Quantized version of StableLM. (#1058)

* Quantized version of StableLM.

* Adapt the stable-lm example to support quantizsed.

* Use some separate hub repo.

* Another repo name tweak.

candle-transformers/src/models/stable_lm.rs

@@ -1,4 +1,3 @@
-#![allow(unused)]
 use crate::models::with_tracing::{linear_no_bias, Linear};
 use candle::{DType, Device, Module, Result, Tensor, D};
 use candle_nn::{Activation, LayerNorm, VarBuilder};
@@ -41,21 +40,21 @@ impl Config {
         }
     }
 
-    fn head_dim(&self) -> usize {
+    pub fn head_dim(&self) -> usize {
         self.hidden_size / self.num_attention_heads
     }
 
-    fn rotary_ndims(&self) -> usize {
+    pub fn rotary_ndims(&self) -> usize {
         (self.head_dim() as f64 * self.rope_pct) as usize
     }
 
-    fn num_kv_groups(&self) -> usize {
+    pub fn num_kv_groups(&self) -> usize {
         self.num_attention_heads / self.num_key_value_heads
     }
 }
 
 #[derive(Debug)]
-struct RotaryEmbedding {
+pub(crate) struct RotaryEmbedding {
     sin: Tensor,
     cos: Tensor,
 }
@@ -66,7 +65,7 @@ fn rotate_half(xs: &Tensor) -> Result<Tensor> {
 }
 
 impl RotaryEmbedding {
-    fn new(dtype: DType, cfg: &Config, dev: &Device) -> Result<Self> {
+    pub(crate) fn new(dtype: DType, cfg: &Config, dev: &Device) -> Result<Self> {
         let dim = cfg.rotary_ndims();
         let max_seq_len = cfg.max_position_embeddings;
         let inv_freq: Vec<_> = (0..dim)
@@ -86,7 +85,7 @@ impl RotaryEmbedding {
         })
     }
 
-    fn apply_rotary_emb_qkv(
+    pub(crate) fn apply_rotary_emb_qkv(
         &self,
         q: &Tensor,
         k: &Tensor,