[test] add nvfp4 DeepSeek-V3-Lite-mtp tests (NVIDIA#5125)

Signed-off-by: Fanrong Li <[email protected]>

Author: lfr-0531

tests/integration/defs/.test_durations

@@ -289,8 +289,8 @@
    "accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_bfloat16[mtp_nextn=2-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False]": 220.57452515885234,
    "accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_bfloat16[mtp_nextn=2-attention_dp=False-cuda_graph=True-overlap_scheduler=False-torch_compile=False]": 202.22269394202158,
    "accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_bfloat16[mtp_nextn=2-attention_dp=False-cuda_graph=False-overlap_scheduler=True-torch_compile=False]": 165.08514453098178,
-   "accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[fp8kv=False-attention_dp=False-cuda_graph=True-overlap_scheduler=False-torch_compile=False]": 252.70569713797886,
-   "accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=True-torch_compile=False]": 85.24235329206567,
+   "accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=0-fp8kv=False-attention_dp=False-cuda_graph=True-overlap_scheduler=False-torch_compile=False]": 252.70569713797886,
+   "accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=0-fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=True-torch_compile=False]": 85.24235329206567,
    "test_e2e.py::test_ptp_quickstart_advanced[Llama3.1-8B-FP8-llama-3.1-model/Llama-3.1-8B-Instruct-FP8]": 81.43792725296225,
    "test_e2e.py::test_ptp_quickstart_advanced_eagle3[Llama-3.1-8b-Instruct-llama-3.1-model/Llama-3.1-8B-Instruct-EAGLE3-LLaMA3.1-Instruct-8B]": 109.26379436196294,
    "test_e2e.py::test_ptp_quickstart_advanced_mixed_precision": 80.88908524392173,
@@ -338,9 +338,9 @@
    "accuracy/test_cli_flow.py::TestLlama3_2_1B::test_smooth_quant_ootb_manage_weights": 216.66169160604477,
    "accuracy/test_cli_flow.py::TestTinyLlama1_1BChat::test_weight_only_manage_weights[int4]": 161.57166086137295,
    "test_unittests.py::test_unittests_v2[unittest/llmapi/test_llm.py -m \"not part0\"]": 1883.5484512336552,
-   "accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False]": 89.92349556891713,
-   "accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[attention_dp-cuda_graph-overlap_scheduler-torch_compile=False]": 175.661773331929,
-   "accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[fp8kv=False-attention_dp=True-cuda_graph=False-overlap_scheduler=False-torch_compile=False]": 90.21807348495349,
+   "accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=0-fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False]": 89.92349556891713,
+   "accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=0-fp8kv=False-attention_dp=True-cuda_graph=True-overlap_scheduler=True-torch_compile=False]": 175.661773331929,
+   "accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=0-fp8kv=False-attention_dp=True-cuda_graph=False-overlap_scheduler=False-torch_compile=False]": 90.21807348495349,
    "accuracy/test_llm_api_pytorch.py::TestLlama3_1_8B::test_nvfp4": 56.31924073398113,
    "test_e2e.py::test_ptp_quickstart_advanced[Llama3.1-8B-NVFP4-nvfp4-quantized/Meta-Llama-3.1-8B]": 56.05445321695879,
    "test_e2e.py::test_trtllm_bench_pytorch_backend_sanity[meta-llama/Llama-3.1-8B-llama-3.1-8b-False-False]": 114.17938271397725,

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

@@ -23,6 +23,13 @@ deepseek-ai/DeepSeek-V3-Lite:
   - quant_algo: NVFP4
     kv_cache_quant_algo: FP8
     accuracy: 63.71
+  - quant_algo: NVFP4
+    spec_dec_algo: MTP
+    accuracy: 63.71
+  - quant_algo: NVFP4
+    kv_cache_quant_algo: FP8
+    spec_dec_algo: MTP
+    accuracy: 63.71
   - quant_algo: FP8_BLOCK_SCALES
     accuracy: 64.74
   - quant_algo: FP8_BLOCK_SCALES

