Feature add lora support for Qwen2 (#3177)

csrc/punica/bgmv/bgmv_config.h

@@ -21,6 +21,7 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
     f(in_T, out_T, W_T, narrow, 2048) \
     f(in_T, out_T, W_T, narrow, 2560) \
     f(in_T, out_T, W_T, narrow, 2752) \
+    f(in_T, out_T, W_T, narrow, 2816) \
     f(in_T, out_T, W_T, narrow, 3072) \
     f(in_T, out_T, W_T, narrow, 3456) \
     f(in_T, out_T, W_T, narrow, 3584) \
@@ -36,6 +37,7 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
     f(in_T, out_T, W_T, narrow, 10240) \
     f(in_T, out_T, W_T, narrow, 11008) \
     f(in_T, out_T, W_T, narrow, 12288) \
+    f(in_T, out_T, W_T, narrow, 13696) \
     f(in_T, out_T, W_T, narrow, 13824) \
     f(in_T, out_T, W_T, narrow, 14336) \
     f(in_T, out_T, W_T, narrow, 16384) \

vllm/model_executor/models/qwen2.py

@@ -46,6 +46,7 @@ from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.model_executor.weight_utils import (default_weight_loader,
                                               hf_model_weights_iterator)
 from vllm.sequence import SamplerOutput
+from vllm.config import LoRAConfig
 
 KVCache = Tuple[torch.Tensor, torch.Tensor]
 
@@ -264,12 +265,35 @@ class Qwen2Model(nn.Module):
 
 
 class Qwen2ForCausalLM(nn.Module):
+    packed_modules_mapping = {
+        "qkv_proj": [
+            "q_proj",
+            "k_proj",
+            "v_proj",
+        ],
+        "gate_up_proj": [
+            "gate_proj",
+            "up_proj",
+        ],
+    }
+
+    # LoRA specific attributes
+    supported_lora_modules = [
+        "qkv_proj",
+        "o_proj",
+        "gate_up_proj",
+        "down_proj",
+    ]
+    embedding_modules = {}
+    embedding_padding_modules = []
 
     def __init__(
         self,
         config: Qwen2Config,
         linear_method: Optional[LinearMethodBase] = None,
+        lora_config: Optional[LoRAConfig] = None,
     ) -> None:
+        del lora_config
         super().__init__()
         self.config = config
         self.linear_method = linear_method