support bitsandbytes quantization with more models (#9148)

vllm/model_executor/models/llama.py

@@ -449,6 +449,19 @@ class LlamaForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
         "lm_head": "output_embeddings"
     }
     embedding_padding_modules = ["lm_head"]
+
+    # 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),