add qwen3-4b accuracy test case

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

Author: crazydemo

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

@@ -106,6 +106,9 @@ deepseek-ai/DeepSeek-V3.2-Exp:
   - quant_algo: NVFP4
     spec_dec_algo: MTP
     accuracy: 95.6
+Qwen3/Qwen3-4B:
+  - spec_dec_algo: Eagle
+    accuracy: 85.823
 Qwen3/Qwen3-8B:
   - accuracy: 87.1114
   - spec_dec_algo: Eagle

tests/integration/defs/accuracy/test_llm_api_pytorch.py

@@ -3294,6 +3294,37 @@ def test_auto_dtype(self):
                           extra_evaluator_kwargs=self.EXTRA_EVALUATOR_KWARGS)
 
 
+class TestQwen3_4B(LlmapiAccuracyTestHarness):
+    MODEL_NAME = "Qwen3/Qwen3-4B"
+
+    def test_eagle3(self):
+        "RCCA: https://nvbugspro.nvidia.com/bug/5698434"
+        pytorch_config = dict(
+            disable_overlap_scheduler=True,
+            cuda_graph_config=CudaGraphConfig(),
+        )
+        kv_cache_config = KvCacheConfig(
+            enable_block_reuse=False,
+            free_gpu_memory_fraction=0.6,
+        )
+
+        eagle_model_dir = f"{llm_models_root()}/Qwen3/Qwen3-4B_eagle3/"
+        target_model_dir = f"{llm_models_root()}/Qwen3/Qwen3-4B"
+
+        draft_len = 3
+        spec_config = EagleDecodingConfig(max_draft_len=draft_len,
+                                          speculative_model_dir=eagle_model_dir)
+
+        llm = LLM(model=target_model_dir,
+                  **pytorch_config,
+                  kv_cache_config=kv_cache_config,
+                  speculative_config=spec_config)
+
+        with llm:
+            task = GSM8K(self.MODEL_NAME)
+            task.evaluate(llm)
+
+
 class TestQwen3_8B(LlmapiAccuracyTestHarness):
     MODEL_NAME = "Qwen3/Qwen3-8B"
 

tests/integration/test_lists/qa/llm_function_core.txt

@@ -523,6 +523,7 @@ accuracy/test_llm_api_pytorch.py::TestGLM4_6::test_nvfp4_multi_gpus[throughput]
 accuracy/test_llm_api_pytorch.py::TestGLM4_6::test_nvfp4_multi_gpus[throughput_trtllm]
 accuracy/test_llm_api_pytorch.py::TestGLM4_6::test_nvfp4_2_model_mtp[2model]
 accuracy/test_llm_api_pytorch.py::TestGLM4_6::test_nvfp4_2_model_mtp[2model_trtllm]
+accuracy/test_llm_api_pytorch.py::TestQwen3_4B::test_eagle3
 accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_fp8_block_scales[latency]
 accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_bf16[multi_gpus_no_cache]
 accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_fp8_block_scales[latency-torch_compile=False]