[Minor][Models] Pass partial_rotary_factor parameter to rope (vllm-project#17266)

Signed-off-by: evian <[email protected]>
Co-authored-by: evian <[email protected]>
Signed-off-by: Agata Dobrzyniewicz <[email protected]>

Author: 2 people

vllm/model_executor/models/llama.py

@@ -130,8 +130,8 @@ def __init__(self,
         self.head_dim = getattr(config, "head_dim",
                                 self.hidden_size // self.total_num_heads)
         # Phi models introduced a partial_rotary_factor parameter in the config
-        partial_rotary_factor = getattr(config, "partial_rotary_factor", 1)
-        self.rotary_dim = int(partial_rotary_factor * self.head_dim)
+        self.partial_rotary_factor = getattr(config, "partial_rotary_factor",
+                                             1)
         self.q_size = self.num_heads * self.head_dim
         self.kv_size = self.num_kv_heads * self.head_dim
         self.scaling = self.head_dim**-0.5
@@ -163,11 +163,12 @@ def __init__(self,
 
         self.rotary_emb = get_rope(
             self.head_dim,
-            rotary_dim=self.rotary_dim,
+            rotary_dim=self.head_dim,
             max_position=max_position_embeddings,
             base=rope_theta,
             rope_scaling=rope_scaling,
             is_neox_style=is_neox_style,
+            partial_rotary_factor=self.partial_rotary_factor,
         )
 
         if hasattr(config, "interleaved_sliding_window"):

vllm/model_executor/models/persimmon.py

@@ -115,9 +115,10 @@ def __init__(self,
 
         self.rotary_emb = get_rope(
             self.head_dim,
-            rotary_dim=int(self.partial_rotary_factor * self.head_dim),
+            rotary_dim=self.head_dim,
             max_position=self.max_position_embeddings,
             base=self.rope_theta,
+            partial_rotary_factor=self.partial_rotary_factor,
         )
         self.scaling = self.head_dim**-0.5
         self.attn = Attention(self.num_heads,

vllm/model_executor/models/stablelm.py

@@ -104,9 +104,8 @@ def __init__(self,
             1, self.total_num_key_value_heads // tp_size)
         self.head_dim = self.hidden_size // self.total_num_heads
         self.max_position_embeddings = config.max_position_embeddings
-        rope_pct = getattr(config, "rope_pct",
-                           getattr(config, "partial_rotary_factor", 1))
-        self.rotary_ndims = int(self.head_dim * rope_pct)
+        self.partial_rotary_factor = getattr(
+            config, "rope_pct", getattr(config, "partial_rotary_factor", 1))
         self.scaling = self.head_dim**-0.5
         self.q_size = self.num_heads * self.head_dim
         self.kv_size = self.num_key_value_heads * self.head_dim
@@ -130,9 +129,10 @@ def __init__(self,
                                         prefix=f"{prefix}.o_proj")
         self.rotary_emb = get_rope(
             self.head_dim,
-            rotary_dim=self.rotary_ndims,
+            rotary_dim=self.head_dim,
             max_position=self.config.max_position_embeddings,
             base=self.config.rope_theta,
+            partial_rotary_factor=self.partial_rotary_factor,
         )
         self.attn = Attention(self.num_heads,
                               self.head_dim,