Fixing a bug from transformers==4.52. config.head_dim is now explicitly set to None (#552)

Signed-off-by: Gregory Shtrasberg <Gregory.Shtrasberg@amd.com>

Author: gshtras

vllm/distributed/kv_transfer/kv_connector/utils.py

@@ -44,8 +44,9 @@ def get_model_args(self, model_executable: torch.nn.Module):
             head_size = model_config.qk_nope_head_dim + \
                 model_config.qk_rope_head_dim
         else:
-            head_size = getattr(model_config, "head_dim",
-                                int(hidden_size // num_attention_heads))
+            head_size = getattr(model_config, "head_dim", None)
+            if head_size is None:
+                head_size = int(hidden_size // num_attention_heads)
 
         return num_heads, head_size
 

vllm/model_executor/models/exaone.py

@@ -127,8 +127,9 @@ def __init__(
             assert tp_size % self.total_num_kv_heads == 0
         self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
         # MistralConfig has an optional head_dim introduced by Mistral-Nemo
-        self.head_dim = getattr(config, "head_dim",
-                                self.hidden_size // self.total_num_heads)
+        self.head_dim = getattr(config, "head_dim", None)
+        if self.head_dim is None:
+            self.head_dim = self.hidden_size // self.total_num_heads
         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

vllm/model_executor/models/granite.py

@@ -122,8 +122,9 @@ def __init__(
             assert tp_size % self.total_num_kv_heads == 0
         self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
         # MistralConfig has an optional head_dim introduced by Mistral-Nemo
-        self.head_dim = getattr(config, "head_dim",
-                                self.hidden_size // self.total_num_heads)
+        self.head_dim = getattr(config, "head_dim", None)
+        if self.head_dim is None:
+            self.head_dim = self.hidden_size // self.total_num_heads
         self.q_size = self.num_heads * self.head_dim
         self.kv_size = self.num_kv_heads * self.head_dim
         self.scaling = config.attention_multiplier

vllm/model_executor/models/minimax_text_01.py

@@ -604,8 +604,9 @@ def __init__(
 
         rope_theta = getattr(config, "rope_theta", 10000)
 
-        head_dim = getattr(config, "head_dim",
-                           config.hidden_size // config.num_attention_heads)
+        head_dim = getattr(config, "head_dim", None)
+        if head_dim is None:
+            head_dim = config.hidden_size // config.num_attention_heads
         if hasattr(config, "max_model_len") and isinstance(
                 config.max_model_len, int):
             max_position_embeddings = min(config.max_position_embeddings,
@@ -861,8 +862,9 @@ def layer_fn(prefix):
                                                  cache_shape=self.cache_shape)
 
         rope_theta = getattr(config, "rope_theta", 10000)
-        head_dim = getattr(config, "head_dim",
-                           config.hidden_size // config.num_attention_heads)
+        head_dim = getattr(config, "head_dim", None)
+        if head_dim is None:
+            head_dim = config.hidden_size // config.num_attention_heads
         if hasattr(config, "max_model_len") and isinstance(
                 config.max_model_len, int):
             max_position_embeddings = min(config.max_position_embeddings,

vllm/model_executor/models/mixtral.py

@@ -138,8 +138,9 @@ def __init__(
             assert tp_size % self.total_num_kv_heads == 0
         self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
         # MixtralConfig has an optional head_dim argument
-        self.head_dim = getattr(config, "head_dim",
-                                self.hidden_size // self.total_num_heads)
+        self.head_dim = getattr(config, "head_dim", None)
+        if self.head_dim is None:
+            self.head_dim = self.hidden_size // self.total_num_heads
         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

vllm/model_executor/models/mixtral_quant.py

@@ -193,8 +193,9 @@ def __init__(
             assert tp_size % self.total_num_kv_heads == 0
         self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
         # MixtralConfig has an optional head_dim argument
-        self.head_dim = getattr(config, "head_dim",
-                                self.hidden_size // self.total_num_heads)
+        self.head_dim = getattr(config, "head_dim", None)
+        if self.head_dim is None:
+            self.head_dim = self.hidden_size // self.total_num_heads
         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

vllm/model_executor/models/nemotron.py

@@ -158,8 +158,9 @@ def __init__(
             assert tp_size % self.total_num_kv_heads == 0
         self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
         # MistralConfig has an optional head_dim introduced by Mistral-Nemo
-        self.head_dim = getattr(config, "head_dim",
-                                self.hidden_size // self.total_num_heads)
+        self.head_dim = getattr(config, "head_dim", None)
+        if self.head_dim is None:
+            self.head_dim = self.hidden_size // self.total_num_heads
         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

vllm/model_executor/models/solar.py

@@ -126,8 +126,9 @@ def __init__(
             assert tp_size % self.total_num_kv_heads == 0
         self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
         # MistralConfig has an optional head_dim introduced by Mistral-Nemo
-        self.head_dim = getattr(config, "head_dim",
-                                self.hidden_size // self.total_num_heads)
+        self.head_dim = getattr(config, "head_dim", None)
+        if self.head_dim is None:
+            self.head_dim = self.hidden_size // self.total_num_heads
         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