module for MLP {
     }
 }
 
-#[derive(Debug)]
+#[derive(Debug, Clone)]
 struct Attention {
     q_proj: Linear,
     k_proj: Linear,
@@ -200,7 +200,7 @@ impl Attention {
     }
 }
 
-#[derive(Debug)]
+#[derive(Debug, Clone)]
 struct DecoderLayer {
     self_attn: Attention,
     mlp: MLP,
@@ -243,7 +243,7 @@ impl DecoderLayer {
     }
 }
 
-#[derive(Debug)]
+#[derive(Debug, Clone)]
 pub struct Model {
     embed_tokens: Embedding,
     layers: Vec<DecoderLayer>,

candle-transformers/src/models/quantized_mixformer.rs

@@ -7,7 +7,7 @@ pub use crate::models::mixformer::Config;
 
 const MAX_SEQ_LEN: usize = 4096;
 
-#[derive(Debug)]
+#[derive(Debug, Clone)]
 struct Embedding {
     wte: crate::quantized_nn::Embedding,
 }
@@ -39,7 +39,7 @@ fn masked_fill(on_false: &Tensor, mask: &Tensor, on_true: f32) -> Result<Tensor>
     Ok(m)
 }
 
-#[derive(Debug)]
+#[derive(Debug, Clone)]
 struct RotaryEmbedding {
     sin: Tensor,
     cos: Tensor,
@@ -105,7 +105,7 @@ impl RotaryEmbedding {
     }
 }
 
-#[derive(Debug)]
+#[derive(Debug, Clone)]
 #[allow(clippy::upper_case_acronyms)]
 struct MLP {
     fc1: Linear,
@@ -132,7 +132,7 @@ impl Module for MLP {
     }
 }
 
-#[derive(Debug)]
+#[derive(Debug, Clone)]
 struct CausalLMHead {
     ln: candle_nn::LayerNorm,
     linear: Linear,
@@ -154,7 +154,7 @@ impl Module for CausalLMHead {
     }
 }
 
-#[derive(Debug)]
+#[derive(Debug, Clone)]
 #[allow(clippy::upper_case_acronyms)]
 struct MHA {
     wqkv: Linear,
@@ -243,7 +243,7 @@ impl MHA {
     }
 }
 
-#[derive(Debug)]
+#[derive(Debug, Clone)]
 struct ParallelBlock {
     ln: candle_nn::LayerNorm,
     mixer: MHA,
@@ -278,7 +278,7 @@ impl ParallelBlock {
     }
 }
 
-#[derive(Debug)]
+#[derive(Debug, Clone)]
 pub struct MixFormerSequentialForCausalLM {
     embedding: Embedding,
     blocks: Vec<ParallelBlock>,

candle-transformers/src/models/quantized_mpt.rs

@@ -7,7 +7,7 @@ use candle_nn::LayerNorm;
 
 pub use super::mpt::Config;
 
-#[derive(Debug)]
+#[derive(Debug, Clone)]
 struct GroupedQueryAttention {
     wqkv: Linear,
     out_proj: Linear,
@@ -101,7 +101,7 @@ impl GroupedQueryAttention {
     }
 }
 
-#[derive(Debug)]
+#[derive(Debug, Clone)]
 struct Ffn {
     up_proj: Linear,
     down_proj: Linear,
@@ -122,7 +122,7 @@ impl Module for Ffn {
     }
 }
 
-#[derive(Debug)]
+#[derive(Debug, Clone)]
 struct MPTBlock {
     norm1: LayerNorm, // Do we need the low-precision variant?
     attn: GroupedQueryAttention,
@@ -155,7 +155,7 @@ impl MPTBlock {
     }
 }
 
-#[derive(Debug)]
+#[derive(Debug, Clone)]
 pub struct Model {
     wte: Embedding,
     blocks: Vec<MPTBlock>,

candle-transformers/src/quantized_nn.rs

@@ -2,7 +2,7 @@ use crate::models::with_tracing::QMatMul;
 use crate::quantized_var_builder::VarBuilder;
 use candle::{Module, Result, Tensor};
 
-#[derive(Debug)]
+#[derive(Debug, Clone)]
 pub struct Embedding {
     inner: candle_nn::Embedding,
     span: tracing::Span,
@@ -28,7 +28,7 @@ impl Module for Embedding {
     }
 }
 
-#[derive(Debug)]
+#[derive(Debug, Clone)]
 pub struct Linear {
     weight: QMatMul,
     bias: Option<Tensor>,
@@ -69,7 +69,7 @@ pub fn linear_no_bias(in_dim: usize, out_dim: usize, vb: VarBuilder) -> Result<L
     Ok(Linear { weight, bias: None })
 }
 
-#[derive(Debug)]
+#[derive(Debug, Clone)]
 pub struct RmsNorm {
     inner: candle_nn::RmsNorm,
     span: tracing::Span,