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

@@ -115,6 +115,9 @@ deepseek-ai/DeepSeek-V3-Lite:
   - accuracy: 71.40
   - quant_algo: NVFP4
     accuracy: 70.60
+  - quant_algo: NVFP4
+    spec_dec_algo: MTP
+    accuracy: 70.60
   - quant_algo: FP8_BLOCK_SCALES
     accuracy: 71.27
   - spec_dec_algo: MTP

tests/integration/defs/accuracy/test_llm_api_pytorch.py

@@ -798,8 +798,12 @@ def test_fp8_block_scales_4gpus_static_eplb(self):
                            (False, False, True, False),
                            (False, False, False, True),
                            (True, False, True, True), (True, True, True, True)])
+    @parametrize_with_ids("mtp_nextn", [0, 2])
+    @parametrize_with_ids("moe_backend", ["CUTLASS", "TRTLLM"])
     def test_nvfp4(self, fp8kv, attention_dp, cuda_graph, overlap_scheduler,
-                   torch_compile):
+                   torch_compile, mtp_nextn, moe_backend):
+        if torch_compile and mtp_nextn > 0:
+            pytest.skip("https://nvbugs/5252313")
         if torch_compile and attention_dp:
             pytest.skip("https://nvbugs/5252559")
         kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.9)
@@ -810,18 +814,24 @@ def test_nvfp4(self, fp8kv, attention_dp, cuda_graph, overlap_scheduler,
             disable_overlap_scheduler=not overlap_scheduler,
             use_cuda_graph=cuda_graph,
             torch_compile_config=torch_compile_config,
+            moe_backend=moe_backend,
         )
+        mtp_config = None
+        if mtp_nextn > 0:
+            mtp_config = MTPDecodingConfig(num_nextn_predict_layers=mtp_nextn)
+
         quant_config = QuantConfig()
         quant_config.quant_algo = QuantAlgo.NVFP4
         if fp8kv:
             quant_config.kv_cache_quant_algo = QuantAlgo.FP8
             pytorch_config["kv_cache_dtype"] = "fp8"
 
-        llm = LLM(f"{llm_models_root()}/DeepSeek-V3-Lite/nvfp4_moe_only",
+        llm = LLM(f"{llm_models_root()}/DeepSeek-V3-Lite/nvfp4_moe_only_mtp",
                   kv_cache_config=kv_cache_config,
                   **pytorch_config,
                   quant_config=quant_config,
-                  enable_attention_dp=attention_dp)
+                  enable_attention_dp=attention_dp,
+                  speculative_config=mtp_config)
 
         assert llm.args.quant_config.quant_algo == QuantAlgo.NVFP4
         if fp8kv:
@@ -850,9 +860,13 @@ def test_nvfp4(self, fp8kv, attention_dp, cuda_graph, overlap_scheduler,
     @pytest.mark.parametrize("tp_size,pp_size,ep_size", [(4, 1, 1), (4, 1, 4),
                                                          (2, 2, 1), (1, 4, 1)],
                              ids=["tp4", "ep4", "tp2pp2", "pp4"])
+    @parametrize_with_ids("mtp_nextn", [0, 2])
+    @parametrize_with_ids("moe_backend", ["CUTLASS", "TRTLLM"])
     def test_nvfp4_4gpus(self, fp8kv, attention_dp, cuda_graph,
                          overlap_scheduler, tp_size, pp_size, ep_size,
-                         torch_compile):
+                         torch_compile, mtp_nextn, moe_backend):
+        if torch_compile and mtp_nextn > 0:
+            pytest.skip("https://nvbugs/5252313")
         if torch_compile and attention_dp:
             pytest.skip("https://nvbugs/5252559")
         if torch_compile and pp_size > 1:
@@ -867,22 +881,28 @@ def test_nvfp4_4gpus(self, fp8kv, attention_dp, cuda_graph,
             disable_overlap_scheduler=not overlap_scheduler,
             use_cuda_graph=cuda_graph,
             torch_compile_config=torch_compile_config,
+            moe_backend=moe_backend,
         )
 
+        mtp_config = None
+        if mtp_nextn > 0:
+            mtp_config = MTPDecodingConfig(num_nextn_predict_layers=mtp_nextn)
+
         quant_config = QuantConfig()
         quant_config.quant_algo = QuantAlgo.NVFP4
         if fp8kv:
             quant_config.kv_cache_quant_algo = QuantAlgo.FP8
             pytorch_config["kv_cache_dtype"] = "fp8"
 
-        llm = LLM(f"{llm_models_root()}/DeepSeek-V3-Lite/nvfp4_moe_only",
+        llm = LLM(f"{llm_models_root()}/DeepSeek-V3-Lite/nvfp4_moe_only_mtp",
                   tensor_parallel_size=tp_size,
                   pipeline_parallel_size=pp_size,
                   moe_expert_parallel_size=ep_size,
                   kv_cache_config=kv_cache_config,
                   **pytorch_config,
                   quant_config=quant_config,
-                  enable_attention_dp=attention_dp)
+                  enable_attention_dp=attention_dp,
+                  speculative_config=mtp_config)
 
         assert llm.args.quant_config.quant_algo == QuantAlgo.NVFP4
         if fp8kv:

tests/integration/test_lists/qa/examples_test_list.txt

@@ -459,7 +459,8 @@ accuracy/test_llm_api_pytorch.py::TestMixtral8x7B::test_fp8_tp2
 accuracy/test_llm_api_pytorch.py::TestMixtral8x7B::test_nvfp4_tp2
 accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_bfloat16[mtp_nextn=0-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False]
 accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_fp8_block_scales[mtp_nextn=0-fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False]
-accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False]
+accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=0-fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False]
+accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=2-fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False]
 accuracy/test_llm_api_pytorch.py::TestMinitron4BBaseInstruct::test_fp8_prequantized
 accuracy/test_llm_api_pytorch.py::TestNemotronNas::test_auto_dtype_tp8
 accuracy/test_llm_api_pytorch.py::TestLlama3_3NemotronSuper49Bv1::test_fp8_prequantized_tp2

tests/integration/test_lists/qa/llm_release_rtx_pro_6000.txt

@@ -11,12 +11,16 @@ accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_bfloat16[mtp_nextn=2-
 accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_bfloat16[mtp_nextn=2-attention_dp=False-cuda_graph=True-overlap_scheduler=False-torch_compile=False]
 accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_bfloat16[mtp_nextn=2-attention_dp=False-cuda_graph=False-overlap_scheduler=True-torch_compile=False]
 accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_bfloat16[mtp_nextn=2-attention_dp=True-cuda_graph=True-overlap_scheduler=True-torch_compile=False]
-accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False]
-accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[fp8kv=True-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False]
-accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[fp8kv=False-attention_dp=True-cuda_graph=False-overlap_scheduler=False-torch_compile=False]
-accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[fp8kv=False-attention_dp=False-cuda_graph=True-overlap_scheduler=False-torch_compile=False]
-accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=True-torch_compile=False]
-accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[fp8kv=True-attention_dp=True-cuda_graph=True-overlap_scheduler=True-torch_compile=False]
+accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=0-fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False]
+accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=0-fp8kv=True-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False]
+accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=0-fp8kv=False-attention_dp=True-cuda_graph=False-overlap_scheduler=False-torch_compile=False]
+accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=0-fp8kv=False-attention_dp=False-cuda_graph=True-overlap_scheduler=False-torch_compile=False]
+accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=0-fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=True-torch_compile=False]
+accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=0-fp8kv=True-attention_dp=True-cuda_graph=True-overlap_scheduler=True-torch_compile=False]
+accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=2-fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False]
+accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=2-fp8kv=True-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False]
+accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=2-fp8kv=False-attention_dp=False-cuda_graph=True-overlap_scheduler=False-torch_compile=False]
+accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=2-fp8kv=True-attention_dp=False-cuda_graph=True-overlap_scheduler=True-torch_compile=False]
 accuracy/test_llm_api_pytorch.py::TestLlama3_1_8B::test_nvfp4
 test_e2e.py::test_ptp_quickstart_advanced_mixed_precision
 test_e2e.py::test_ptp_quickstart_advanced[Llama3.1-8B-BF16-llama-3.1-model/Meta-Llama-3.1-8B]

