cleanup conversion to use single qkv matrix

phi2/README.md

@@ -1,7 +1,7 @@
 # Phi-2
 
-Phi-2 is a 2.7B parameter language model released by Microsoft[^1] with
-performance that rivals much larger models. It was trained on a mixture of
+Phi-2 is a 2.7B parameter language model released by Microsoft with
+performance that rivals much larger models.[^1] It was trained on a mixture of
 GPT-4 outputs and clean web text.
 
 Phi-2 efficiently runs on Apple silicon devices with 8GB of memory in 16-bit

phi2/convert.py

@@ -1,34 +1,5 @@
 from transformers import AutoModelForCausalLM
-
-import numpy
-
-
-def split_attention_matrix(state_dict, key) -> dict:
-    # "transformer.h.0.mixer"
-    _, model_dim = state_dict[key + ".weight"].shape
-    # (3 * model_dim, model_dim)
-    Wqkv_weight_key = key + ".weight"
-    Wq_weight = state_dict[Wqkv_weight_key][:model_dim, :]
-    Wk_weight = state_dict[Wqkv_weight_key][model_dim : 2 * model_dim, :]
-    Wv_weight = state_dict[Wqkv_weight_key][2 * model_dim :, :]
-
-    # (3 * model_dim)
-    Wqkv_bias_key = key + ".bias"
-    Wq_bias = state_dict[Wqkv_bias_key][:model_dim]
-    Wk_bias = state_dict[Wqkv_bias_key][model_dim : 2 * model_dim]
-    Wv_bias = state_dict[Wqkv_bias_key][2 * model_dim :]
-
-    out_key = key.replace("mixer.Wqkv", "self_attention")
-
-    return {
-        out_key + ".query_proj.weight": Wq_weight,
-        out_key + ".query_proj.bias": Wq_bias,
-        out_key + ".key_proj.weight": Wk_weight,
-        out_key + ".key_proj.bias": Wk_bias,
-        out_key + ".value_proj.weight": Wv_weight,
-        out_key + ".value_proj.bias": Wv_bias,
-    }
-
+import numpy as np
 
 def replace_key(key: str) -> str:
     if "wte.weight" in key:
@@ -36,10 +7,6 @@ def replace_key(key: str) -> str:
 
     if ".mlp" in key:
         key = key.replace(".mlp", "")
-
-    if ".mixer.out_proj" in key:
-        key = key.replace(".mixer", ".self_attention")
-
     return key
 
 
@@ -48,19 +15,8 @@ def convert():
         "microsoft/phi-2", torch_dtype="auto", trust_remote_code=True
     )
     state_dict = model.state_dict()
-    keys = list(state_dict.keys())
-
-    for key in keys:
-        if ".mixer.Wqkv.weight" not in key:
-            continue
-        key_stub = key.rstrip(".weight")
-        state_dict.update(split_attention_matrix(state_dict, key_stub))
-
-        del state_dict[key_stub + ".weight"]
-        del state_dict[key_stub + ".bias"]
-
     weights = {replace_key(k): v.numpy() for k, v in state_dict.items()}
-    numpy.savez("weights.npz", **weights)
+    np.savez("weights.npz", **weights)
 
 
 if __name__ == "__main__":

phi2/phi2.py

@@ -31,15 +31,12 @@ class RoPEAttention(nn.Module):
         self.num_heads = num_heads
 
         self.rope = nn.RoPE(rotary_dim, traditional=False)
-        self.query_proj = nn.Linear(dims, dims)
-        self.key_proj = nn.Linear(dims, dims)
-        self.value_proj = nn.Linear(dims, dims)
+        self.Wqkv = nn.Linear(dims, 3 * dims)
         self.out_proj = nn.Linear(dims, dims)
 
-    def __call__(self, queries, keys, values, mask=None, cache=None):
-        queries = self.query_proj(queries)
-        keys = self.key_proj(keys)
-        values = self.value_proj(values)
+    def __call__(self, x, mask=None, cache=None):
+        qkv = self.Wqkv(x)
+        queries, keys, values = mx.split(qkv, 3, axis=-1)
 
         # Extract some shapes
         num_heads = self.num_heads
@@ -81,7 +78,7 @@ class ParallelBlock(nn.Module):
         super().__init__()
         dims = config.model_dim
         mlp_dims = dims * 4
-        self.self_attention = RoPEAttention(dims, config.num_heads, config.rotary_dim)
+        self.mixer = RoPEAttention(dims, config.num_heads, config.rotary_dim)
         self.ln = LayerNorm(dims)
         self.fc1 = nn.Linear(dims, mlp_dims)
         self.fc2 = nn.Linear(mlp_dims, dims)
@@ -89,7 +86,7 @@ class ParallelBlock(nn.Module):
 
     def __call__(self, x, mask, cache):
         h = self.ln(x)
-        attn_h, cache = self.self_attention(h, h, h, mask, cache)
+        attn_h, cache = self.mixer(h, mask, cache)
         ff_h = self.fc2(self.act(self.fc1(h)))
         return attn_h + ff_h + x, cache
 

phi2/requirements.txt

@@ -1,3 +1,4 @@
 einops
 mlx
+numpy
 transformers