From 0b083ef5b3ee743e1a05c31caa7d8a9efa831c1a Mon Sep 17 00:00:00 2001 From: "Xin He (SW-GPU)" <200704525+xinhe-nv@users.noreply.github.com> Date: Tue, 22 Jul 2025 17:50:30 +0800 Subject: [PATCH 1/2] fix test_fp8_tp4 Signed-off-by: Xin He (SW-GPU) <200704525+xinhe-nv@users.noreply.github.com> --- tests/integration/defs/accuracy/test_llm_api_pytorch.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/tests/integration/defs/accuracy/test_llm_api_pytorch.py b/tests/integration/defs/accuracy/test_llm_api_pytorch.py index 204094787043..4993fc7a35b5 100644 --- a/tests/integration/defs/accuracy/test_llm_api_pytorch.py +++ b/tests/integration/defs/accuracy/test_llm_api_pytorch.py @@ -358,7 +358,10 @@ def test_auto_dtype_tp8(self): @pytest.mark.skip_device_not_contain(["H100", "H200", "B200"]) def test_fp8_tp4(self): model_path = f"{llm_models_root()}/modelopt-hf-model-hub/Llama-3.3-70B-Instruct-fp8" - with LLM(model_path, tensor_parallel_size=4) as llm: + kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.6) + with LLM(model_path, + tensor_parallel_size=4, + kv_cache_config=kv_cache_config) as llm: assert llm.args.quant_config.quant_algo == QuantAlgo.FP8 task = MMLU(self.MODEL_NAME) task.evaluate(llm) From 2b5a103f1d49b451f50a1a22cda3bc4d9d9fc752 Mon Sep 17 00:00:00 2001 From: "Xin He (SW-GPU)" <200704525+xinhe-nv@users.noreply.github.com> Date: Thu, 24 Jul 2025 17:27:59 +0800 Subject: [PATCH 2/2] skip TestLlama3_3_70BInstruct on CG4 Signed-off-by: Xin He (SW-GPU) <200704525+xinhe-nv@users.noreply.github.com> --- tests/integration/defs/accuracy/test_llm_api_pytorch.py | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/tests/integration/defs/accuracy/test_llm_api_pytorch.py b/tests/integration/defs/accuracy/test_llm_api_pytorch.py index 4993fc7a35b5..4848b2d02f08 100644 --- a/tests/integration/defs/accuracy/test_llm_api_pytorch.py +++ b/tests/integration/defs/accuracy/test_llm_api_pytorch.py @@ -339,6 +339,8 @@ def test_fp8_prequantized(self): @pytest.mark.timeout(7200) +@pytest.mark.skip_less_host_memory(1000000) +# 1TB is basic requirement for large model tests. CG4 120G only has 800G host memory, and 480G is shared with GPUs. the test will cause the system crash. class TestLlama3_3_70BInstruct(LlmapiAccuracyTestHarness): MODEL_NAME = "meta-llama/Llama-3.3-70B-Instruct" @@ -355,7 +357,7 @@ def test_auto_dtype_tp8(self): extra_evaluator_kwargs=dict(apply_chat_template=True)) @pytest.mark.skip_less_device(4) - @pytest.mark.skip_device_not_contain(["H100", "H200", "B200"]) + @skip_pre_hopper def test_fp8_tp4(self): model_path = f"{llm_models_root()}/modelopt-hf-model-hub/Llama-3.3-70B-Instruct-fp8" kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.6) @@ -372,7 +374,7 @@ def test_fp8_tp4(self): extra_evaluator_kwargs=dict(apply_chat_template=True)) @pytest.mark.skip_less_device(4) - @pytest.mark.skip_device_not_contain(["B200"]) + @skip_pre_blackwell def test_nvfp4_tp4(self): model_path = f"{llm_models_root()}/modelopt-hf-model-hub/Llama-3.3-70B-Instruct-fp4" with LLM(model_path, tensor_parallel_size=4) as llm: