[Bugfix] Fix MiniCPMV and Mllama BNB bug (#9917)

Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>

vllm/model_executor/layers/resampler.py

@@ -41,6 +41,7 @@ from torch import nn
 from torch.nn.init import trunc_normal_
 
 from vllm.model_executor.layers.linear import ReplicatedLinear
+from vllm.model_executor.layers.quantization import QuantizationConfig
 
 DEFAULT_LN = partial(nn.LayerNorm, eps=1e-6)
 
@@ -154,15 +155,15 @@ class BaseResampler(nn.Module):
         A tensor with the shape of (grid_size**2, embed_dim)
     """
 
-    def __init__(
-        self,
+    def __init__(self,
                  num_queries: int,
                  embed_dim: int,
                  num_heads: int,
                  kv_dim: Optional[int] = None,
                  norm_layer: Callable[[int], nn.LayerNorm] = DEFAULT_LN,
                  do_post_projection: bool = True,
-    ) -> None:
+                 quant_config: Optional[QuantizationConfig] = None,
+                 prefix: str = "") -> None:
         super().__init__()
 
         self.num_queries = num_queries
@@ -172,7 +173,11 @@ class BaseResampler(nn.Module):
         self.query = nn.Parameter(torch.zeros(self.num_queries, embed_dim))
         trunc_normal_(self.query, std=0.02)
         if kv_dim is not None and kv_dim != embed_dim:
-            self.kv_proj = ReplicatedLinear(kv_dim, embed_dim, bias=False)
+            self.kv_proj = ReplicatedLinear(kv_dim,
+                                            embed_dim,
+                                            bias=False,
+                                            quant_config=quant_config,
+                                            prefix=prefix)
         else:
             # Maintain the same return value with ReplicatedLinear.forward
             self.kv_proj = lambda *args, **kwargs: (  # type: ignore # noqa 
@@ -209,8 +214,7 @@ class Resampler2(BaseResampler):
     present in minicpmv2.0, but not qwen-vl.
     """
 
-    def __init__(
-        self,
+    def __init__(self,
                  grid_size: int,
                  embed_dim: int,
                  num_heads: int,
@@ -218,13 +222,16 @@ class Resampler2(BaseResampler):
                  norm_layer: Callable[[int], nn.LayerNorm] = DEFAULT_LN,
                  adaptive: bool = False,
                  do_post_projection: bool = True,
-    ) -> None:
+                 quant_config: Optional[QuantizationConfig] = None,
+                 prefix: str = "") -> None:
         super().__init__(grid_size**2,
                          embed_dim,
                          num_heads,
                          kv_dim,
                          norm_layer,
-                         do_post_projection=do_post_projection)
+                         do_post_projection=do_post_projection,
+                         quant_config=quant_config,
+                         prefix=prefix)
 
         self.adaptive = adaptive
         pos_embed_arr = get_2d_sincos_pos_embed(embed_dim,

vllm/model_executor/model_loader/loader.py

@@ -28,6 +28,7 @@ from vllm.distributed import (get_tensor_model_parallel_rank,
                               get_tensor_model_parallel_world_size)
 from vllm.envs import VLLM_USE_MODELSCOPE
 from vllm.logger import init_logger
+from vllm.model_executor.layers.linear import ReplicatedLinear
 from vllm.model_executor.layers.quantization.base_config import (
     QuantizationConfig)
 from vllm.model_executor.model_loader.tensorizer import (
@@ -771,6 +772,8 @@ class BitsAndBytesModelLoader(BaseModelLoader):
         with open(config_file_path, "r") as f:
             config = json.load(f)
             self.target_modules = config["target_modules"]
+        # Save the module names without sharding.
+        self.unsharded_weights_modules: List[str] = []
 
     def _get_config_file(self, qlora_adapter: str) -> str:
         is_local = os.path.isdir(qlora_adapter)
@@ -990,8 +993,13 @@ class BitsAndBytesModelLoader(BaseModelLoader):
             if any(target_module in weight_name for target_module in
                    self.target_modules) and weight_name.endswith(".weight"):
                 weight_name = weight_name.replace(".weight", ".qweight")
-
-                if any(module in weight_name
+                # Without sharding
+                if any(
+                        weight_name.startswith(module)
+                        for module in self.unsharded_weights_modules):
+                    weight_sub_tensor = weight_tensor
+                # Shard by column
+                elif any(module in weight_name
                          for module in self.column_parallel_weights_modules):
 
