issue/248 - change default attn backend to classic impl

README.md

@@ -176,5 +176,5 @@ python scripts/test_ppl.py --model-path MODEL_PATH [--ndev NDEV] [--max-batch MA
       ```
     - 选择attention后端 (使用flash attention后端需要先在InfiniCore完成相关配置和编译)
       ```bash
-      python examples/bench.py --nvidia --model=<model-path> --enable-paged-attn [--attn=flash-attn | --attn=default]
+      python examples/bench.py --nvidia --model=<model-path> --enable-paged-attn [--attn=default | --attn=flash-attn]
       ```

examples/bench.py

@@ -255,7 +255,7 @@ def get_args():
     parser.add_argument(
         "--attn",
         type=str,
-        default="flash-attn",
+        default="default",
         choices=["default", "flash-attn"],
         help="attention backend to use: 'default' or 'flash-attn'",
     )
@@ -285,7 +285,7 @@ class TestModel:
         skip_load=False,
         cache_config=None,
         enable_graph=False,
-        attn_backend="flash-attn",
+        attn_backend="default",
     ) -> None:
         model_path = os.path.expanduser(model_path)
         # ---------------------------------------------------------------------------- #

examples/jiuge.py

@@ -145,7 +145,7 @@ def get_args():
     parser.add_argument(
         "--attn",
         type=str,
-        default="flash-attn",
+        default="default",
         choices=["default", "flash-attn"],
         help="attention backend to use: 'default' or 'flash-attn'",
     )
@@ -164,7 +164,7 @@ def test(
     top_k=1,
     top_p=1.0,
     temperature=1.0,
-    attn_backend="flash-attn",
+    attn_backend="default",
 ):
     model_path = os.path.expanduser(model_path)
     # ---------------------------------------------------------------------------- #

python/infinilm/infer_engine.py

@@ -29,7 +29,7 @@ class InferEngine(_infinilm.InferEngine):
         distributed_config=DistConfig(1),
         cache_config=None,
         enable_graph_compiling=False,
-        attention_backend="flash-attn",
+        attention_backend="default",
     ):
         self.config = AutoConfig.from_pretrained(model_path)