[nightly ut] reduce memory usage and skip deepseek on machines without avx512_bf16 support (#3664)

Author: blzheng

tests/cpu/hf_configs/mllama/config.json

@@ -39,7 +39,7 @@
         "forced_bos_token_id": null,
         "forced_eos_token_id": null,
         "hidden_act": "silu",
-        "hidden_size": 4096,
+        "hidden_size": 1024,
         "id2label": {
             "0": "LABEL_0",
             "1": "LABEL_1"

tests/cpu/test_ipex_optimize_transformers_nightly.py

@@ -302,7 +302,7 @@ def model_replacement_check(
         elif m.name == "jamba":
             model.config.dtype = dtype
         model.eval()
-        ref_m = copy.deepcopy(model)
+        ref_m = model
         ipex_m = copy.deepcopy(model)
         ipex_m = ipex.llm.optimize(
             ipex_m, dtype=dtype, deployment_mode=deployment_mode, inplace=True
@@ -428,6 +428,11 @@ def test_model_replacement(self):
         ):
             if torchcompile and deployment_mode:
                 continue
+            if (
+                m.name in ["deepseekv2", "deepseekv3"]
+                and not core.isa_has_avx512_bf16_support()
+            ):
+                continue
             self.model_replacement_check(m, dtype, jit, torchcompile, ret_dict)
         _disable_tpp()