[Model] Add BNB quantization support for Mllama (#9720)

Author: Isotr0py

vllm/model_executor/layers/quantization/bitsandbytes.py

@@ -3,6 +3,7 @@
 import torch
 
 from vllm.model_executor.layers.linear import (LinearBase, LinearMethodBase,
+                                               UnquantizedLinearMethod,
                                                set_weight_attrs)
 from vllm.model_executor.layers.quantization.base_config import (
     QuantizationConfig)
@@ -23,7 +24,7 @@ def __init__(
         bnb_4bit_use_double_quant: bool = False,
         llm_int8_enable_fp32_cpu_offload: bool = False,
         llm_int8_has_fp16_weight: bool = False,
-        llm_int8_skip_modules: Optional[Any] = None,
+        llm_int8_skip_modules: Optional[List[str]] = None,
         llm_int8_threshold: float = 0.0,
     ) -> None:
 
@@ -34,11 +35,15 @@ def __init__(
         self.bnb_4bit_use_double_quant = bnb_4bit_use_double_quant
         self.llm_int8_enable_fp32_cpu_offload = llm_int8_enable_fp32_cpu_offload
         self.llm_int8_has_fp16_weight = llm_int8_has_fp16_weight
-        self.llm_int8_skip_modules = llm_int8_skip_modules
+        self.llm_int8_skip_modules = llm_int8_skip_modules or []
         self.llm_int8_threshold = llm_int8_threshold
 
     def __repr__(self) -> str:
-        return "BitsAndBytesConfig"
+        return (f"BitsAndBytesConfig(load_in_8bit={self.load_in_8bit}, "
+                f"load_in_4bit={self.load_in_4bit}, "
+                f"bnb_4bit_compute_dtype={self.bnb_4bit_compute_dtype}, "
+                f"bnb_4bit_quant_type={self.bnb_4bit_quant_type}, "
+                f"llm_int8_skip_modules={self.llm_int8_skip_modules})")
 
     @classmethod
     def get_name(self) -> str:
@@ -102,15 +107,21 @@ def get_safe_value(config, keys, default_value=None):
             llm_int8_threshold=llm_int8_threshold)
 
     def get_quant_method(self, layer: torch.nn.Module,
-                         prefix: str) -> Optional["BitsAndBytesLinearMethod"]:
+                         prefix: str) -> Optional["LinearMethodBase"]:
         if isinstance(layer, LinearBase):
+            if is_layer_skipped_bnb(prefix, self.llm_int8_skip_modules):
+                return UnquantizedLinearMethod()
             return BitsAndBytesLinearMethod(self)
         return None
 
     def get_scaled_act_names(self) -> List[str]:
         return []
 
 
+def is_layer_skipped_bnb(prefix: str, llm_int8_skip_modules: List[str]):
+    return any(module_name in prefix for module_name in llm_int8_skip_modules)
+
+
 class BitsAndBytesLinearMethod(LinearMethodBase):
     """Linear method for BitsAndBytes.
 
@@ -211,6 +222,11 @@ def _apply_8bit_weight(
         from bitsandbytes import MatmulLtState, matmul
 
         original_type = x.dtype
+        original_shape = x.shape
+        reshape_after_matmul = False
+        if x.ndim > 2:
+            x = x.reshape(-1, x.size(-1))
+            reshape_after_matmul = True
         bf_x = x.to(torch.bfloat16)
 
         qweight = layer.qweight
@@ -265,6 +281,9 @@ def _apply_8bit_weight(
 
         out = out.to(original_type)
 
+        if reshape_after_matmul:
+            out = out.view(*original_shape[:-1], out.size(-1))
+
         if bias is not None:
             out += bias
 
@@ -282,6 +301,11 @@ def _apply_4bit_weight(
         from bitsandbytes import matmul_4bit
 
         original_type = x.dtype
+        original_shape = x.shape
+        reshape_after_matmul = False
+        if x.ndim > 2:
+            x = x.reshape(-1, x.size(-1))
+            reshape_after_matmul = True
         bf_x = x.to(torch.bfloat16)
 
         qweight = layer.qweight
@@ -310,6 +334,9 @@ def _apply_4bit_weight(
 
         out = out.to(original_type)
 
+        if reshape_after_matmul:
+            out = out.view(*original_shape[:-1], out.size(-1))
+
         if bias is not None:
             out += bias
 

vllm/model_executor/model_loader/loader.py

@@ -899,6 +899,19 @@ def _get_quantized_weights_iterator(
         return self._unquantized_generator(hf_weights_files, use_safetensors,
                                            quant_state_dict), quant_state_dict
 
+    def _is_8bit_weight_name(self, weight_name: str):
+        quantized_suffix = {".scb", ".weight_format"}
+        return any(weight_name.lower().endswith(suffix)
+                   for suffix in quantized_suffix)
+
+    def _is_4bit_weight_name(self, weight_name: str):
+        quantized_suffix = {
+            "absmax", "quant_map", "nested_absmax", "nested_quant_map",
+            "bitsandbytes"
+        }
+        suffix = weight_name.split(".")[-1]
+        return any(q_suffix in suffix for q_suffix in quantized_suffix)
+
     def _quantized_8bit_generator(self, hf_weights_files, use_safetensors,
                                   quant_state_dict) -> Generator:
         for weight_name, weight_tensor in self._hf_weight_iter(
@@ -912,7 +925,7 @@ def _quantized_8bit_generator(self, hf_weights_files, use_safetensors,
         for weight_name, weight_tensor in self._hf_weight_iter(
                 hf_weights_files, use_safetensors):
 
