[Model] Support bitsandbytes for MiniCPMV (vllm-project#9891)

Signed-off-by: mgoin <[email protected]>
Signed-off-by: Loc Huynh <[email protected]>

Author: mgoin

vllm/model_executor/models/minicpmv.py

@@ -810,6 +810,28 @@ class MiniCPMV2_5(MiniCPMVBaseModel, SupportsLoRA):
         # resampler
         "kv_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),
+    }
+
     embedding_modules = {}
     embedding_padding_modules = []
 
@@ -931,6 +953,27 @@ class MiniCPMV2_6(MiniCPMVBaseModel, SupportsLoRA):
         "kv_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),
+    }
+
     embedding_modules = {}
     embedding_padding_modules = []