temp(clip): 继续实现 clip 模型，部分实现 transformer

Signed-off-by: YdrMaster <ydrml@hotmail.com>

Author: YdrMaster

models/clip/common-cpu/src/test_infer.rs models/clip/common-cpu/src/infer.rsmodels/clip/common-cpu/src/test_infer.rs renamed to models/clip/common-cpu/src/infer.rs

@@ -25,7 +25,7 @@ fn test_infer() {
     println!("{meta:#?}");
 
     let &ClipMeta {
-        dt_embd,
+        dt,
 
         d_image,
         d_patch,
@@ -42,7 +42,7 @@ fn test_infer() {
     let time = Instant::now();
     let slices = image
         .slice_uhd(9, d_image, d_patch)
-        .normalize(dt_embd, image_mean, image_std);
+        .normalize(dt, image_mean, image_std);
     println!("slice image {:?}", time.elapsed());
 
     let weights = Weights::new(&storage);

models/clip/common-cpu/src/lib.rs

@@ -1,6 +1,6 @@
-use clip::{ClipStorage, WeightLoader};
-use operators::{common_cpu::Cpu, conv, QueueOf, TopoNode};
-use std::marker::PhantomData;
+use clip::{BlkWeight, ClipBlkStorage, ClipStorage, Tensor, WeightLoader};
+use operators::{common_cpu::Cpu, conv, ByteOf, QueueOf, TopoNode};
+use std::{marker::PhantomData, ops::Deref};
 
 pub struct Operators<N = Cpu>(PhantomData<N>);
 
@@ -21,7 +21,16 @@ where
     type TopoNode = Cpu;
     type Conv = conv::common_cpu::ConvIm2Col;
     type AddRows = op!(add_rows);
+    type Rearrange = op!(rearrange);
     type LayerNorm = op!(layer_norm);
+    type MatMul = op!(mat_mul);
+
+    fn debug<T>(tensor: &Tensor<T>)
+    where
+        T: Deref<Target = [ByteOf<Self::Hardware>]>,
+    {
+        println!("{tensor}")
+    }
 }
 
 impl<'w> Weights<'w> {
@@ -32,37 +41,67 @@ impl<'w> Weights<'w> {
 
 impl WeightLoader for Weights<'_> {
     type Hardware = Cpu;
-    type Weight<'s>
+    type Memory<'s>
         = &'s [u8]
     where
         Self: 's;
 
+    fn load_blk(
+        &self,
+        which: BlkWeight,
+        iblk: usize,
+        _queue: &QueueOf<Self::Hardware>,
+    ) -> [Self::Memory<'_>; 2] {
+        let ClipBlkStorage {
+            attn_norm_w,
+            attn_norm_b,
+            attn_qkv_w,
+            attn_qkv_b,
+            attn_o_w,
+            attn_o_b,
+            ffn_norm_w,
+            ffn_norm_b,
+            ffn_up_w,
+            ffn_up_b,
+            ffn_down_w,
+            ffn_down_b,
+        } = &self.0.blocks[iblk];
+        match which {
+            BlkWeight::AttnNorm => [attn_norm_w, attn_norm_b],
+            BlkWeight::AttnQKV => [attn_qkv_w, attn_qkv_b],
+            BlkWeight::AttnO => [attn_o_w, attn_o_b],
+            BlkWeight::FfnNorm => [ffn_norm_w, ffn_norm_b],
+            BlkWeight::FfnUp => [ffn_up_w, ffn_up_b],
+            BlkWeight::FfnDown => [ffn_down_w, ffn_down_b],
+        }
+    }
+
     #[inline]
-    fn patch_embd<'a>(&'a self, _queue: &'a QueueOf<Self::Hardware>) -> [Self::Weight<'a>; 2] {
+    fn patch_embd<'a>(&'a self, _queue: &'a QueueOf<Self::Hardware>) -> [Self::Memory<'a>; 2] {
         [self.0.patch_embd_w, self.0.patch_embd_b]
     }
 
     #[inline]
-    fn pos_embd<'a>(&'a self, _queue: &'a QueueOf<Self::Hardware>) -> Self::Weight<'a> {
+    fn pos_embd<'a>(&'a self, _queue: &'a QueueOf<Self::Hardware>) -> Self::Memory<'a> {
         self.0.pos_embd
     }
 
     #[inline]
     fn pre_norm<'a>(
         &'a self,
         _queue: &'a QueueOf<Self::Hardware>,
-    ) -> Option<[Self::Weight<'a>; 2]> {
+    ) -> Option<[Self::Memory<'a>; 2]> {
         self.0.pre_norm
     }
 
     #[inline]
     fn post_norm<'a>(
         &'a self,
         _queue: &'a QueueOf<Self::Hardware>,
-    ) -> Option<[Self::Weight<'a>; 2]> {
+    ) -> Option<[Self::Memory<'a>; 2]> {
         self.0.post_norm
     }
 }
 
 #[cfg(test)]
-mod test_infer;
+mod infer;