chore: add workaround for quant

llms/yayi2/convert.py

@@ -84,6 +84,23 @@ def convert(args):
     state_dict = {k.replace("v_proj", "wv"): v for k, v in state_dict.items()}
     state_dict = {k.replace("o_proj", "wo"): v for k, v in state_dict.items()}
 
+    # quantization workaround for kv proj
+    keys_to_delete =[]
+    new_state_dict ={}
+    for k, v in state_dict.items():
+        if k.endswith('wk.weight'):
+            prefix = k[:-len('wk.weight')]
+            wv = prefix +"wv.weight"
+            if wv in state_dict:
+                wkwv = torch.cat([v, state_dict[wv]],dim=0)
+                new_key = prefix + "wkwv.weight"
+                new_state_dict[new_key] = wkwv
+                keys_to_delete.extend([k,wv])
+
+    for key in keys_to_delete:
+        del state_dict[key]
+
+    state_dict.update(new_state_dict)
     weights = {k: v.numpy() for k, v in state_dict.items()}
 
     keep_keys = set(
@@ -152,3 +169,4 @@ if __name__ == "__main__":
     with open(mlx_path / "config.json", "w") as f:
         config["model_type"] = "yayi"
         json.dump(config, f, indent=4)
+

llms/yayi2/yayi.py

@@ -1,6 +1,7 @@
 import argparse
 import json
 from dataclasses import dataclass
+import math
 from pathlib import Path
 from typing import Optional, Tuple
 
@@ -23,6 +24,21 @@ class ModelArgs:
     rope_traditional: bool = False
 
 
+class KVLinear(nn.Module):
+    def __init__(self, input_dims: int, output_dims: int):
+        super().__init__()
+        scale = math.sqrt(1 / input_dims)
+        self.weight = mx.random.uniform(
+            low=-scale,
+            high=scale,
+            shape=(output_dims, input_dims),
+        )
+
+    def __call__(self, x):
+        k, v = mx.split(x @ self.weight.T, 2, axis=-1)
+        return k, v
+
+
 class RMSNorm(nn.Module):
     def __init__(self, dims: int, eps: float = 1e-5):
         super().__init__()
@@ -49,8 +65,8 @@ class Attention(nn.Module):
         self.wq = nn.Linear(
             args.hidden_size, args.num_attention_heads * self.head_dim, bias=False
         )
-        self.wk = nn.Linear(args.hidden_size, int(self.head_dim), bias=False)
-        self.wv = nn.Linear(args.hidden_size, int(self.head_dim), bias=False)
+        self.wkwv = KVLinear(args.hidden_size, int(self.head_dim) * 2)
+
         self.wo = nn.Linear(
             args.num_attention_heads * self.head_dim, args.hidden_size, bias=False
         )
@@ -66,7 +82,8 @@ class Attention(nn.Module):
     ) -> mx.array:
         B, L, _ = x.shape
 
-        q, k, v = self.wq(x), self.wk(x), self.wv(x)
+        q = self.wq(x)
+        k, v = self.wkwv(x)
 
         q = q.reshape(B, L, self.num_attention_heads, self.head_dim).transpose(
             0, 2, 1, 3
@@ -189,7 +206,9 @@ def load_model(model_path: str):
     weights = mx.load(str(model_path / "weights.npz"))
     if quantization is not None:
         nn.QuantizedLinear.quantize_module(model, **quantization)
-    model.update(tree_unflatten(list(weights.items())))
+    parameteres = tree_unflatten(list(weights.items()))
+
+    model.update(parameteres)
 
     tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
     return model, tokenizer
@@ -206,7 +225,7 @@ if __name__ == "__main__":
     parser.add_argument(
         "--prompt",
         help="The message to be processed by the model",
-        default="The winter in Beijing is",
+        default="The winter in Beijing is ",
     )
     parser.add_argument(
         "--max-tokens",