improve tests

Signed-off-by: Xin He (SW-GPU) <[email protected]>

Author: xinhe-nv

tests/integration/defs/accuracy/test_disaggregated_serving.py

@@ -523,6 +523,7 @@ def test_ctx_pp_gen_tp_asymmetric(self, ctx_pp, gen_tp, testset):
 
 @pytest.mark.skip_less_device_memory(140000)
 @pytest.mark.timeout(3600)
+@pytest.mark.skip_less_device(8)
 class TestLlama4ScoutInstruct(LlmapiAccuracyTestHarness):
     MODEL_NAME = "meta-llama/Llama-4-Scout-17B-16E-Instruct"
     MODEL_PATH = f"{llm_models_root()}/llama4-models/Llama-4-Scout-17B-16E-Instruct"
@@ -561,6 +562,7 @@ def test_auto_dtype(self, overlap_scheduler):
 
 
 @pytest.mark.timeout(3600)
+@pytest.mark.skip_less_device(8)
 class TestDeepSeekV3Lite(LlmapiAccuracyTestHarness):
     MODEL_NAME = "deepseek-ai/DeepSeek-V3-Lite"
     MODEL_PATH = f"{llm_models_root()}/DeepSeek-V3-Lite/bf16"

tests/integration/defs/accuracy/test_llm_api_pytorch.py

@@ -250,7 +250,7 @@ def test_eagle3(self, overlap_scheduler, eagle3_one_model):
                                               enable_padding=True),
         )
         kv_cache_config = KvCacheConfig(
-            enable_block_reuse=True
+            enable_block_reuse=True, free_gpu_memory_fraction=0.8
         )  # both one-model and two-model supports this feature
 
         eagle_model_dir = f"{llm_models_root()}/EAGLE3-LLaMA3.1-Instruct-8B"
@@ -280,7 +280,8 @@ def test_ngram(self):
             cuda_graph_config=CudaGraphConfig(batch_sizes=[1]),
         )
 
-        kv_cache_config = KvCacheConfig(enable_block_reuse=False)
+        kv_cache_config = KvCacheConfig(enable_block_reuse=False,
+                                        free_gpu_memory_fraction=0.8)
 
         spec_config = NGramDecodingConfig(
             max_draft_len=4,