diff --git a/tests/integration/defs/accuracy/test_llm_api_pytorch.py b/tests/integration/defs/accuracy/test_llm_api_pytorch.py
index 22d04b26145..4098e2565b3 100644
--- a/tests/integration/defs/accuracy/test_llm_api_pytorch.py
+++ b/tests/integration/defs/accuracy/test_llm_api_pytorch.py
@@ -759,13 +759,16 @@ def test_auto_dtype(self):
         kv_cache_config = KvCacheConfig(
             enable_block_reuse=False,
             enable_partial_reuse=False,
+            free_gpu_memory_fraction=0.5,
         )
         # We use FlashInfer as the attention backend for Gemma3 VLM to support custom mask for images.
         # So, testing with it here.
         with LLM(self.MODEL_PATH,
                  kv_cache_config=kv_cache_config,
                  attn_backend="FLASHINFER",
-                 cuda_graph_config=None) as llm:
+                 cuda_graph_config=None,
+                 max_batch_size=128,
+                 max_seq_len=4096) as llm:
             task = CnnDailymail(self.MODEL_NAME)
             task.evaluate(llm)
             task = MMLU(self.MODEL_NAME)
diff --git a/tests/integration/defs/test_e2e.py b/tests/integration/defs/test_e2e.py
index bda61ce177b..0d47cc7c6f7 100644
--- a/tests/integration/defs/test_e2e.py
+++ b/tests/integration/defs/test_e2e.py
@@ -2232,7 +2232,8 @@ def test_ptp_quickstart_advanced_mixed_precision(llm_root, llm_venv):
                  marks=pytest.mark.skip_less_device_memory(80000)),
     pytest.param("gemma-3-27b-it",
                  "gemma/gemma-3-27b-it",
-                 marks=pytest.mark.skip_less_device_memory(80000)),
+                 marks=(skip_post_blackwell,
+                        pytest.mark.skip_less_device_memory(80000))),
 ])
 def test_ptp_quickstart_multimodal(llm_root, llm_venv, model_name, model_path,
                                    modality, use_cuda_graph):
@@ -2368,6 +2369,8 @@ def test_ptp_quickstart_multimodal(llm_root, llm_venv, model_name, model_path,
         cmd.append("--image_format=pil")
         cmd.append("--attention_backend=FLASHINFER")
         cmd.append("--disable_kv_cache_reuse")
+        cmd.append("--kv_cache_fraction=0.5")
+        cmd.append("--max_seq_len=1024")
 
     output = llm_venv.run_cmd(cmd, caller=check_output)
 
@@ -2528,9 +2531,10 @@ def test_ptp_quickstart_multimodal_phi4mm(llm_root, llm_venv, modality):
 @pytest.mark.skip_less_device(2)
 @pytest.mark.skip_less_device_memory(80000)
 @pytest.mark.parametrize("model_name,model_path", [
-    ("gemma-3-27b-it", "gemma/gemma-3-27b-it"),
     ("mistral-small-3.1-24b-instruct", "Mistral-Small-3.1-24B-Instruct-2503"),
     ("Phi-4-multimodal-instruct", "multimodals/Phi-4-multimodal-instruct"),
+    pytest.param(
+        "gemma-3-27b-it", "gemma/gemma-3-27b-it", marks=skip_post_blackwell),
 ])
 def test_ptp_quickstart_multimodal_2gpu(llm_root, llm_venv, model_name,
                                         model_path):
@@ -2602,6 +2606,8 @@ def test_ptp_quickstart_multimodal_2gpu(llm_root, llm_venv, model_name,
         cmd.append("--image_format=pil")
         cmd.append("--attention_backend=FLASHINFER")
         cmd.append("--disable_kv_cache_reuse")
+        cmd.append("--kv_cache_fraction=0.5")
+        cmd.append("--max_seq_len=1024")
     elif model_name == "Phi-4-multimodal-instruct":
         # Set max_seq_len to 4096 to use short rope factor.
         cmd.append("--max_seq_len=4096")
@@ -2630,9 +2636,10 @@ def test_ptp_quickstart_multimodal_2gpu(llm_root, llm_venv, model_name,
 
 @pytest.mark.skip_less_device_memory(80000)
 @pytest.mark.parametrize("model_name,model_path", [
-    ("gemma-3-27b-it", "gemma/gemma-3-27b-it"),
     ("mistral-small-3.1-24b-instruct", "Mistral-Small-3.1-24B-Instruct-2503"),
     ("Phi-4-multimodal-instruct", "multimodals/Phi-4-multimodal-instruct"),
+    pytest.param(
+        "gemma-3-27b-it", "gemma/gemma-3-27b-it", marks=skip_post_blackwell),
 ])
 def test_ptp_quickstart_multimodal_multiturn(llm_root, llm_venv, model_name,
                                              model_path):
@@ -2698,6 +2705,8 @@ def test_ptp_quickstart_multimodal_multiturn(llm_root, llm_venv, model_name,
         cmd.append("--image_format=pil")
         cmd.append("--attention_backend=FLASHINFER")
         cmd.append("--disable_kv_cache_reuse")
+        cmd.append("--kv_cache_fraction=0.5")
+        cmd.append("--max_seq_len=1024")
     elif model_name == "Phi-4-multimodal-instruct":
         # Set max_seq_len to 4096 to use short rope factor.
         cmd.append("--max_seq_len=4096")
diff --git a/tests/integration/test_lists/waives.txt b/tests/integration/test_lists/waives.txt
index ad1efc24b40..8c87ead0af0 100644
--- a/tests/integration/test_lists/waives.txt
+++ b/tests/integration/test_lists/waives.txt
@@ -281,7 +281,6 @@ full:GB200/examples/test_qwen.py::test_llm_qwen_7b_multi_gpus_summary[qwen1.5_7b
 full:GB200/examples/test_qwen.py::test_llm_qwen_7b_multi_gpus_summary[qwen2_7b_instruct-enable_fmha_fp32_acc-enable_plugin-tp2pp2-nb:4] SKIP (https://nvbugs/5247837)
 full:GB200/examples/test_qwen.py::test_llm_qwen_7b_multi_gpus_summary[qwen2_vl_7b_instruct-enable_fmha_fp32_acc-enable_plugin-tp2pp2-nb:4] SKIP (https://nvbugs/5359696)
 full:GB200/examples/test_qwen.py::test_llm_qwen_7b_multi_gpus_summary[qwen2.5_7b_chat-enable_fmha_fp32_acc-enable_plugin-tp2pp2-nb:4] SKIP (https://nvbugs/5247837)
-accuracy/test_llm_api_pytorch.py::TestMistralSmall24B::test_auto_dtype[/scratch.trt_llm_data/llm-models/Mistral-Small-3.1-24B-Instruct-2503-fp8-FP8] SKIP (https://nvbugs/5462007)
 accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_bfloat16[mtp_nextn=0-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False] SKIP (https://nvbugs/5410391)
 accuracy/test_llm_api.py::TestMistral_Nemo_12B_Base::test_fp8 SKIP (https://nvbugs/5413197)
 accuracy/test_cli_flow.py::TestLlama3_8BInstructGradient1048k::test_long_context_ppl SKIP (https://nvbugs/5413362)