[None][infra] add eagle3 one model accuracy tests (NVIDIA#6264)

jhaotingc · lancelly · commit e82d393d26b8 · 2025-08-06T03:01:59.000Z
Signed-off-by: Jhao-Ting Chen &lt;jhaotingc@nvidia.com&gt;
Signed-off-by: Lanyu Liao &lt;lancelly@users.noreply.github.com&gt;
diff --git a/tests/integration/defs/accuracy/references/gsm8k.yaml b/tests/integration/defs/accuracy/references/gsm8k.yaml
@@ -2,6 +2,8 @@ meta-llama/Llama-3.1-8B-Instruct:
   - accuracy: 74.20
   - spec_dec_algo: NGram
     accuracy: 74.20
+  - spec_dec_algo: Eagle
+    accuracy: 74.20
   - quant_algo: FP8
     accuracy: 74.30
   - quant_algo: FP8
diff --git a/tests/integration/defs/accuracy/test_disaggregated_serving.py b/tests/integration/defs/accuracy/test_disaggregated_serving.py
@@ -336,36 +336,44 @@ def test_ngram(self):
             task = GSM8K(self.MODEL_NAME)
             task.evaluate(llm)
 
-    @pytest.mark.parametrize("overlap_scheduler", [False])
-    def test_eagle3(self, overlap_scheduler):
+    @parametrize_with_ids("overlap_scheduler", [True, False])
+    @parametrize_with_ids("eagle3_one_model", [True, False])
+    def test_eagle3(self, overlap_scheduler, eagle3_one_model):
         speculative_decoding_config = {
             "decoding_type": "Eagle",
             "max_draft_len": 4,
             "speculative_model_dir":
             f"{llm_models_root()}/EAGLE3-LLaMA3.1-Instruct-8B",
-            "eagle3_one_model": False
-        }
-        kv_cache_config = {
-            "free_gpu_memory_fraction": 0.5,
-            "enable_block_reuse": False
+            "eagle3_one_model": eagle3_one_model
         }
         ctx_server_config = {
-            "disable_overlap_scheduler": True,
+            "disable_overlap_scheduler":
+            True,  # BS=1 does not need overlap scheduling
             "speculative_config": speculative_decoding_config,
-            "kv_cache_config": kv_cache_config,
+            "kv_cache_config": {
+                "free_gpu_memory_fraction": 0.5,
+                "enable_block_reuse": True  # reuse on context requests
+            },
             "max_num_tokens": 13393 * 2,
+            "max_batch_size": 1,
             "cache_transceiver_config": {
                 "backend": "default"
-            }
+            },
+            "cuda_graph_config": None,
         }
         gen_server_config = {
             "disable_overlap_scheduler": not overlap_scheduler,
             "speculative_config": speculative_decoding_config,
-            "kv_cache_config": kv_cache_config,
+            "kv_cache_config": {
+                "free_gpu_memory_fraction": 0.5,
+                "enable_block_reuse": False
+            },
             "max_num_tokens": 13393 * 2,
+            "max_batch_size": 16,
             "cache_transceiver_config": {
                 "backend": "default"
-            }
+            },
+            "cuda_graph_config": None,
         }
         disaggregated_server_config = {
             "hostname": "localhost",
diff --git a/tests/integration/defs/accuracy/test_llm_api_pytorch.py b/tests/integration/defs/accuracy/test_llm_api_pytorch.py
@@ -238,19 +238,27 @@ def test_fp8_beam_search(self):
                           extra_acc_spec="beam_width=4")
 
     @skip_pre_hopper
-    def test_eagle3(self):
+    @parametrize_with_ids("overlap_scheduler", [True, False])
+    @parametrize_with_ids("eagle3_one_model", [True, False])
+    def test_eagle3(self, overlap_scheduler, eagle3_one_model):
         pytorch_config = dict(
-            disable_overlap_scheduler=True,
-            cuda_graph_config=CudaGraphConfig(batch_sizes=[1]),
+            max_batch_size=
+            1,  # add max_batch_size to avoid error in overlap scheduler
+            disable_overlap_scheduler=not overlap_scheduler,
+            cuda_graph_config=CudaGraphConfig(max_batch_size=1,
+                                              enable_padding=True),
         )
-        kv_cache_config = KvCacheConfig(enable_block_reuse=False)
+        kv_cache_config = KvCacheConfig(
+            enable_block_reuse=True
+        )  # both one-model and two-model supports this feature
 
         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)
+                                          speculative_model_dir=eagle_model_dir,
+                                          eagle3_one_model=eagle3_one_model)
 
         with LLM(model=target_model_dir,
                  **pytorch_config,
@@ -259,6 +267,8 @@ def test_eagle3(self):
                  build_config=None) as llm:
             task = MMLU(self.MODEL_NAME)
             task.evaluate(llm)
+            task = GSM8K(self.MODEL_NAME)
+            task.evaluate(llm)
 
     @skip_pre_hopper
     def test_ngram(self):
diff --git a/tests/integration/test_lists/qa/examples_test_list.txt b/tests/integration/test_lists/qa/examples_test_list.txt
@@ -441,7 +441,8 @@ 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_eagle3[eagle3_one_model=True-overlap_scheduler=True]
+accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_eagle3[eagle3_one_model=False-overlap_scheduler=False]
 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]
diff --git a/tests/integration/test_lists/qa/llm_release_digits_func.txt b/tests/integration/test_lists/qa/llm_release_digits_func.txt
@@ -19,5 +19,6 @@ test_e2e.py::test_ptp_quickstart_multimodal[NVILA-15B-FP16-NVILA-15B-image-False
 test_e2e.py::test_ptp_quickstart_multimodal[NVILA-15B-FP16-NVILA-15B-video-False]
 
 accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_fp8_llm_sampler
-accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_eagle3
+accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_eagle3[eagle3_one_model=False-overlap_scheduler=False]
+accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_eagle3[eagle3_one_model=True-overlap_scheduler=True]
 accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_ngram
diff --git a/tests/integration/test_lists/qa/llm_sanity_test.txt b/tests/integration/test_lists/qa/llm_sanity_test.txt
@@ -21,7 +21,8 @@ accuracy/test_llm_api_pytorch.py::TestGemma3_1BInstruct::test_auto_dtype
 accuracy/test_llm_api_pytorch.py::TestKanana_Instruct::test_auto_dtype
 accuracy/test_llm_api_pytorch.py::TestLlama3_1_8B::test_nvfp4
 accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_chunked_prefill[attn_backend=FLASHINFER]
-accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_eagle3
+accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_eagle3[eagle3_one_model=False-overlap_scheduler=False]
+accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_eagle3[eagle3_one_model=True-overlap_scheduler=True]
 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_guided_decoding_4gpus[llguidance]
diff --git a/tests/integration/test_lists/test-db/l0_dgx_h100.yml b/tests/integration/test_lists/test-db/l0_dgx_h100.yml
@@ -38,9 +38,10 @@ l0_dgx_h100:
   - accuracy/test_disaggregated_serving.py::TestGemma3_1BInstruct::test_auto_dtype[False]
   - accuracy/test_disaggregated_serving.py::TestGemma3_1BInstruct::test_auto_dtype[True]
   - accuracy/test_disaggregated_serving.py::TestLlama3_1_8BInstruct::test_ngram
-  - accuracy/test_disaggregated_serving.py::TestLlama3_1_8BInstruct::test_eagle3[False]
   - accuracy/test_disaggregated_serving.py::TestQwen3_8B::test_auto_dtype[False]
   - accuracy/test_disaggregated_serving.py::TestQwen3_8B::test_auto_dtype[True]
+  - accuracy/test_disaggregated_serving.py::TestLlama3_1_8BInstruct::test_eagle3[eagle3_one_model=False-overlap_scheduler=False]
+  - accuracy/test_disaggregated_serving.py::TestLlama3_1_8BInstruct::test_eagle3[eagle3_one_model=True-overlap_scheduler=True]
   - accuracy/test_disaggregated_serving.py::TestLlama3_1_8BInstruct::test_tp_pp_symmetric[GSM8K-tp1pp2]
   - accuracy/test_disaggregated_serving.py::TestLlama3_1_8BInstruct::test_tp_pp_symmetric[MMLU-tp1pp2]
   - accuracy/test_disaggregated_serving.py::TestLlama3_1_8BInstruct::test_tp_pp_symmetric[GSM8K-tp2pp1]
diff --git a/tests/integration/test_lists/test-db/l0_h100.yml b/tests/integration/test_lists/test-db/l0_h100.yml
@@ -34,6 +34,8 @@ l0_h100:
   - 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]
   - accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_fp8[fp8kv=True-attn_backend=TRTLLM-torch_compile=True]
+  - accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_eagle3[eagle3_one_model=False-overlap_scheduler=False]
+  - accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_eagle3[eagle3_one_model=True-overlap_scheduler=True]
   - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_fp8_block_scales[mtp=disable-fp8kv=True-attention_dp=False-cuda_graph=True-overlap_scheduler=True-torch_compile=True]
   - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_fp8_block_scales[mtp=eagle-fp8kv=True-attention_dp=True-cuda_graph=True-overlap_scheduler=True-torch_compile=False]
   - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_fp8_block_scales[mtp=vanilla-fp8kv=True-attention_dp=False-cuda_graph=True-overlap_scheduler=True-torch_compile=True]
diff --git a/tests/integration/test_lists/waives.txt b/tests/integration/test_lists/waives.txt
@@ -251,7 +251,6 @@ test_e2e.py::test_ptp_quickstart_multimodal[gemma-3-27b-it-gemma/gemma-3-27b-it-
 test_e2e.py::test_ptp_quickstart_multimodal[gemma-3-27b-it-gemma/gemma-3-27b-it-image-False] SKIP (https://nvbugs/5401114)
 examples/test_recurrentgemma.py::test_llm_recurrentgemma_1gpu[use_cpp_session-recurrentgemma-2b-use_paged_cache-int4_awq-float16-enable_attn_plugin-enable_gemm_plugin] SKIP (https://nvbugs/5401233)
 examples/test_recurrentgemma.py::test_llm_recurrentgemma_2gpu[recurrentgemma-2b] SKIP (https://nvbugs/5401233)
-accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_eagle3 SKIP (https://nvbugs/5409414)
 accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_fp8_beam_search SKIP (https://nvbugs/5409415)
 accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_ngram SKIP (https://nvbugs/5409414)
 test_e2e.py::test_openai_multi_chat_example SKIP (https://nvbugs/5409416)
