[TRTLLM-6991][chore] add DeepSeek-R1 FP8 accuracy tests on Blackwell (#6710)

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

Author: lfr-0531

jenkins/L0_Test.groovy

@@ -1894,12 +1894,13 @@ def launchTestJobs(pipeline, testFilter, dockerNode=null)
 
     multiNodesSBSAConfigs = [
         // Each stage test 1 testcase with 8 GPUs and 2 nodes.
-        "GB200-8_GPUs-2_Nodes-PyTorch-Post-Merge-1": ["gb200-multi-node", "l0_gb200_multi_nodes", 1, 6, 8, 2],
-        "GB200-8_GPUs-2_Nodes-PyTorch-Post-Merge-2": ["gb200-multi-node", "l0_gb200_multi_nodes", 2, 6, 8, 2],
-        "GB200-8_GPUs-2_Nodes-PyTorch-Post-Merge-3": ["gb200-multi-node", "l0_gb200_multi_nodes", 3, 6, 8, 2],
-        "GB200-8_GPUs-2_Nodes-PyTorch-Post-Merge-4": ["gb200-multi-node", "l0_gb200_multi_nodes", 4, 6, 8, 2],
-        "GB200-8_GPUs-2_Nodes-PyTorch-Post-Merge-5": ["gb200-multi-node", "l0_gb200_multi_nodes", 5, 6, 8, 2],
-        "GB200-8_GPUs-2_Nodes-PyTorch-Post-Merge-6": ["gb200-multi-node", "l0_gb200_multi_nodes", 6, 6, 8, 2],
+        "GB200-8_GPUs-2_Nodes-PyTorch-Post-Merge-1": ["gb200-multi-node", "l0_gb200_multi_nodes", 1, 7, 8, 2],
+        "GB200-8_GPUs-2_Nodes-PyTorch-Post-Merge-2": ["gb200-multi-node", "l0_gb200_multi_nodes", 2, 7, 8, 2],
+        "GB200-8_GPUs-2_Nodes-PyTorch-Post-Merge-3": ["gb200-multi-node", "l0_gb200_multi_nodes", 3, 7, 8, 2],
+        "GB200-8_GPUs-2_Nodes-PyTorch-Post-Merge-4": ["gb200-multi-node", "l0_gb200_multi_nodes", 4, 7, 8, 2],
+        "GB200-8_GPUs-2_Nodes-PyTorch-Post-Merge-5": ["gb200-multi-node", "l0_gb200_multi_nodes", 5, 7, 8, 2],
+        "GB200-8_GPUs-2_Nodes-PyTorch-Post-Merge-6": ["gb200-multi-node", "l0_gb200_multi_nodes", 6, 7, 8, 2],
+        "GB200-8_GPUs-2_Nodes-PyTorch-Post-Merge-7": ["gb200-multi-node", "l0_gb200_multi_nodes", 7, 7, 8, 2],
     ]
     fullSet += multiNodesSBSAConfigs.keySet()
 

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

@@ -74,6 +74,9 @@ deepseek-ai/DeepSeek-R1:
   - quant_algo: FP8_BLOCK_SCALES
     spec_dec_algo: MTP
     accuracy: 95.413
+  - quant_algo: FP8_BLOCK_SCALES
+    kv_cache_quant_algo: FP8
+    accuracy: 95.413
 Qwen3/Qwen3-8B:
   - accuracy: 87.1114
   - quant_algo: FP8

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

@@ -163,6 +163,9 @@ deepseek-ai/DeepSeek-R1:
   - quant_algo: FP8_BLOCK_SCALES
     spec_dec_algo: MTP
     accuracy: 87.573
+  - quant_algo: FP8_BLOCK_SCALES
+    kv_cache_quant_algo: FP8
+    accuracy: 87.573
 Qwen3/Qwen3-8B:
   - quant_algo: W4A8_MXFP4_FP8
     accuracy: 72.70

tests/integration/defs/accuracy/test_llm_api_pytorch.py

@@ -1663,10 +1663,17 @@ def test_nvfp4_multi_gpus(self, tp_size, pp_size, ep_size, mtp_nextn, fp8kv,
     def test_fp8_blockscale(self, tp_size, pp_size, ep_size, mtp_nextn, fp8kv,
                             attention_dp, cuda_graph, overlap_scheduler,
                             max_batch_size):
-        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.9)
+        if get_sm_version() == 100:
+            moe_config = MoeConfig(backend="DEEPGEMM", max_num_tokens=16384)
+            kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.6)
+        else:
+            moe_config = MoeConfig()
+            kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.9)
+
         pytorch_config = dict(
             disable_overlap_scheduler=not overlap_scheduler,
             cuda_graph_config=CudaGraphConfig() if cuda_graph else None,
+            moe_config=moe_config,
         )
 
         if fp8kv:

tests/integration/test_lists/qa/llm_function_full.txt

@@ -498,6 +498,7 @@ accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_nvfp4_multi_gpus[latency_
 accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_nvfp4_multi_gpus[throughput]
 accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_nvfp4_multi_gpus[throughput_tp8]
 accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_nvfp4_multi_gpus[throughput_tp4]
+accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_fp8_blockscale[throughput]
 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-torch_compile=False]
 accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_fp8_block_scales[latency-torch_compile=True]

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

@@ -16,6 +16,7 @@ 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::TestDeepSeekR1::test_fp8_blockscale[throughput] 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)