[https://nvbugs/5437405][fix] cherry-pick PR 7000 (qwen3 235b eagle3 ci) (#7702)

Signed-off-by: bhsueh <[email protected]>

Author: byshiue

tests/integration/defs/accuracy/test_llm_api_pytorch.py

@@ -2295,11 +2295,12 @@ def test_fp8(self, tp_size, pp_size, ep_size, attention_dp, cuda_graph,
         [
             (8, 1, 8, True, True, True, "CUTLASS", False),
             (8, 1, 8, True, True, True, "TRTLLM", False),
-            (8, 1, 8, False, False, False, "TRTLLM", True),
+            (8, 1, 8, True, True, True, "TRTLLM", True),
         ],
         ids=[
-            "latency_moe_cutlass", "latency_moe_trtllm",
-            "latency_moe_trtllm_eagle3"
+            "latency_moe_cutlass",
+            "latency_moe_trtllm",
+            "latency_moe_trtllm_eagle3",
         ],
     )
     def test_nvfp4(self, tp_size, pp_size, ep_size, attention_dp, cuda_graph,
@@ -2339,6 +2340,50 @@ def test_nvfp4(self, tp_size, pp_size, ep_size, attention_dp, cuda_graph,
             task = GSM8K(self.MODEL_NAME)
             task.evaluate(llm)
 
+    @skip_pre_blackwell
+    @pytest.mark.skip_less_mpi_world_size(4)
+    @pytest.mark.parametrize(
+        "tp_size,pp_size,ep_size,attention_dp,cuda_graph,overlap_scheduler,moe_backend,eagle3",
+        [
+            (4, 1, 4, False, False, False, "TRTLLM",
+             True),  # TP8 has bug when we use TRTLLM moe backend and eagle3
+        ],
+        ids=[
+            "latency_moe_trtllm_eagle3",
+        ],
+    )
+    def test_nvfp4_4gpus(self, tp_size, pp_size, ep_size, attention_dp,
+                         cuda_graph, overlap_scheduler, moe_backend, eagle3):
+
+        pytorch_config = dict(
+            disable_overlap_scheduler=not overlap_scheduler,
+            cuda_graph_config=CudaGraphConfig() if cuda_graph else None,
+            moe_config=MoeConfig(backend=moe_backend))
+
+        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.4,
+                                        enable_block_reuse=not eagle3)
+        spec_config = None
+        if eagle3:
+            spec_config = EagleDecodingConfig(
+                max_draft_len=2,
+                speculative_model_dir=
+                f"{llm_models_root()}/Qwen3/qwen3-235B-eagle3/",
+                eagle3_one_model=True)
+        with LLM(
+                f"{llm_models_root()}/Qwen3/saved_models_Qwen3-235B-A22B_nvfp4_hf",
+                tensor_parallel_size=tp_size,
+                pipeline_parallel_size=pp_size,
+                moe_expert_parallel_size=ep_size,
+                **pytorch_config,
+                enable_attention_dp=attention_dp,
+                kv_cache_config=kv_cache_config,
+                speculative_config=spec_config) as llm:
+
+            task = MMLU(self.MODEL_NAME)
+            task.evaluate(llm)
+            task = GSM8K(self.MODEL_NAME)
+            task.evaluate(llm)
+
 
 class TestPhi4MiniInstruct(LlmapiAccuracyTestHarness):
     MODEL_NAME = "microsoft/Phi-4-mini-instruct"

tests/integration/test_lists/qa/llm_function_full.txt

@@ -537,7 +537,7 @@ 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::TestQwen3_235B_A22B::test_nvfp4_4gpus[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

tests/integration/test_lists/qa/llm_function_sanity.txt

@@ -100,7 +100,7 @@ accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[latency_moe_cutl
 accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[latency_moe_cutlass-torch_compile=True]
 accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[latency_moe_trtllm-torch_compile=False]
 accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[latency_moe_trtllm-torch_compile=True]
-accuracy/test_llm_api_pytorch.py::TestQwen3_235B_A22B::test_nvfp4[latency_moe_trtllm_eagle3]
+accuracy/test_llm_api_pytorch.py::TestQwen3_235B_A22B::test_nvfp4_4gpus[latency_moe_trtllm_eagle3]
 accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_fp8_block_scales[latency]
 accuracy/test_llm_api_pytorch.py::TestPhi4MiniInstruct::test_auto_dtype
 accuracy/test_disaggregated_serving.py::TestQwen3_8B::test_nixl_backend

tests/integration/test_lists/test-db/l0_gb200.yml

@@ -66,3 +66,4 @@ l0_gb200:
   - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4_4gpus[moe_backend=CUTLASS-mtp_nextn=2-pp4-fp8kv=True-attention_dp=True-cuda_graph=True-overlap_scheduler=True-torch_compile=False]
   - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4_4gpus[moe_backend=TRTLLM-mtp_nextn=2-tp4-fp8kv=True-attention_dp=True-cuda_graph=True-overlap_scheduler=True-torch_compile=False]
   - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4_4gpus[moe_backend=TRTLLM-mtp_nextn=2-ep4-fp8kv=True-attention_dp=True-cuda_graph=True-overlap_scheduler=True-torch_compile=False]
+  - accuracy/test_llm_api_pytorch.py::TestQwen3_235B_A22B::test_nvfp4_4gpus[latency_moe_trtllm_eagle3] TIMEOUT (90)

tests/integration/test_lists/test-db/l0_gb200_multi_nodes.yml

@@ -16,7 +16,6 @@ l0_gb200_multi_nodes:
   - accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_nvfp4_multi_gpus[latency] TIMEOUT (180)
   - accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_nvfp4_multi_gpus[throughput_tp8] TIMEOUT (180)
   - accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_nvfp4_multi_gpus[latency_trtllmgen] TIMEOUT (180)
-  - accuracy/test_llm_api_pytorch.py::TestQwen3_235B_A22B::test_nvfp4[latency_moe_cutlass] TIMEOUT (180)
-  - accuracy/test_llm_api_pytorch.py::TestQwen3_235B_A22B::test_nvfp4[latency_moe_trtllm] TIMEOUT (180)
-  - accuracy/test_llm_api_pytorch.py::TestQwen3_235B_A22B::test_nvfp4[latency_moe_trtllm_eagle3] TIMEOUT (180)
+  - accuracy/test_llm_api_pytorch.py::TestQwen3_235B_A22B::test_nvfp4[latency_moe_cutlass] TIMEOUT (90)
+  - accuracy/test_llm_api_pytorch.py::TestQwen3_235B_A22B::test_nvfp4[latency_moe_trtllm] TIMEOUT (90)
   - accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_bf16[multi_gpus_no_cache] TIMEOUT (180)