add llama4 maverick eagle3 accuracy case

Signed-off-by: Ivy Zhang <[email protected]>

Author: crazydemo

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

@@ -68,6 +68,10 @@ meta-llama/Llama-3.3-70B-Instruct:
     accuracy: 81.02
 meta-llama/Llama-4-Maverick-17B-128E-Instruct:
   - accuracy: 86.40
+  - quant_algo: FP8
+    kv_cache_quant_algo: FP8
+    spec_dec_algo: Eagle3
+    accuracy: 86.40
 meta-llama/Llama-4-Scout-17B-16E-Instruct:
   - accuracy: 80.00
   - quant_algo: NVFP4

tests/integration/defs/accuracy/test_llm_api_pytorch.py

@@ -464,6 +464,30 @@ def test_chunked_prefill(self, attn_backend):
             task = MMLU(self.MODEL_NAME)
             task.evaluate(llm)
 
+    @skip_pre_hopper
+    @pytest.mark.skip_less_mpi_world_size(8)
+    @pytest.mark.parametrize("tp_size,pp_size,ep_size", [(8, 1, 1)],
+                             ids=["tp8"])
+    def test_fp8_eagle3(self, cuda_graph, tp_size, pp_size, ep_size):
+        model_path = f"{llm_models_root()}/llama4-models/nvidia/Llama-4-Maverick-17B-128E-Instruct-FP8"
+        eagle_model_dir = f"{llm_models_root()}/Llama-4-Maverick-17B-128E-Eagle3"
+        spec_config = EagleDecodingConfig(max_draft_len=3,
+                                          speculative_model_dir=eagle_model_dir)
+        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.75)
+        pytorch_config = dict(
+            disable_overlap_scheduler=not cuda_graph,
+            cuda_graph_config=CudaGraphConfig(max_batch_size=8),
+            enable_attention_dp=False)
+        with LLM(model_path,
+                 kv_cache_config=kv_cache_config,
+                 tensor_parallel_size=tp_size,
+                 pipeline_parallel_size=pp_size,
+                 moe_expert_parallel_size=ep_size,
+                 **pytorch_config,
+                 speculative_config=spec_config) as llm:
+            task = MMLU(self.MODEL_NAME)
+            task.evaluate(llm)
+
 
 class TestLlama4ScoutInstruct(LlmapiAccuracyTestHarness):
     MODEL_NAME = "meta-llama/Llama-4-Scout-17B-16E-Instruct"

tests/integration/test_lists/qa/examples_test_list.txt

@@ -450,15 +450,16 @@ accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_guided_decoding_
 accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_guided_decoding_4gpus[llguidance]
 accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_fp8_tp4
 accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_nvfp4_tp4
-accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_eagle3[one_model]
-accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_eagle3[two_model]
+accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_eagle3_tp8[one_model]
+accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_eagle3_tp8[two_model]
 accuracy/test_llm_api_pytorch.py::TestMistral7B::test_auto_dtype
 accuracy/test_llm_api_pytorch.py::TestGemma3_1BInstruct::test_auto_dtype
 accuracy/test_llm_api_pytorch.py::TestLlama4MaverickInstruct::test_auto_dtype[tp8-cuda_graph=False]
 accuracy/test_llm_api_pytorch.py::TestLlama4MaverickInstruct::test_auto_dtype[tp8ep4-cuda_graph=True]
 accuracy/test_llm_api_pytorch.py::TestLlama4MaverickInstruct::test_auto_dtype[tp8ep8-cuda_graph=True]
 accuracy/test_llm_api_pytorch.py::TestLlama4MaverickInstruct::test_chunked_prefill[attn_backend=FLASHINFER]
 accuracy/test_llm_api_pytorch.py::TestLlama4MaverickInstruct::test_chunked_prefill[attn_backend=TRTLLM]
+accuracy/test_llm_api_pytorch.py::TestLlama4MaverickInstruct::test_fp8_eagle3[tp8]
 accuracy/test_llm_api_pytorch.py::TestLlama4ScoutInstruct::test_auto_dtype[tp8-cuda_graph=False]
 accuracy/test_llm_api_pytorch.py::TestLlama4ScoutInstruct::test_auto_dtype[tp8ep4-cuda_graph=True]
 accuracy/test_llm_api_pytorch.py::TestLlama4ScoutInstruct::test_auto_dtype[tp8ep8-cuda_graph=True]