[Gemma2] add bitsandbytes support for Gemma2 (vllm-project#8338)

Signed-off-by: Alvant <alvasian@yandex.ru>
compressa-ai · Oct 26, 2024 · 3c36559 · 3c36559
1 parent 1ca7cb3
commit 3c36559
Showing 1 changed file with 8 additions and 0 deletions.
diff --git a/vllm/model_executor/models/gemma2.py b/vllm/model_executor/models/gemma2.py
@@ -312,6 +312,14 @@ class Gemma2ForCausalLM(nn.Module, SupportsLoRA):
     # Gemma does not apply LoRA to the embedding layer.
     embedding_modules = {}
     embedding_padding_modules = []
+    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,