                     total_size = weight_tensor.size(-1)
@@ -999,7 +1007,7 @@ class BitsAndBytesModelLoader(BaseModelLoader):
                     end_index = total_size // tp_size * (tp_rank + 1)
                     weight_sub_tensor = weight_tensor[...,
                                                       start_index:end_index]
-
+                # Shard by row
                 else:
                     total_size = weight_tensor.size(0)
                     start_index = total_size // tp_size * tp_rank
@@ -1053,7 +1061,15 @@ class BitsAndBytesModelLoader(BaseModelLoader):
                 model.column_parallel_weights_modules
         else:
             self.column_parallel_weights_modules = []
-
+        # Some modules like `ReplicatedLinear` should not have their weights
+        # sharded. The reason for implementing it this way is to avoid new
+        # static variable in the model implementation.
+        # TODO: Can we reduce the static variables needed for BNB based on
+        #  model information?
+        self.unsharded_weights_modules = [
+            name for name, module in model.named_modules()
+            if isinstance(module, (ReplicatedLinear, ))
+        ]
         self.model_type = type(model).__name__
 
         logger.info("Loading weights with BitsAndBytes quantization. "
@@ -1100,7 +1116,13 @@ class BitsAndBytesModelLoader(BaseModelLoader):
             for shard_name, (
                     weight_name, index
             ) in model.bitsandbytes_stacked_params_mapping.items():
-                if shard_name in quant_param_name:
+
+                shard_pos = quant_param_name.find(shard_name)
+                # Some models, such as MiniCPM V2.5/2.6, contain both
+                # module names 'kv_proj' and 'qkv_proj'. To prevent 'kv_proj'
+                # from being incorrectly identified as being present in
+                # 'vpm.encoder.layers.0.self_attn.qkv_proj.qweight
+                if shard_pos > 0 and quant_param_name[shard_pos - 1] == ".":
                     shard_index = index
                     quant_param_name = quant_param_name.replace(
                         shard_name, weight_name)

vllm/model_executor/models/minicpmv.py

@@ -131,16 +131,22 @@ DEFAULT_LN = partial(nn.LayerNorm, eps=1e-6)
 
 class Resampler2_5(BaseResampler):
 
-    def __init__(
-            self,
+    def __init__(self,
                  num_queries: int,
                  embed_dim: int,
                  num_heads: int,
                  kv_dim: Optional[int] = None,
                  norm_layer: Callable[[int], nn.LayerNorm] = DEFAULT_LN,
                  max_size: Tuple[int, int] = (70, 70),
-    ) -> None:
-        super().__init__(num_queries, embed_dim, num_heads, kv_dim, norm_layer)
+                 quant_config: Optional[QuantizationConfig] = None,
+                 prefix: str = "") -> None:
+        super().__init__(num_queries,
+                         embed_dim,
+                         num_heads,
+                         kv_dim,
+                         norm_layer,
+                         quant_config=quant_config,
+                         prefix=prefix)
 
         self.max_size = max_size
         self._set_2d_pos_cache(self.max_size)
@@ -404,7 +410,10 @@ class MiniCPMVBaseModel(nn.Module, SupportsMultiModal, SupportsPP):
         self.vision_dim = (self.vpm.embed_dim if self.version == (2, 0) else
                            self.vpm.embeddings.embed_dim)
         self.embed_dim = self.config.hidden_size
-        self.resampler = self.init_resampler(self.embed_dim, self.vision_dim)
+        self.resampler = self.init_resampler(self.embed_dim,
+                                             self.vision_dim,
+                                             quant_config=quant_config,
+                                             prefix="resampler")
         self.resampler.to(device="cuda", dtype=param_dtype)
         # TODO: why is there _KEYS_TO_MODIFY_MAPPING? lm_head should be in llm
         self.lm_head = ParallelLMHead(config.vocab_size,
@@ -666,7 +675,11 @@ class MiniCPMVBaseModel(nn.Module, SupportsMultiModal, SupportsPP):
     ) -> nn.Module:
         raise NotImplementedError
 
-    def init_resampler(self, embed_dim: int, vision_dim: int) -> nn.Module:
+    def init_resampler(self,
+                       embed_dim: int,
+                       vision_dim: int,
+                       quant_config: Optional[QuantizationConfig] = None,
+                       prefix: str = "") -> nn.Module:
         raise NotImplementedError
 
     def get_vision_embedding(
@@ -743,16 +756,21 @@ class MiniCPMV2_0(MiniCPMVBaseModel):
     def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
         return self.model.embed_tokens(input_ids)
 
