support bitsandbytes quantization with more models (#9148)

vllm/model_executor/models/gemma.py

@@ -332,6 +332,28 @@ class GemmaForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
         "gate_up_proj",
         "down_proj",
     ]
+
+    # BitandBytes specific attributes
+    default_bitsandbytes_target_modules = [
+        ".gate_proj.",
+        ".down_proj.",
+        ".up_proj.",
+        ".q_proj.",
+        ".k_proj.",
+        ".v_proj.",
+        ".o_proj.",
+    ]
+    # in TP, these weights are partitioned along the column dimension (dim=-1)
+    column_parallel_weights_modules = [".down_proj.", ".o_proj."]
+    bitsandbytes_stacked_params_mapping = {
+        # shard_name, weight_name, index
+        "q_proj": ("qkv_proj", 0),
+        "k_proj": ("qkv_proj", 1),
+        "v_proj": ("qkv_proj", 2),
+        "gate_proj": ("gate_up_proj", 0),
+        "up_proj": ("gate_up_proj", 1),
+    }
+
     # Gemma does not apply LoRA to the embedding layer.
     embedding_modules = {}
     embedding_padding_modules = []