remove phi3 empty cache (intel#10997)

python/llm/src/ipex_llm/transformers/convert.py

@@ -188,13 +188,6 @@ def is_linear_module(module):
     return result, (in_features, out_features, mp_group)
 
 
-def empty_cache_post(module, input, output):
-    try:
-        torch.xpu.empty_cache()
-    except:  # cpu
-        pass
-
-
 def convert_gptq(module, awq=False, llm_awq=False, act_order=False):
     from ipex_llm.transformers.low_bit_linear import get_block_size
     Q4_1 = get_block_size("asym_int4")
@@ -1524,8 +1517,6 @@ def safe_bmm_fwd(*args, **kwargs):
         from ipex_llm.transformers.models.phi3 import model_forward_wrapper
         model_forward = model_forward_wrapper(module.Phi3Model.forward)
         convert_forward(model, module.Phi3Model, model_forward)
-        # Empty cache after the first attention to run long context.
-        model.model.layers[0].self_attn.register_forward_hook(empty_cache_post)
     elif model.config.model_type == 'yuan':
         modeling_module_name = model.__class__.__module__
         module = importlib.import_module(modeling_module_name)