-            if not weight_name.endswith((".weight", ".bias")):
+            if self._is_8bit_weight_name(weight_name):
                 continue
 
             qweight_name = weight_name.replace(".weight", ".qweight")
@@ -932,7 +945,7 @@ def _quantized_4bit_generator(self, hf_weights_files, use_safetensors,
                                                use_safetensors)
         temp_state_dict = {}
         for weight_name, weight_tensor in weight_iterator:
-            if weight_name.endswith((".weight", ".bias")):
+            if not self._is_4bit_weight_name(weight_name):
                 continue
             # bitsandbytes library requires
             # weight.quant_state.bitsandbytes__* in CPU
@@ -956,7 +969,7 @@ def _parse_quant_state(param_name: str,
         for weight_name, weight_tensor in self._hf_weight_iter(
                 hf_weights_files, use_safetensors):
 
-            if not weight_name.endswith((".weight", ".bias")):
+            if self._is_4bit_weight_name(weight_name):
                 continue
 
             if (f"{weight_name}.quant_state.bitsandbytes__nf4" \

vllm/model_executor/models/mllama.py

@@ -325,7 +325,10 @@ def forward(self, hidden_state: torch.Tensor,
 # TODO: support other attention backends for attention in vision model
 class MllamaVisionSdpaAttention(nn.Module):
 
-    def __init__(self, config: config_mllama.MllamaVisionConfig):
+    def __init__(self,
+                 config: config_mllama.MllamaVisionConfig,
+                 quant_config: Optional[QuantizationConfig] = None,
+                 prefix: str = ""):
         super().__init__()
 
         model_parallel_size = get_tensor_model_parallel_world_size()
@@ -341,12 +344,16 @@ def __init__(self, config: config_mllama.MllamaVisionConfig):
             self.head_dim,
             self.num_heads,
             bias=False,
+            quant_config=quant_config,
+            prefix=f"{prefix}.qkv_proj",
         )
         self.o_proj = RowParallelLinear(
             self.num_heads * self.head_dim,
             self.embed_dim,
             bias=False,
             input_is_parallel=True,
+            quant_config=quant_config,
+            prefix=f"{prefix}.o_proj",
         )
 
     def forward(
@@ -393,7 +400,8 @@ def __init__(
         self.is_gated = is_gated
         self.intermediate_size = config.intermediate_size
 
-        self.self_attn = MllamaVisionSdpaAttention(config)
+        self.self_attn = MllamaVisionSdpaAttention(
+            config, quant_config=quant_config, prefix=f"{prefix}.self_attn")
         self.mlp = CLIPMLP(config,
                            quant_config=quant_config,
                            prefix=f"{prefix}.mlp")
@@ -1002,6 +1010,7 @@ def __init__(
             org_num_embeddings=config.vocab_size,
             padding_size=DEFAULT_VOCAB_PADDING_SIZE,
             quant_config=quant_config,
+            prefix=f"{prefix}.lm_head",
         )
 
     def forward(
@@ -1037,6 +1046,26 @@ def forward(
 @INPUT_REGISTRY.register_dummy_encoder_data(dummy_encoder_data_for_mllama)
 @INPUT_REGISTRY.register_input_processor(input_processor_for_mllama)
 class MllamaForConditionalGeneration(nn.Module, SupportsMultiModal):
+    # 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),
+    }
 
     def __init__(self,
                  config: config_mllama.MllamaConfig,
@@ -1061,10 +1090,13 @@ def __init__(self,
             quant_config=quant_config,
             prefix="language_model",
         )
-        self.multi_modal_projector = nn.Linear(
+        self.multi_modal_projector = ColumnParallelLinear(
             config.vision_config.vision_output_dim,
             config.text_config.hidden_size,
             bias=True,
+            quant_config=quant_config,
+            gather_output=True,
+            prefix="multi_modal_projector",
         )
         self.logits_processor = LogitsProcessor(config.output_hidden_states,
                                                 config.text_config.vocab_size)
@@ -1128,7 +1160,7 @@ def _parse_and_validate_image_input(self, **kwargs: object):
                 raise ValueError("No images provided.")
             max_num_tiles = max(
                 max([len(x) for x in y[0]]) for y in pixel_values)
-            device = self.multi_modal_projector.weight.device
+            device = next(self.multi_modal_projector.parameters()).device
             bsz = len(pixel_values)
             out_num_tiles = []
             out_images = torch.zeros(
@@ -1204,7 +1236,7 @@ def get_cross_attention_states(
         cross_attention_states = self.vision_model(pixel_values,
                                                    aspect_ratio_ids,
                                                    aspect_ratio_mask)
-        cross_attention_states = self.multi_modal_projector(
+        cross_attention_states, _ = self.multi_modal_projector(
             cross_attention_states)
 
         bsz, _, _, _, image_token_dim = tuple(cross_attention_states.shape)