[None][test] Add one-model and overlap-scheduling to eagle tests for GPTOSS (NVIDIA#9312)

Signed-off-by: Dongfeng Yu <[email protected]>

Author: dongfengy

tests/integration/defs/accuracy/test_llm_api_pytorch.py

@@ -3956,16 +3956,24 @@ def test_w4_chunked_prefill(self, kv_cache_dtype, moe_backend, mocker):
                           extra_evaluator_kwargs=extra_evaluator_kwargs)
 
     @pytest.mark.skip_less_device(4)
+    @pytest.mark.parametrize("overlap_scheduler", [True, False],
+                             ids=["overlap_scheduler", "no_overlap_scheduler"])
+    @pytest.mark.parametrize("one_model", [True, False],
+                             ids=["one_model", "two_model"])
     @pytest.mark.parametrize(
         "moe_backend",
         ["CUTLASS",
          pytest.param("TRTLLM", marks=skip_pre_blackwell), "TRITON"],
         ids=["cutlass", "trtllm", "triton"])
-    def test_eagle3(self, moe_backend, mocker):
+    def test_eagle3(self, moe_backend, one_model, overlap_scheduler, mocker):
         if moe_backend == "TRITON":
             if not IS_TRITON_KERNELS_AVAILABLE:
                 pytest.skip("Triton kernels are not available")
 
+        if get_sm_version() == 90 and moe_backend == "CUTLASS":
+            pytest.skip(
+                "https://nvbugs/5636916: Remaining Hopper Eagle Accuracy Issue")
+
         MAX_OUTPUT_LEN = 128179
         MAX_INPUT_LEN = 32768
 
@@ -3977,16 +3985,16 @@ def test_eagle3(self, moe_backend, mocker):
         mocker.patch.object(GPQADiamond, "MAX_INPUT_LEN", MAX_INPUT_LEN)
 
         # https://nvbugs/5590408: 2-Model overlap scheduling has accuracy issue
-        pytorch_config = dict(disable_overlap_scheduler=True,
+        pytorch_config = dict(disable_overlap_scheduler=not overlap_scheduler,
                               cuda_graph_config=CudaGraphConfig())
-        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.6,
+        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.4,
                                         dtype="auto")
 
         eagle_model_dir = f"{llm_models_root()}/gpt_oss/gpt-oss-120b-Eagle3"
         draft_len = 3
         spec_config = EagleDecodingConfig(max_draft_len=draft_len,
                                           speculative_model_dir=eagle_model_dir,
-                                          eagle3_one_model=False)
+                                          eagle3_one_model=one_model)
 
         max_seq_len = MAX_INPUT_LEN + MAX_OUTPUT_LEN
         llm = LLM(self.MODEL_PATH,

tests/integration/test_lists/qa/llm_function_core.txt

@@ -564,9 +564,18 @@ accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_chunked_prefill[cutlass-au
 accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_chunked_prefill[trtllm-auto]
 accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_chunked_prefill[triton-auto]
 accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_chunked_prefill[trtllm-fp8]
-accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[trtllm]
-accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[cutlass]
-accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[triton]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[cutlass-one_model-overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[cutlass-one_model-no_overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[cutlass-two_model-overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[cutlass-two_model-no_overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[trtllm-one_model-overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[trtllm-one_model-no_overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[trtllm-two_model-overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[trtllm-two_model-no_overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[triton-one_model-overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[triton-one_model-no_overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[triton-two_model-overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[triton-two_model-no_overlap_scheduler]
 accuracy/test_disaggregated_serving.py::TestLlama3_1_8BInstruct::test_auto_dtype[False-False-False]
 accuracy/test_disaggregated_serving.py::TestLlama3_1_8BInstruct::test_auto_dtype[True-True-True]
 accuracy/test_disaggregated_serving.py::TestLlama3_1_8BInstruct::test_ngram

tests/integration/test_lists/qa/llm_function_core_sanity.txt

@@ -101,9 +101,18 @@ accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_chunked_prefill[cutlass-au
 accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_chunked_prefill[trtllm-auto]
 accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_chunked_prefill[triton-auto]
 accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_chunked_prefill[trtllm-fp8]
-accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[trtllm]
-accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[cutlass]
-accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[triton]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[cutlass-one_model-overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[cutlass-one_model-no_overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[cutlass-two_model-overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[cutlass-two_model-no_overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[trtllm-one_model-overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[trtllm-one_model-no_overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[trtllm-two_model-overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[trtllm-two_model-no_overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[triton-one_model-overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[triton-one_model-no_overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[triton-two_model-overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[triton-two_model-no_overlap_scheduler]
 accuracy/test_llm_api_pytorch.py::TestMistralSmall24B::test_auto_dtype
 accuracy/test_llm_api_pytorch.py::TestKanana_Instruct::test_auto_dtype
 accuracy/test_llm_api_pytorch.py::TestKimiK2::test_fp8_blockscale[latency]

