feat: capture decode-step KV cache in auto-generated module configs

.github/workflows/test_pull_request.yaml

@@ -553,8 +553,11 @@ jobs:
         config:
           - granite_4_1_8b_spyre.yaml
           - granite_3_3_8b_instruct_spyre.yaml
+	  - Meta-Llama-3.1-8B-Instruct_spyre.yaml
           - Ministral-3-14B-Instruct-2512.yaml
           - Mistral-Small-3.2-24B-Instruct-2506.yaml
+	  - Qwen2.5-7B-Instruct_spyre.yaml
+	  - gpt_oss_20b_spyre.yaml
 
     steps:
       - name: Checkout composite actions

tests/test_modules_custom.py

@@ -451,6 +451,13 @@ def test_eager_vs_compile(self, device, dtype, module_info, training):
         3. Runs module in compile mode on Spyre
         4. Compares outputs between eager CPU, eager Spyre, and compile Spyre
         """
+
+        cls_name = module_info.module_cls.__name__
+        if cls_name.endswith("DecoderLayer"):
+            self.skipTest(
+                f"{cls_name}: skip the whole decoder layer due to time-consuming module"
+            )
+
         module_inputs = module_info.module_inputs_func(
             module_info, device=device, dtype=dtype, requires_grad=False, training=False
         )
@@ -581,6 +588,13 @@ def test_with_cpu(self, device, dtype, module_info, training):
         """
         run_compile = os.getenv("TEST_COMPILE_WITH_CPU", "1") == "1"
         run_eager = os.getenv("TEST_EAGER_WITH_CPU", "0") == "1"
+
+        cls_name = module_info.module_cls.__name__
+        if cls_name.endswith("DecoderLayer"):
+            self.skipTest(
+                f"{cls_name}: skip the whole decoder layer due to time-consuming module"
+            )
+
         module_inputs = module_info.module_inputs_func(
             module_info,
             device=device,