Phi3 support.

server/text_generation_server/models/__init__.py

@@ -327,7 +327,7 @@ def get_model(
                 trust_remote_code=trust_remote_code,
             )
 
-    elif model_type == "llama" or model_type == "baichuan":
+    elif model_type == "llama" or model_type == "baichuan" or model_type == "phi3":
         if FLASH_ATTENTION:
             return FlashLlama(
                 model_id,

server/text_generation_server/models/custom_modeling/flash_llama_modeling.py

@@ -101,6 +101,13 @@ def load_attention(config, prefix, weights):
                 weights=weights,
                 bias=False,
             )
+        elif config.model_type == "phi3":
+            return TensorParallelColumnLinear.load(
+                config,
+                prefix=f"{prefix}.qkv_proj",
+                weights=weights,
+                bias=False,
+            )
         else:
             return TensorParallelColumnLinear.load_multi(
                 config,
@@ -257,6 +264,14 @@ class LlamaMLP(nn.Module):
             )
         )
         # Fuse gate and up proj
+        if config.model_type == "phi3":
+            self.gate_up_proj = TensorParallelColumnLinear.load(
+                config,
+                prefix=f"{prefix}.gate_up_proj",
+                weights=weights,
+                bias=False,
+            )
+        else:
             self.gate_up_proj = TensorParallelColumnLinear.load_multi(
                 config,
                 prefixes=[f"{prefix}.gate_proj", f"{prefix}.up_proj"],