tests/integration/test_lists/qa/llm_function_nim.txt

@@ -342,9 +342,18 @@ accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_chunked_prefill[cutlass-au
 accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_chunked_prefill[trtllm-auto]
 accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_chunked_prefill[triton-auto]
 accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_chunked_prefill[trtllm-fp8]
-accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[trtllm]
-accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[cutlass]
-accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[triton]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[cutlass-one_model-overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[cutlass-one_model-no_overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[cutlass-two_model-overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[cutlass-two_model-no_overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[trtllm-one_model-overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[trtllm-one_model-no_overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[trtllm-two_model-overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[trtllm-two_model-no_overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[triton-one_model-overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[triton-one_model-no_overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[triton-two_model-overlap_scheduler]
+accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[triton-two_model-no_overlap_scheduler]
 accuracy/test_llm_api_pytorch.py::TestQwen3_235B_A22B::test_fp8[throughput_latency]
 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]

tests/integration/test_lists/test-db/l0_dgx_b200.yml

@@ -50,8 +50,8 @@ l0_dgx_b200:
   - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_4gpus[dp4-trtllm-fp8]
   - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4a16[dp4-auto]
   - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4a16[dp4-fp8]
-  - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[trtllm]
-  - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[cutlass]
+  - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[trtllm-two_model-overlap_scheduler]
+  - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[cutlass-one_model-overlap_scheduler]
   - disaggregated/test_disaggregated.py::test_disaggregated_benchmark_on_diff_backends[DeepSeek-V3-Lite-bf16]
   - disaggregated/test_disaggregated.py::test_disaggregated_benchmark_on_diff_backends[llama-3.1-8b-instruct-hf-fp8]
   - disaggregated/test_disaggregated.py::test_disaggregated_deepseek_v3_lite_fp8_nixl[DeepSeek-V3-Lite-fp8]
@@ -196,6 +196,12 @@ l0_dgx_b200:
   - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_4gpus[ep4-trtllm-fp8]
   - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_4gpus[dp4-cutlass-auto]
   - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_4gpus[dp4-triton-auto]
-  - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[trtllm]
-  - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[cutlass]
+  - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[trtllm-one_model-overlap_scheduler]
+  - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[trtllm-one_model-no_overlap_scheduler]
+  - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[trtllm-two_model-overlap_scheduler]
+  - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[trtllm-two_model-no_overlap_scheduler]
+  - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[cutlass-one_model-overlap_scheduler]
+  - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[cutlass-one_model-no_overlap_scheduler]
+  - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[cutlass-two_model-overlap_scheduler]
+  - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[cutlass-two_model-no_overlap_scheduler]
   - disaggregated/test_disaggregated.py::test_disaggregated_benchmark_on_diff_backends[llama-v3-8b-hf]

tests/integration/test_lists/test-db/l0_dgx_h100.yml

@@ -185,8 +185,14 @@ l0_dgx_h100:
   - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_4gpus[dp4-cutlass-auto]
   - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_4gpus[dp4-triton-auto]
   - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4a16[dp4-auto]
-  - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[cutlass]
-  - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[triton]
+  - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[cutlass-one_model-overlap_scheduler]
+  - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[cutlass-one_model-no_overlap_scheduler]
+  - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[cutlass-two_model-overlap_scheduler]
+  - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[cutlass-two_model-no_overlap_scheduler]
+  - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[triton-one_model-overlap_scheduler]
+  - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[triton-one_model-no_overlap_scheduler]
+  - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[triton-two_model-overlap_scheduler]
+  - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[triton-two_model-no_overlap_scheduler]
 - condition:
     ranges:
       system_gpu_count:

tests/integration/test_lists/waives.txt

@@ -346,7 +346,6 @@ accuracy/test_llm_api_pytorch.py::TestQwen3_235B_A22B::test_fp8[throughput_laten
 accuracy/test_llm_api_pytorch.py::TestQwen3_235B_A22B::test_nvfp4[latency_moe_trtllm_attention_dp] SKIP (https://nvbugs/5637220)
 llmapi/test_llm_examples.py::test_llmapi_example_multilora SKIP (https://nvbugs/5636857)
 unittest/_torch/modules/test_mla_helix.py::test_mla_helix_distributed SKIP (https://nvbugspro.nvidia.com/bug/5637012)
-accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3[cutlass] SKIP (https://nvbugs/5636916)
 accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_bfloat16_4gpus[tp4-attn_backend=TRTLLM-torch_compile=False] SKIP (https://nvbugs/5616182)
 examples/test_phi.py::test_llm_phi_quantization_1gpu[Phi-3-small-128k-instruct-fp8-bfloat16] SKIP (https://nvbugs/5465143)
 examples/test_multimodal.py::test_llm_multimodal_general[Mistral-Small-3.1-24B-Instruct-2503-pp:1-tp:1-bfloat16-bs:8-cpp_e2e:False-nb:1] SKIP (https://nvbugs/5644684)