Fixing mistral nemo. (#2276)

server/text_generation_server/models/__init__.py

@@ -762,8 +762,6 @@ def get_model(
                 default_dtype=torch.bfloat16,
                 trust_remote_code=trust_remote_code,
                 lora_adapter_ids=lora_adapter_ids,
-                # hidden_size / num_attention_heads is wrong in `google/gemma-2-9b-it`
-                head_size=config_dict["head_dim"],
             )
         elif sharded:
             raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("Sharded Gemma2"))

server/text_generation_server/models/custom_modeling/flash_mistral_modeling.py

@@ -117,6 +117,9 @@ class MistralAttention(torch.nn.Module):
         )
         self.num_heads = config.num_attention_heads
         self.hidden_size = config.hidden_size
+        if hasattr(config, "head_dim"):
+            self.head_size = config.head_dim
+        else:
             self.head_size = self.hidden_size // self.num_heads
 
         self.rotary_emb = PositionRotaryEmbedding.static(

server/text_generation_server/models/flash_causal_lm.py

@@ -925,6 +925,11 @@ class FlashCausalLM(Model):
         assert self.num_kv_heads > 0
 
         if head_size is None:
+            # Some models use GQA and different sizes for o_proj
+            # and q_proj, that allows for that.
+            if hasattr(config, "head_dim"):
+                self.head_size = config.head_dim
+            else:
                 self.head_size = config.hidden_size // config.num_attention_heads
         else:
             self.head_size = head_size