[NVIDIA#10707][fix] AutoDeploy: Super accuracy test fixes

- Initial PR NVIDIA#10308 added wrong test name in L0 config files - fix
- Add fp8 test
- Add (disabled) fp4 test
- Slightly decrease bf16 mmlu to accommodate autodeploy test

Signed-off-by: Gal Hubara Agam <96368689+galagam@users.noreply.github.com>

Author: galagam

tests/integration/defs/accuracy/references/mmlu.yaml

@@ -349,10 +349,10 @@ mistral/Mistral-Large-3-675B:
   - spec_dec_algo: Eagle
     accuracy: 85.30
 nvidia/Nemotron-Super-V3:
-  - accuracy: 81.07
+  - accuracy: 80.00
   - quant_algo: FP8
     kv_cache_quant_algo: FP8
-    accuracy: 78.22
+    accuracy: 77.80
   - quant_algo: NVFP4
     kv_cache_quant_algo: FP8
     accuracy: 77.56

tests/integration/defs/accuracy/test_llm_api_autodeploy.py

@@ -20,6 +20,7 @@
 from tensorrt_llm.quantization import QuantAlgo
 from tensorrt_llm.sampling_params import SamplingParams
 
+from ..conftest import get_device_count, llm_models_root
 from .accuracy_core import GSM8K, MMLU, CnnDailymail, LlmapiAccuracyTestHarness
 
 
@@ -244,6 +245,9 @@ class TestNemotronSuperV3(LlmapiAccuracyTestHarness):
 
     MODEL_NAME = "nvidia/Nemotron-Super-V3"
     MODEL_PATH_BF16 = f"{llm_models_root()}/Nemotron-Super-3-120B-A12B-dev"
+    MODEL_PATH_FP8 = f"{llm_models_root()}/Nemotron-SuperV3-phase1-mtp-fp8-fp8kv"
+    MODEL_PATH_FP4 = f"{llm_models_root()}/Nemotron-SuperV3-phase1-mtp-nvfp4-fp8kv"
+
     # Set minimum possible seq len + small buffer, for test speed & memory usage
     MAX_SEQ_LEN = max(MMLU.MAX_INPUT_LEN + MMLU.MAX_OUTPUT_LEN,
                       GSM8K.MAX_INPUT_LEN + GSM8K.MAX_OUTPUT_LEN)
@@ -289,3 +293,45 @@ def test_bf16(self):
             task.evaluate(llm, sampling_params=sampling_params)
             task = GSM8K(self.MODEL_NAME)
             task.evaluate(llm)
+
+    @pytest.mark.skip_less_device_memory(180000)
+    @pytest.mark.skip_less_device(4)
+    @pytest.mark.parametrize("world_size", [4, 8])
+    def test_fp8(self, world_size):
+        if get_device_count() < world_size:
+            pytest.skip("Not enough devices for world size, skipping test")
+        kwargs = self.get_default_kwargs()
+        sampling_params = self.get_default_sampling_params()
+        with AutoDeployLLM(model=self.MODEL_PATH_FP8,
+                           tokenizer=self.MODEL_PATH_FP8,
+                           world_size=world_size,
+                           **kwargs) as llm:
+            # Manually set quant_config for FP8 model to get the accuracy threshold
+            llm.args.quant_config.quant_algo = QuantAlgo.FP8
+            llm.args.quant_config.kv_cache_quant_algo = QuantAlgo.FP8
+
+            task = MMLU(self.MODEL_NAME)
+            task.evaluate(llm, sampling_params=sampling_params)
+            task = GSM8K(self.MODEL_NAME)
+            task.evaluate(llm)
+
+    @pytest.mark.skip("Skipping FP4 test until it is supported")
+    @pytest.mark.skip_less_device_memory(180000)
+    @pytest.mark.parametrize("world_size", [1, 4, 8])
+    def test_fp4(self, world_size):
+        if get_device_count() < world_size:
+            pytest.skip("Not enough devices for world size, skipping test")
+        kwargs = self.get_default_kwargs()
+        sampling_params = self.get_default_sampling_params()
+        with AutoDeployLLM(model=self.MODEL_PATH_FP4,
+                           tokenizer=self.MODEL_PATH_FP4,
+                           world_size=world_size,
+                           **kwargs) as llm:
+            # Manually set quant_config for FP4 model to get the accuracy threshold
+            llm.args.quant_config.quant_algo = QuantAlgo.NVFP4
+            llm.args.quant_config.kv_cache_quant_algo = QuantAlgo.NVFP4
+
+            task = MMLU(self.MODEL_NAME)
+            task.evaluate(llm, sampling_params=sampling_params)
+            task = GSM8K(self.MODEL_NAME)
+            task.evaluate(llm)

tests/integration/test_lists/test-db/l0_dgx_b200.yml

@@ -218,4 +218,5 @@ l0_dgx_b200:
   tests:
   - unittest/_torch/auto_deploy/unit/multigpu
   - accuracy/test_llm_api_autodeploy.py::TestLlama3_1_8B::test_auto_dtype[False-4]
-  - accuracy/test_llm_api_autodeploy.py::TestNemotronMOE::test_bf16
+  - accuracy/test_llm_api_autodeploy.py::TestNemotronSuperV3::test_bf16[4]
+  - accuracy/test_llm_api_autodeploy.py::TestNemotronSuperV3::test_fp8[4]

tests/integration/test_lists/test-db/l0_dgx_h100.yml

@@ -322,4 +322,5 @@ l0_dgx_h100:
   tests:
   - unittest/_torch/auto_deploy/unit/multigpu
   - accuracy/test_llm_api_autodeploy.py::TestLlama3_1_8B::test_auto_dtype[False-4]
-  - accuracy/test_llm_api_autodeploy.py::TestNemotronMOE::test_bf16
+  - accuracy/test_llm_api_autodeploy.py::TestNemotronSuperV3::test_bf16[4]
+  - accuracy/test_llm_api_autodeploy.py::TestNemotronSuperV3::test_fp8[4]