[None][fix] fix Llama3 eagle3 test case OOM (NVIDIA#6832)

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

Author: crazydemo

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

@@ -184,7 +184,8 @@ meta-llama/Llama-3.2-3B:
     kv_cache_quant_algo: FP8
     accuracy: 33.629
 meta-llama/Llama-3.3-70B-Instruct:
-  - spec_dec_algo: Eagle
+  - quant_algo: FP8
+    spec_dec_algo: Eagle
     accuracy: 33.244
   - quant_algo: NVFP4
     kv_cache_quant_algo: FP8

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

@@ -59,7 +59,8 @@ meta-llama/Llama-3.2-3B:
     accuracy: 60.60
 meta-llama/Llama-3.3-70B-Instruct:
   - accuracy: 81.31
-  - spec_dec_algo: Eagle
+  - quant_algo: FP8
+    spec_dec_algo: Eagle
     accuracy: 81.31
   - quant_algo: NVFP4
     kv_cache_quant_algo: FP8

tests/integration/defs/accuracy/test_llm_api_pytorch.py

@@ -480,25 +480,27 @@ def test_auto_dtype_tp8(self):
             task.evaluate(llm,
                           extra_evaluator_kwargs=dict(apply_chat_template=True))
 
+    @skip_pre_hopper
     @pytest.mark.skip_less_mpi_world_size(8)
     @parametrize_with_ids("eagle3_one_model", [True, False])
-    def test_eagle3_tp8(self, eagle3_one_model):
-        model_path = f"{llm_models_root()}/llama-3.3-models/Llama-3.3-70B-Instruct"
+    def test_fp8_eagle3_tp8(self, eagle3_one_model):
+        model_path = f"{llm_models_root()}/modelopt-hf-model-hub/Llama-3.3-70B-Instruct-fp8"
         eagle_model_dir = f"{llm_models_root()}/EAGLE3-LLaMA3.3-Instruct-70B"
         kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.6)
         spec_config = EagleDecodingConfig(max_draft_len=4,
                                           speculative_model_dir=eagle_model_dir,
                                           eagle3_one_model=eagle3_one_model)
-        pytorch_config = dict(disable_overlap_scheduler=True, )
+        pytorch_config = dict(
+            disable_overlap_scheduler=True,
+            cuda_graph_config=CudaGraphConfig(max_batch_size=1))
         with LLM(model_path,
+                 max_batch_size=16,
                  tensor_parallel_size=8,
                  speculative_config=spec_config,
                  kv_cache_config=kv_cache_config,
                  **pytorch_config) as llm:
             task = CnnDailymail(self.MODEL_NAME)
             task.evaluate(llm)
-            task = MMLU(self.MODEL_NAME)
-            task.evaluate(llm)
 
     @pytest.mark.skip_less_device(4)
     @skip_pre_hopper

tests/integration/test_lists/qa/llm_function_full.txt

@@ -460,8 +460,8 @@ accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_guided_decoding_
 accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_guided_decoding_with_ngram[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_tp8[eagle3_one_model=True]
-accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_eagle3_tp8[eagle3_one_model=False]
+accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_fp8_eagle3_tp8[eagle3_one_model=True]
+accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_fp8_eagle3_tp8[eagle3_one_model=False]
 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::TestMistralSmall24B::test_auto_dtype

tests/integration/test_lists/qa/llm_function_sanity.txt

@@ -75,10 +75,10 @@ accuracy/test_llm_api_pytorch.py::TestLlama3_2_1B::test_auto_dtype
 accuracy/test_llm_api_pytorch.py::TestLlama3_2_1B::test_fp8_prequantized
 accuracy/test_llm_api_pytorch.py::TestLlama3_2_3B::test_auto_dtype
 accuracy/test_llm_api_pytorch.py::TestLlama3_2_3B::test_fp8_prequantized
-accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_eagle3_tp8[eagle3_one_model=False]
-accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_eagle3_tp8[eagle3_one_model=True]
 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_fp8_eagle3_tp8[eagle3_one_model=True]
+accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_fp8_eagle3_tp8[eagle3_one_model=False]
 accuracy/test_llm_api_pytorch.py::TestLlama4MaverickInstruct::test_auto_dtype[tp4-cuda_graph=False]
 accuracy/test_llm_api_pytorch.py::TestLlama4MaverickInstruct::test_auto_dtype[tp4ep2-cuda_graph=True]
 accuracy/test_llm_api_pytorch.py::TestLlama4MaverickInstruct::test_auto_dtype[tp4ep4-cuda_graph=True]

tests/integration/test_lists/waives.txt

@@ -306,8 +306,6 @@ disaggregated/test_disaggregated.py::test_disaggregated_diff_max_tokens[TinyLlam
 disaggregated/test_disaggregated.py::test_disaggregated_deepseek_v3_lite_fp8_tp1_single_gpu_mtp[DeepSeek-V3-Lite-fp8] SKIP (https://nvbugs/5465642)
 examples/test_multimodal.py::test_llm_multimodal_general[Mistral-Small-3.1-24B-Instruct-2503-pp:1-tp:1-bfloat16-bs:1-cpp_e2e:False-nb:1] SKIP (https://nvbugs/5431146)
 accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_fp8_blockscale[latency] SKIP (https://nvbugs/5464461)
-full:H100/accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_eagle3_tp8[eagle3_one_model=True] SKIP (https://nvbugs/5467815)
-full:H100/accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_eagle3_tp8[eagle3_one_model=False] SKIP (https://nvbugs/5467815)
 full:H100/accuracy/test_llm_api_pytorch.py::TestLlama4ScoutInstruct::test_fp8[tp4-cuda_graph=True] SKIP (https://nvbugs/5467815)
 full:H100/accuracy/test_llm_api_pytorch.py::TestLlama4ScoutInstruct::test_fp8_chunked_prefill[tp4ep4-cuda_graph=True] SKIP (https://nvbugs/5467815)
 accuracy/test_disaggregated_serving.py::TestQwen3_30B_A3B::test_mixed_ctx_gen_model[ctxpp2gentp2] SKIP (https://nvbugs/5470769)