-    def init_resampler(self, embed_dim: int, vision_dim: int) -> nn.Module:
+    def init_resampler(self,
+                       embed_dim: int,
+                       vision_dim: int,
+                       quant_config: Optional[QuantizationConfig] = None,
+                       prefix: str = "") -> nn.Module:
         with set_default_torch_dtype(torch.float16):
-            resampler = Resampler2(
-                embed_dim=embed_dim,
+            resampler = Resampler2(embed_dim=embed_dim,
                                    num_heads=embed_dim // 128,
-                grid_size=int(math.sqrt(self.config.query_num)),
+                                   grid_size=int(
+                                       math.sqrt(self.config.query_num)),
                                    kv_dim=vision_dim,
                                    adaptive=False,
                                    do_post_projection=True,
-            )
+                                   quant_config=quant_config,
+                                   prefix=prefix)
 
         return resampler
 
@@ -825,9 +843,21 @@ class MiniCPMV2_5(MiniCPMVBaseModel, SupportsLoRA):
         ".k_proj.",
         ".v_proj.",
         ".o_proj.",
+        # vision encoder
+        ".fc1.",
+        ".fc2.",
+        # Currently, vllm does not support BNB quantization for the `out_proj`
+        # of the resampler, so it's necessary to distinguish between the
+        # vision encoder and the resampler's out_proj. The same applies to
+        # MiniCPMV2_6.
+        ".self_attn.out_proj.",  #  vision encoder out_proj
+        # resampler
+        ".kv_proj.",
     ]
     # in TP, these weights are partitioned along the column dimension (dim=-1)
-    column_parallel_weights_modules = [".down_proj.", ".o_proj."]
+    column_parallel_weights_modules = [
+        ".down_proj.", ".o_proj.", ".self_attn.out_proj.", ".fc2."
+    ]
     bitsandbytes_stacked_params_mapping = {
         # shard_name, weight_name, index
         "q_proj": ("qkv_proj", 0),
@@ -877,14 +907,18 @@ class MiniCPMV2_5(MiniCPMVBaseModel, SupportsLoRA):
             model.encoder.layers = model.encoder.layers[:-1]
         return model
 
-    def init_resampler(self, embed_dim: int, vision_dim: int) -> nn.Module:
+    def init_resampler(self,
+                       embed_dim: int,
+                       vision_dim: int,
+                       quant_config: Optional[QuantizationConfig] = None,
+                       prefix: str = "") -> nn.Module:
         with set_default_torch_dtype(torch.float16):
-            resampler = Resampler2_5(
-                num_queries=self.config.query_num,
+            resampler = Resampler2_5(num_queries=self.config.query_num,
                                      embed_dim=embed_dim,
                                      num_heads=embed_dim // 128,
                                      kv_dim=vision_dim,
-            )
+                                     quant_config=quant_config,
+                                     prefix=prefix)
         return resampler
 
     def get_vision_embedding(
@@ -967,9 +1001,17 @@ class MiniCPMV2_6(MiniCPMVBaseModel, SupportsLoRA):
         ".k_proj.",
         ".v_proj.",
         ".o_proj.",
+        # vision encoder
+        ".fc1.",
+        ".fc2.",
+        ".self_attn.out_proj.",
+        # resampler
+        ".kv_proj.",
     ]
     # in TP, these weights are partitioned along the column dimension (dim=-1)
-    column_parallel_weights_modules = [".down_proj.", ".o_proj."]
+    column_parallel_weights_modules = [
+        ".down_proj.", ".o_proj.", ".self_attn.out_proj.", ".fc2."
+    ]
     bitsandbytes_stacked_params_mapping = {
         # shard_name, weight_name, index
         "q_proj": ("qkv_proj", 0),
@@ -1019,15 +1061,19 @@ class MiniCPMV2_6(MiniCPMVBaseModel, SupportsLoRA):
             model.encoder.layers = model.encoder.layers[:-1]
         return model
 
-    def init_resampler(self, embed_dim: int, vision_dim: int) -> nn.Module:
+    def init_resampler(self,
+                       embed_dim: int,
+                       vision_dim: int,
+                       quant_config: Optional[QuantizationConfig] = None,
+                       prefix: str = "") -> nn.Module:
         with set_default_torch_dtype(torch.float16):
             # The resampler in 2.6 remains consistent with the one in 2.5.
-            resampler = Resampler2_5(
-                num_queries=self.config.query_num,
+            resampler = Resampler2_5(num_queries=self.config.query_num,
                                      embed_dim=embed_dim,
                                      num_heads=embed_dim // 128,
                                      kv_dim=vision_dim,
-            )
+                                     quant_config=quant_config,
+                                     prefix=prefix)
         return resampler
 
     def get_vision_embedding(

vllm/model_executor/models/mllama.py

@@ -1056,9 +1056,14 @@ class MllamaForConditionalGeneration(nn.Module, SupportsMultiModal):
         ".k_proj.",
         ".v_proj.",
         ".o_proj.",
+        ".fc1.",
+        ".fc2.",
+        # The `multi_modal_projector` is at the top level of the model,
+        # so we can't add a dot in front of it.
+        "multi_modal_projector."
     ]
     # in TP, these weights are partitioned along the column dimension (dim=-1)
-    column_parallel_weights_modules = [".down_proj.", ".o_proj."]
+    column_parallel_weights_modules = [".down_proj.", ".o_proj.", ".fc2."]
     bitsandbytes_stacked_params_mapping = {
         # shard_name, weight_name, index
         "q_proj": ("qkv_proj", 0),