add qwen3 eagle test

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

Author: crazydemo

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

@@ -1,7 +1,5 @@
 meta-llama/Llama-3.3-70B-Instruct:
   - accuracy: 45.96
-  - spec_dec_algo: Eagle
-    accuracy: 45.96
   - quant_algo: NVFP4
     kv_cache_quant_algo: FP8
     accuracy: 45.55

tests/integration/defs/accuracy/references/gsm8k.yaml

@@ -9,8 +9,6 @@ meta-llama/Llama-3.1-8B-Instruct:
     accuracy: 72.85
 meta-llama/Llama-3.3-70B-Instruct:
   - accuracy: 83.78
-  - spec_dec_algo: Eagle
-    accuracy: 83.78
   - quant_algo: NVFP4
     kv_cache_quant_algo: FP8
     accuracy: 75.61

tests/integration/defs/accuracy/test_llm_api_pytorch.py

@@ -229,6 +229,29 @@ def test_fp8_beam_search(self):
                           sampling_params=sampling_params,
                           extra_acc_spec="beam_width=4")
 
+    @skip_pre_hopper
+    def test_eagle3(self):
+        pytorch_config = dict(
+            disable_overlap_scheduler=True,
+            cuda_graph_config=CudaGraphConfig(batch_sizes=[1]),
+        )
+        kv_cache_config = KvCacheConfig(enable_block_reuse=False)
+
+        eagle_model_dir = f"{llm_models_root()}/EAGLE3-LLaMA3.1-Instruct-8B"
+        target_model_dir = f"{llm_models_root()}/llama-3.1-model/Llama-3.1-8B-Instruct"
+
+        draft_len = 4
+        spec_config = EagleDecodingConfig(max_draft_len=draft_len,
+                                          speculative_model_dir=eagle_model_dir)
+
+        with LLM(model=target_model_dir,
+                 **pytorch_config,
+                 kv_cache_config=kv_cache_config,
+                 speculative_config=spec_config,
+                 build_config=None) as llm:
+            task = MMLU(self.MODEL_NAME)
+            task.evaluate(llm)
+
     @skip_pre_hopper
     def test_ngram(self):
         pytorch_config = dict(
@@ -341,7 +364,7 @@ def test_auto_dtype_tp8(self):
                           extra_evaluator_kwargs=dict(apply_chat_template=True))
 
     @pytest.mark.skip_less_mpi_world_size(8)
-    @pytest.mark.parametrize("eagle3_one_model", [True, False])
+    @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"
         eagle_model_dir = f"{llm_models_root()}/EAGLE3-LLaMA3.3-Instruct-70B"

tests/integration/defs/examples/test_llama.py

@@ -4069,8 +4069,7 @@ def test_llm_api_lookahead_decoding_1gpu(model_name, model_path):
     """
     from defs.conftest import llm_models_root
 
-    from tensorrt_llm._tensorrt_engine import LLM
-    from tensorrt_llm.llmapi import (BuildConfig, KvCacheConfig,
+    from tensorrt_llm.llmapi import (LLM, BuildConfig, KvCacheConfig,
                                      LookaheadDecodingConfig, SamplingParams)
     build_config = BuildConfig(max_batch_size=128,
                                max_input_len=2048,

tests/integration/test_lists/qa/examples_test_list.txt

@@ -441,15 +441,16 @@ accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_chunked_prefill[
 accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_chunked_prefill[attn_backend=TRTLLM]
 accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_fp8_llm_sampler
 accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_fp8_beam_search
+accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_eagle3
 accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_ngram
 accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_guided_decoding[xgrammar]
 accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_guided_decoding[llguidance]
 accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_guided_decoding_4gpus[xgrammar]
 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_tp8[one_model]
-accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_eagle3_tp8[two_model]
+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::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]
@@ -497,6 +498,7 @@ accuracy/test_llm_api_pytorch.py::TestQwen3_235B_A22B::test_fp8[throughput_laten
 accuracy/test_llm_api_pytorch.py::TestQwen3_235B_A22B::test_fp8[latency]
 accuracy/test_llm_api_pytorch.py::TestQwen3_235B_A22B::test_nvfp4[latency_moe_cutlass]
 accuracy/test_llm_api_pytorch.py::TestQwen3_235B_A22B::test_nvfp4[latency_moe_trtllm]
+accuracy/test_llm_api_pytorch.py::TestQwen3_235B_A22B::test_nvfp4[latency_moe_trtllm_eagle3]
 accuracy/test_llm_api_pytorch.py::TestKanana_Instruct::test_auto_dtype
 accuracy/test_llm_api_pytorch.py::TestBielik11BInstruct::test_auto_dtype
 accuracy/test_llm_api_pytorch.py::TestBielik11BInstruct::test_fp8