fix

tests/generation/test_utils.py

@@ -2298,7 +2298,7 @@ class GenerationTesterMixin:
         NOTE: despite the test logic being the same, different implementations actually need different decorators, hence
         this separate function.
         """
-        max_new_tokens = 30
+        max_new_tokens = 5
         support_flag = {
             "sdpa": "_supports_sdpa",
             "flash_attention_2": "_supports_flash_attn_2",
@@ -2351,8 +2351,6 @@ class GenerationTesterMixin:
                     attn_implementation="eager",
                 ).to(torch_device)
                 res_eager = model_eager.generate(**inputs_dict, **generate_kwargs)
-                del model_eager
-                gc.collect()
 
                 model_attn = model_class.from_pretrained(
                     tmpdirname,
@@ -2360,8 +2358,6 @@ class GenerationTesterMixin:
                     attn_implementation=attn_implementation,
                 ).to(torch_device)
                 res_attn = model_attn.generate(**inputs_dict, **generate_kwargs)
-                del model_attn
-                gc.collect()
 
                 self._check_similar_generate_outputs(res_eager, res_attn, atol=1e-3, rtol=1e-3)