[Misc]Further reduce BNB static variable (#10597)

Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
2024-11-27 14:54:12 +08:00
parent e85250b1d1
commit 15cc2a9f1a
14 changed files with 131 additions and 219 deletions
--- a/vllm/model_executor/models/qwen.py
+++ b/vllm/model_executor/models/qwen.py
@@ -1028,12 +1028,7 @@ class QWenLLM(QWenBaseModel):
    embedding_modules = {}
    embedding_padding_modules = []

-    default_bitsandbytes_target_modules = [
-        ".c_attn.",
-        ".c_proj.",
-        ".w1.",
-        ".w2.",
-    ]
+    # BitandBytes specific attributes
    bitsandbytes_stacked_params_mapping = {
        # shard_name, weight_name, index
        "w2": ("gate_up_proj", 0),