tests/integration/test_lists/qa/llm_sanity_test.txt

@@ -128,7 +128,8 @@ accuracy/test_llm_api_pytorch.py::TestNemotronUltra::test_fp8_prequantized[tp8-c
 accuracy/test_llm_api_pytorch.py::TestNemotronH::test_reasoning_fp8_prequantized
 accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_bfloat16[mtp_nextn=0-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False]
 accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_fp8_block_scales[mtp_nextn=0-fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False]
-accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False]
+accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=0-fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False]
+accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=2-fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False]
 accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_fp8_block_scales[latency]
 accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_fp8_block_scales[latency]
 accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[latency_moe_cutlass]

tests/integration/test_lists/test-db/l0_b200.yml

@@ -29,15 +29,23 @@ l0_b200:
   - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_bfloat16[mtp_nextn=2-attention_dp=False-cuda_graph=True-overlap_scheduler=False-torch_compile=False]
   - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_bfloat16[mtp_nextn=2-attention_dp=False-cuda_graph=False-overlap_scheduler=True-torch_compile=False]
   - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_bfloat16[mtp_nextn=2-attention_dp=True-cuda_graph=True-overlap_scheduler=True-torch_compile=False]
-  - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False]
-  - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=True]
-  - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[fp8kv=True-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False]
-  - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[fp8kv=True-attention_dp=False-cuda_graph=True-overlap_scheduler=True-torch_compile=True]
-  - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[fp8kv=False-attention_dp=True-cuda_graph=False-overlap_scheduler=False-torch_compile=False]
-  - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[fp8kv=False-attention_dp=True-cuda_graph=False-overlap_scheduler=False-torch_compile=True]
-  - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[fp8kv=False-attention_dp=False-cuda_graph=True-overlap_scheduler=False-torch_compile=False]
-  - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=True-torch_compile=False]
-  - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[fp8kv=True-attention_dp=True-cuda_graph=True-overlap_scheduler=True-torch_compile=False]
+  - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=0-fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False]
+  - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=0-fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=True]
+  - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=0-fp8kv=True-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False]
+  - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=0-fp8kv=True-attention_dp=False-cuda_graph=True-overlap_scheduler=True-torch_compile=True]
+  - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=0-fp8kv=False-attention_dp=True-cuda_graph=False-overlap_scheduler=False-torch_compile=False]
+  - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=0-fp8kv=False-attention_dp=True-cuda_graph=False-overlap_scheduler=False-torch_compile=True]
+  - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=0-fp8kv=False-attention_dp=False-cuda_graph=True-overlap_scheduler=False-torch_compile=False]
+  - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=0-fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=True-torch_compile=False]
+  - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=0-fp8kv=True-attention_dp=True-cuda_graph=True-overlap_scheduler=True-torch_compile=False]
+  - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=TRTLLM-mtp_nextn=0-fp8kv=False-attention_dp=False-cuda_graph=True-overlap_scheduler=False-torch_compile=False]
+  - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=TRTLLM-mtp_nextn=0-fp8kv=True-attention_dp=False-cuda_graph=True-overlap_scheduler=True-torch_compile=False]
+  - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=2-fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False]
+  - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=2-fp8kv=True-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False]
+  - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=2-fp8kv=False-attention_dp=False-cuda_graph=True-overlap_scheduler=False-torch_compile=False]
+  - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=2-fp8kv=True-attention_dp=False-cuda_graph=True-overlap_scheduler=True-torch_compile=False]
+  - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=TRTLLM-mtp_nextn=2-fp8kv=False-attention_dp=False-cuda_graph=True-overlap_scheduler=False-torch_compile=False]
+  - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=TRTLLM-mtp_nextn=2-fp8kv=True-attention_dp=False-cuda_graph=True-overlap_scheduler=True-torch_compile=False]
   - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_no_kv_cache_reuse[quant_dtype=none-mtp_nextn=2-fp8kv=False-attention_dp=True-cuda_graph=True-overlap_scheduler=True]
   - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_no_kv_cache_reuse[quant_dtype=nvfp4-mtp_nextn=0-fp8kv=True-attention_dp=True-cuda_graph=True-overlap_scheduler=True]
   - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[latency_moe_cutlass]