[Misc] Merge bitsandbytes_stacked_params_mapping and packed_modules_mapping (#11924)

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

vllm/model_executor/models/qwen.py

@@ -1028,13 +1028,6 @@ class QWenLLM(QWenBaseModel):
     embedding_modules = {}
     embedding_padding_modules = []
 
-    # BitandBytes specific attributes
-    bitsandbytes_stacked_params_mapping = {
-        # shard_name, weight_name, index
-        "w2": ("gate_up_proj", 0),
-        "w1": ("gate_up_proj", 1),
-    }
-
 
 class QWenVL(QWenBaseModel, SupportsMultiModal):
     packed_modules_mapping = {

vllm/model_executor/models/qwen2.py

@@ -418,16 +418,6 @@ class Qwen2ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
     embedding_modules = {}
     embedding_padding_modules = []
 
-    # BitandBytes specific attributes
-    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, *, vllm_config: VllmConfig, prefix: str = ""):
         super().__init__()
         config = vllm_config.model_config.hf_config

vllm/model_executor/models/qwen2_vl.py

@@ -1038,16 +1038,6 @@ class Qwen2VLForConditionalGeneration(nn.Module, SupportsMultiModal,
     embedding_modules = {}
     embedding_padding_modules = []
 
-    # BitandBytes specific attributes
-    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),
-    }
-
     # To ensure correct weight loading and mapping.
     hf_to_vllm_mapper = WeightsMapper(orig_to_new_prefix={
         "lm_head.": "language_model.lm_head.",

vllm/model_executor/models/solar.py

@@ -401,14 +401,6 @@ class SolarForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
         "lm_head": "output_embeddings",
     }
     embedding_padding_modules = ["lm_head"]
-    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, *, vllm_config: VllmConfig, prefix: str = ""):
         super().__init__()