add test

Signed-off-by: Enwei Zhu <[email protected]>

Author: syuoni

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

@@ -1,2 +1,6 @@
 meta-llama/Llama-3.1-8B-Instruct:
   - accuracy: 74.00
+  - spec_dec_algo: Eagle
+    accuracy: 74.00
+  - spec_dec_algo: NGram
+    accuracy: 74.00

tests/integration/defs/accuracy/test_llm_api_pytorch.py

@@ -302,9 +302,7 @@ def test_ngram(self):
     @pytest.mark.parametrize("backend", ["xgrammar", "llguidance"])
     def test_guided_decoding(self, backend: str, mocker):
         mocker.patch.dict(os.environ, {"TRTLLM_XGUIDANCE_LENIENT": "1"})
-        llm = LLM(self.MODEL_PATH,
-                  guided_decoding_backend=backend,
-                  cuda_graph_config=CudaGraphConfig())
+        llm = LLM(self.MODEL_PATH, guided_decoding_backend=backend)
         with llm:
             task = JsonModeEval(self.MODEL_NAME)
             task.evaluate(llm)
@@ -316,12 +314,42 @@ def test_guided_decoding_4gpus(self, backend: str, mocker):
         mocker.patch.dict(os.environ, {"TRTLLM_XGUIDANCE_LENIENT": "1"})
         with LLM(self.MODEL_PATH,
                  guided_decoding_backend=backend,
-                 cuda_graph_config=CudaGraphConfig(),
                  tensor_parallel_size=2,
                  pipeline_parallel_size=2) as llm:
             task = JsonModeEval(self.MODEL_NAME)
             task.evaluate(llm)
 
+    @skip_pre_hopper
+    @pytest.mark.parametrize("backend", ["xgrammar", "llguidance"])
+    def test_guided_decoding_with_eagle3(self, backend: str, mocker):
+        mocker.patch.dict(os.environ, {"TRTLLM_XGUIDANCE_LENIENT": "1"})
+        spec_config = EagleDecodingConfig(
+            max_draft_len=3,
+            speculative_model_dir=
+            f"{llm_models_root()}/EAGLE3-LLaMA3.1-Instruct-8B",
+            eagle3_one_model=False)
+        llm = LLM(self.MODEL_PATH,
+                  guided_decoding_backend=backend,
+                  speculative_config=spec_config,
+                  disable_overlap_scheduler=True)
+        with llm:
+            task = JsonModeEval(self.MODEL_NAME)
+            task.evaluate(llm)
+
+    @skip_pre_hopper
+    @pytest.mark.parametrize("backend", ["xgrammar", "llguidance"])
+    def test_guided_decoding_with_ngram(self, backend: str, mocker):
+        mocker.patch.dict(os.environ, {"TRTLLM_XGUIDANCE_LENIENT": "1"})
+        spec_config = NGramDecodingConfig(max_draft_len=3,
+                                          max_matching_ngram_size=3)
+        llm = LLM(self.MODEL_PATH,
+                  guided_decoding_backend=backend,
+                  speculative_config=spec_config,
+                  disable_overlap_scheduler=True)
+        with llm:
+            task = JsonModeEval(self.MODEL_NAME)
+            task.evaluate(llm)
+
 
 class TestLlama3_2_1B(LlmapiAccuracyTestHarness):
     MODEL_NAME = "meta-llama/Llama-3.2-1B"

tests/integration/test_lists/test-db/l0_h100.yml

@@ -30,6 +30,7 @@ l0_h100:
   - accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_bfloat16[attn_backend=TRTLLM-torch_compile=True]
   - accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_chunked_prefill[attn_backend=TRTLLM] TIMEOUT (90)
   - 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_with_eagle3[xgrammar]
   - accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_fp8[fp8kv=False-attn_backend=TRTLLM-torch_compile=False]
   - accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_fp8[fp8kv=False-attn_backend=TRTLLM-torch_compile=True]
   - accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_fp8[fp8kv=True-attn_backend=TRTLLM-torch_compile=False]
@@ -208,6 +209,9 @@ l0_h100:
   - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_fp8_block_scales[latency-torch_compile=False]
   - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_fp8_block_scales[latency-torch_compile=True]
   - 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_with_eagle3[llguidance]
+  - accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_guided_decoding_with_ngram[xgrammar]
+  - accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_guided_decoding_with_ngram[llguidance]
   - test_e2e.py::test_ptp_quickstart_multimodal[mistral-small-3.1-24b-instruct-Mistral-Small-3.1-24B-Instruct-2503-image-True]
   - test_e2e.py::test_ptp_quickstart_multimodal[mistral-small-3.1-24b-instruct-Mistral-Small-3.1-24B-Instruct-2503-mixture_text_image-True]
 - condition: