Merge a8813edb44 into 6df2c8a074

2026-01-13 22:18:36 +08:00 · 2026-01-13 21:32:11 +08:00 · 2026-01-13 21:32:11 +08:00 · c9b01f8278
commit c9b01f8278
parent 6df2c8a074 a8813edb44
14 changed files with 803 additions and 170 deletions
--- a/jenkins/L0_Test.groovy
+++ b/jenkins/L0_Test.groovy
@ -3267,12 +3267,9 @@ def launchTestJobs(pipeline, testFilter)
        "DGX_B300-4_GPUs-PyTorch-Post-Merge-1": ["b300-x4", "l0_dgx_b300", 1, 2, 4],
        "DGX_B300-4_GPUs-PyTorch-Post-Merge-2": ["b300-x4", "l0_dgx_b300", 2, 2, 4],
        // PerfSanity post-merge tests
-        // "DGX_B200-4_GPUs-PyTorch-PerfSanity-Post-Merge-1": ["b200-x4", "l0_dgx_b200_perf_sanity", 1, 3, 4],
-        // "DGX_B200-4_GPUs-PyTorch-PerfSanity-Post-Merge-2": ["b200-x4", "l0_dgx_b200_perf_sanity", 2, 3, 4],
-        // "DGX_B200-4_GPUs-PyTorch-PerfSanity-Post-Merge-3": ["b200-x4", "l0_dgx_b200_perf_sanity", 3, 3, 4],
-        // "DGX_B200-8_GPUs-PyTorch-PerfSanity-Post-Merge-1": ["b200-x8", "l0_dgx_b200_perf_sanity", 1, 3, 8],
-        // "DGX_B200-8_GPUs-PyTorch-PerfSanity-Post-Merge-2": ["b200-x8", "l0_dgx_b200_perf_sanity", 2, 3, 8],
-        // "DGX_B200-8_GPUs-PyTorch-PerfSanity-Post-Merge-3": ["b200-x8", "l0_dgx_b200_perf_sanity", 3, 3, 8],
+        "DGX_B200-8_GPUs-PyTorch-PerfSanity-Post-Merge-1": ["b200-x8-lbd", "l0_dgx_b200_perf_sanity", 1, 3, 8],
+        "DGX_B200-8_GPUs-PyTorch-PerfSanity-Post-Merge-2": ["b200-x8-lbd", "l0_dgx_b200_perf_sanity", 2, 3, 8],
+        "DGX_B200-8_GPUs-PyTorch-PerfSanity-Post-Merge-3": ["b200-x8-lbd", "l0_dgx_b200_perf_sanity", 3, 3, 8],
    ]
    fullSet += x86SlurmTestConfigs.keySet()

@ -3306,8 +3303,7 @@ def launchTestJobs(pipeline, testFilter)
        // "GB300-PyTorch-1": ["gb300-single", "l0_gb300", 1, 1],
        // "GB300-4_GPUs-PyTorch-Post-Merge-1": ["gb300-x4", "l0_gb300_multi_gpus", 1, 1, 4],
        // PerfSanity pre-merge tests
-        "GB200-4_GPUs-PyTorch-PerfSanity-1": ["gb200-x4-oci", "l0_gb200_multi_gpus_perf_sanity", 1, 2, 4],
-        "GB200-4_GPUs-PyTorch-PerfSanity-2": ["gb200-x4-oci", "l0_gb200_multi_gpus_perf_sanity", 2, 2, 4],
+        "GB200-4_GPUs-PyTorch-PerfSanity-1": ["gb200-x4-oci", "l0_gb200_multi_gpus_perf_sanity", 1, 1, 4],
        // PerfSanity post-merge tests
        "GB200-4_GPUs-PyTorch-PerfSanity-Post-Merge-1": ["gb200-x4-oci", "l0_gb200_multi_gpus_perf_sanity", 1, 3, 4],
        "GB200-4_GPUs-PyTorch-PerfSanity-Post-Merge-2": ["gb200-x4-oci", "l0_gb200_multi_gpus_perf_sanity", 2, 3, 4],
@ -3324,8 +3320,12 @@ def launchTestJobs(pipeline, testFilter)
        "GB200-8_GPUs-2_Nodes-PyTorch-Post-Merge-2": ["gb200-oci-trtllm", "l0_gb200_multi_nodes", 2, 3, 8, 2],
        "GB200-8_GPUs-2_Nodes-PyTorch-Post-Merge-3": ["gb200-oci-trtllm", "l0_gb200_multi_nodes", 3, 3, 8, 2],
        // PerfSanity post-merge tests
-        "GB200-8_GPUs-2_Nodes-PyTorch-PerfSanity-Post-Merge-1": ["gb200-oci-trtllm", "l0_gb200_multi_nodes_aggr_perf_sanity_2_nodes", 1, 2, 8, 2],
-        "GB200-8_GPUs-2_Nodes-PyTorch-PerfSanity-Post-Merge-2": ["gb200-oci-trtllm", "l0_gb200_multi_nodes_aggr_perf_sanity_2_nodes", 2, 2, 8, 2],
+        // Multi-Node perf tests run one pytest per stage to avoid OOM.
+        "GB200-8_GPUs-2_Nodes-PyTorch-PerfSanity-Post-Merge-1": ["gb200-oci-trtllm", "l0_gb200_multi_nodes_aggr_perf_sanity_2_nodes", 1, 5, 8, 2],
+        "GB200-8_GPUs-2_Nodes-PyTorch-PerfSanity-Post-Merge-2": ["gb200-oci-trtllm", "l0_gb200_multi_nodes_aggr_perf_sanity_2_nodes", 2, 5, 8, 2],
+        "GB200-8_GPUs-2_Nodes-PyTorch-PerfSanity-Post-Merge-3": ["gb200-oci-trtllm", "l0_gb200_multi_nodes_aggr_perf_sanity_2_nodes", 3, 5, 8, 2],
+        "GB200-8_GPUs-2_Nodes-PyTorch-PerfSanity-Post-Merge-4": ["gb200-oci-trtllm", "l0_gb200_multi_nodes_aggr_perf_sanity_2_nodes", 4, 5, 8, 2],
+        "GB200-8_GPUs-2_Nodes-PyTorch-PerfSanity-Post-Merge-5": ["gb200-oci-trtllm", "l0_gb200_multi_nodes_aggr_perf_sanity_2_nodes", 5, 5, 8, 2],
        "GB200-12_GPUs-3_Nodes-PyTorch-PerfSanity-Disagg-Post-Merge-1": ["gb200-oci-trtllm", "l0_gb200_multi_nodes_disagg_perf_sanity_3_nodes", 1, 1, 12, 3],
        // "GB200-24_GPUs-6_Nodes-PyTorch-PerfSanity-Disagg-Post-Merge-1": ["gb200-oci-trtllm", "l0_gb200_multi_nodes_disagg_perf_sanity_6_nodes", 1, 2, 24, 6],
        // "GB200-24_GPUs-6_Nodes-PyTorch-PerfSanity-Disagg-Post-Merge-2": ["gb200-oci-trtllm", "l0_gb200_multi_nodes_disagg_perf_sanity_6_nodes", 2, 2, 24, 6],
--- a/tests/integration/defs/perf/test_perf_sanity.py
+++ b/tests/integration/defs/perf/test_perf_sanity.py
@ -54,7 +54,9 @@ MODEL_PATH_DICT = {
    "deepseek_r1_0528_fp8": "DeepSeek-R1/DeepSeek-R1-0528/",
    "deepseek_r1_0528_fp4": "DeepSeek-R1/DeepSeek-R1-0528-FP4/",
    "deepseek_r1_0528_fp4_v2": "DeepSeek-R1/DeepSeek-R1-0528-FP4-v2/",
+    "deepseek_v32_fp4": "DeepSeek-V3.2-Exp-FP4-v2",
    "gpt_oss_120b_fp4": "gpt_oss/gpt-oss-120b",
+    "k2_thinking_fp4": "Kimi-K2-Thinking-NVFP4",
 }

 SUPPORTED_GPU_TYPE = [
@ -852,19 +854,24 @@ class PerfSanityTestConfig:

        def get_gpu_type() -> str:
            try:
+                # GB200 uses dgx_b200 for wrongly adding dgx_b200 to opensearch in the past.
+                mapping = {
+                    "GB200": "dgx_b200",
+                    "GB300": "gb300",
+                    "B200": "b200",
+                    "B300": "b300",
+                }
                output = subprocess.check_output(
-                    ["nvidia-smi", "-L"], stderr=subprocess.DEVNULL, text=True
+                    "nvidia-smi -q | grep 'Product Name' | head -1",
+                    shell=True,
+                    stderr=subprocess.DEVNULL,
+                    text=True,
                )
-                first_line = output.strip().split("\n")[0]
-                gpu_models = SUPPORTED_GPU_TYPE
-                for model in gpu_models:
-                    if model in first_line:
-                        if model.startswith("B") and not model.startswith("GB"):
-                            return f"dgx_{model.lower()}"
-                        return model.lower()
+                model = output.split()[-1]
+                return mapping.get(model, "unsupported")
            except (subprocess.CalledProcessError, FileNotFoundError, IndexError):
                print_error("Failed to get GPU type")
-            return ""
+            return "unsupported"

        assert len(labels) > 1, "perf_sanity test must have a config file!"
        is_disagg = "disagg" in labels[0]
--- a/tests/integration/test_lists/test-db/l0_dgx_b200_perf_sanity.yml
+++ b/tests/integration/test_lists/test-db/l0_dgx_b200_perf_sanity.yml
@ -15,25 +15,21 @@ l0_dgx_b200_perf_sanity:
      backend: pytorch
      orchestrator: mpi
  tests:
-  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp8_blackwell-r1_fp8_dep8_mtp1_1k1k] TIMEOUT (180)
-  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp8_blackwell-r1_fp8_tep8_mtp3_1k1k] TIMEOUT (180)
-  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp8_blackwell-r1_fp8_tp8_mtp3_1k1k] TIMEOUT (180)
-
- condition:
-    ranges:
-      system_gpu_count:
-        gte: 4
-        lte: 4
-    wildcards:
-      gpu:
-      - '*b200*'
-      linux_distribution_name: ubuntu*
-      cpu: x86_64
-    terms:
-      stage: post_merge
-      backend: pytorch
-      orchestrator: mpi
-  tests:
-  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp4_v2_blackwell-r1_fp4_v2_dep4_mtp1_1k1k] TIMEOUT (180)
-  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp4_v2_blackwell-r1_fp4_v2_tep4_mtp3_1k1k] TIMEOUT (180)
-  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp4_v2_blackwell-r1_fp4_v2_tp4_mtp3_1k1k] TIMEOUT (180)
+  # deepseek_r1_fp8_blackwell
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp8_blackwell-r1_fp8_tp8_mtp3_1k1k] TIMEOUT (60)
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp8_blackwell-r1_fp8_dep8_mtp1_1k1k] TIMEOUT (60)
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp8_blackwell-r1_fp8_tp8_mtp3_8k1k] TIMEOUT (60)
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp8_blackwell-r1_fp8_dep8_mtp1_8k1k] TIMEOUT (60)
+  # deepseek_r1_fp4_v2_blackwell
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp4_v2_blackwell-r1_fp4_v2_tp4_mtp3_1k1k] TIMEOUT (60)
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp4_v2_blackwell-r1_fp4_v2_tp4_mtp3_8k1k] TIMEOUT (60)
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp4_v2_blackwell-r1_fp4_v2_dep8_mtp1_1k1k] TIMEOUT (60)
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp4_v2_blackwell-r1_fp4_v2_dep8_mtp1_8k1k] TIMEOUT (60)
+  # deepseek_v32_fp4_blackwell
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_v32_fp4_blackwell-v32_fp4_tep8_mtp3_8k1k] TIMEOUT (60)
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_v32_fp4_blackwell-v32_fp4_dep8_mtp1_8k1k] TIMEOUT (60)
+  # k2_thinking_fp4_blackwell
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-k2_thinking_fp4_blackwell-k2_thinking_fp4_tep8_8k1k] TIMEOUT (60)
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-k2_thinking_fp4_blackwell-k2_thinking_fp4_dep8_8k1k] TIMEOUT (60)
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-k2_thinking_fp4_blackwell-k2_thinking_fp4_tep8_32k8k] TIMEOUT (90)
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-k2_thinking_fp4_blackwell-k2_thinking_fp4_dep8_32k8k] TIMEOUT (90)
--- a/tests/integration/test_lists/test-db/l0_gb200_multi_gpus_perf_sanity.yml
+++ b/tests/integration/test_lists/test-db/l0_gb200_multi_gpus_perf_sanity.yml
@ -14,12 +14,14 @@ l0_gb200_multi_gpus_perf_sanity:
      stage: pre_merge
      backend: pytorch
  tests:
+  # deepseek_r1_fp4_v2_grace_blackwell
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp4_v2_grace_blackwell-r1_fp4_v2_dep4_mtp1_1k1k] TIMEOUT (90)
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp4_v2_grace_blackwell-r1_fp4_v2_tep4_mtp3_1k1k] TIMEOUT (60)
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp4_v2_grace_blackwell-r1_fp4_v2_tp4_mtp3_1k1k] TIMEOUT (60)
  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp4_v2_grace_blackwell-r1_fp4_v2_dep4_mtp1_8k1k] TIMEOUT (90)
  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp4_v2_grace_blackwell-r1_fp4_v2_tep4_mtp3_8k1k] TIMEOUT (60)
  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp4_v2_grace_blackwell-r1_fp4_v2_tp4_mtp3_8k1k] TIMEOUT (60)
-  - perf/test_perf_sanity.py::test_e2e[aggr_upload-gpt_oss_120b_fp4_grace_blackwell-gpt_oss_fp4_dep4_1k8k] TIMEOUT (90)
-  - perf/test_perf_sanity.py::test_e2e[aggr_upload-gpt_oss_120b_fp4_grace_blackwell-gpt_oss_fp4_dep2_1k1k] TIMEOUT (90)
-  - perf/test_perf_sanity.py::test_e2e[aggr_upload-gpt_oss_120b_fp4_grace_blackwell-gpt_oss_fp4_tep2_1k8k] TIMEOUT (90)
+
 - condition:
    ranges:
      system_gpu_count:
@ -34,11 +36,22 @@ l0_gb200_multi_gpus_perf_sanity:
      stage: post_merge
      backend: pytorch
  tests:
-  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp4_v2_grace_blackwell-r1_fp4_v2_dep4_mtp1_1k1k] TIMEOUT (90)
-  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp4_v2_grace_blackwell-r1_fp4_v2_tep4_mtp3_1k1k] TIMEOUT (60)
-  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp4_v2_grace_blackwell-r1_fp4_v2_tp4_mtp3_1k1k] TIMEOUT (60)
-  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp4_v2_grace_blackwell-r1_fp4_v2_dep4_mtp1_1k8k] TIMEOUT (120)
+  # deepseek_r1_fp4_v2_grace_blackwell
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp4_v2_grace_blackwell-r1_fp4_v2_dep4_mtp1_1k8k] TIMEOUT (6060)
  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp4_v2_grace_blackwell-r1_fp4_v2_tep4_mtp3_1k8k] TIMEOUT (60)
  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp4_v2_grace_blackwell-r1_fp4_v2_tp4_mtp3_1k8k] TIMEOUT (60)
-  - perf/test_perf_sanity.py::test_e2e[aggr_upload-gpt_oss_120b_fp4_grace_blackwell-gpt_oss_fp4_tp2_1k8k] TIMEOUT (90)
-  - perf/test_perf_sanity.py::test_e2e[aggr_upload-gpt_oss_120b_fp4_grace_blackwell-gpt_oss_fp4_tp4_eagle3_1k1k] TIMEOUT (90)
+  # deepseek_v32_fp4_grace_blackwell
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_v32_fp4_grace_blackwell-v32_fp4_tep4_mtp3_1k1k] TIMEOUT (60)
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_v32_fp4_grace_blackwell-v32_fp4_dep4_mtp1_1k1k] TIMEOUT (60)
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_v32_fp4_grace_blackwell-v32_fp4_tep4_mtp3_8k1k] TIMEOUT (60)
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_v32_fp4_grace_blackwell-v32_fp4_dep4_mtp1_8k1k] TIMEOUT (60)
+  # k2_thinking_fp4_grace_blackwell
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-k2_thinking_fp4_grace_blackwell-k2_thinking_fp4_tep4_8k1k] TIMEOUT (90)
+  # k2_thinking_fp4_grace_blackwell - DEP4 with CUTLASS (hang)
+  # - perf/test_perf_sanity.py::test_e2e[aggr_upload-k2_thinking_fp4_grace_blackwell-k2_thinking_fp4_dep4_8k1k] TIMEOUT (90)
+  # gpt_oss_120b_fp4_grace_blackwell
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-gpt_oss_120b_fp4_grace_blackwell-gpt_oss_fp4_tp2_1k8k] TIMEOUT (60)
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-gpt_oss_120b_fp4_grace_blackwell-gpt_oss_fp4_dep4_1k8k] TIMEOUT (60)
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-gpt_oss_120b_fp4_grace_blackwell-gpt_oss_fp4_dep2_1k1k] TIMEOUT (60)
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-gpt_oss_120b_fp4_grace_blackwell-gpt_oss_fp4_tep2_1k8k] TIMEOUT (60)
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-gpt_oss_120b_fp4_grace_blackwell-gpt_oss_fp4_tp4_eagle3_1k1k] TIMEOUT (60)
--- a/tests/integration/test_lists/test-db/l0_gb200_multi_nodes_aggr_perf_sanity_2_nodes.yml
+++ b/tests/integration/test_lists/test-db/l0_gb200_multi_nodes_aggr_perf_sanity_2_nodes.yml
@ -13,5 +13,10 @@ l0_gb200_multi_nodes_aggr_perf_sanity_2_nodes:
      stage: post_merge
      backend: pytorch
  tests:
-  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp4_v2_2_nodes_grace_blackwell-r1_fp4_v2_dep8_mtp1] TIMEOUT (90)
+  # deepseek_r1_fp4_v2_2_nodes_grace_blackwell
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp4_v2_2_nodes_grace_blackwell-r1_fp4_v2_dep8_mtp1_1k1k] TIMEOUT (90)
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp4_v2_2_nodes_grace_blackwell-r1_fp4_v2_dep8_mtp1_8k1k] TIMEOUT (90)
  - perf/test_perf_sanity.py::test_e2e[aggr_upload-deepseek_r1_fp4_v2_2_nodes_grace_blackwell-r1_fp4_v2_tep8_mtp3] TIMEOUT (90)
+  # k2_thinking_fp4_2_nodes_grace_blackwell
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-k2_thinking_fp4_2_nodes_grace_blackwell-k2_thinking_fp4_tep8_32k8k] TIMEOUT (90)
+  - perf/test_perf_sanity.py::test_e2e[aggr_upload-k2_thinking_fp4_2_nodes_grace_blackwell-k2_thinking_fp4_dep8_32k8k] TIMEOUT (90)
--- a/tests/scripts/perf-sanity/deepseek_r1_fp4_v2_2_nodes_grace_blackwell.yaml
+++ b/tests/scripts/perf-sanity/deepseek_r1_fp4_v2_2_nodes_grace_blackwell.yaml
@ -2,18 +2,18 @@ metadata:
  model_name: deepseek_r1_0528_fp4_v2
  supported_gpus:
  - GB200
-  - GB300
 hardware:
  gpus_per_node: 4
 server_configs:
-  - name: "r1_fp4_v2_dep8_mtp1"
+  # 1k1k configs - DEP8 with CUTLASS, MTP1
+  - name: "r1_fp4_v2_dep8_mtp1_1k1k"
    model_name: "deepseek_r1_0528_fp4_v2"
    trust_remote_code: true
    tensor_parallel_size: 8
    moe_expert_parallel_size: 8
    pipeline_parallel_size: 1
    max_batch_size: 512
-    max_num_tokens: 3136
+    max_num_tokens: 8192
    attn_backend: "TRTLLM"
    enable_attention_dp: true
    attention_dp_config:
@ -28,16 +28,56 @@ server_configs:
    kv_cache_config:
      dtype: 'fp8'
      enable_block_reuse: false
-      free_gpu_memory_fraction: 0.5
+      free_gpu_memory_fraction: 0.8
+    speculative_config:
+      decoding_type: 'MTP'
+      num_nextn_predict_layers: 1
    client_configs:
-      - name: "con32_iter12_1k1k"
-        concurrency: 32
-        iterations: 12
+      - name: "con1024_iter10_1k1k"
+        concurrency: 1024
+        iterations: 10
        isl: 1024
        osl: 1024
-        random_range_ratio: 0.8
+        random_range_ratio: 0.2
        backend: "openai"

+  # 8k1k configs - DEP8 with CUTLASS, MTP1
+  - name: "r1_fp4_v2_dep8_mtp1_8k1k"
+    model_name: "deepseek_r1_0528_fp4_v2"
+    trust_remote_code: true
+    tensor_parallel_size: 8
+    moe_expert_parallel_size: 8
+    pipeline_parallel_size: 1
+    max_batch_size: 512
+    max_num_tokens: 12288
+    attn_backend: "TRTLLM"
+    enable_attention_dp: true
+    attention_dp_config:
+      batching_wait_iters: 0
+      enable_balance: true
+      timeout_iters: 60
+    moe_config:
+      backend: 'CUTLASS'
+    cuda_graph_config:
+      enable_padding: true
+      max_batch_size: 512
+    kv_cache_config:
+      dtype: 'fp8'
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.8
+    speculative_config:
+      decoding_type: 'MTP'
+      num_nextn_predict_layers: 1
+    client_configs:
+      - name: "con1024_iter10_8k1k"
+        concurrency: 1024
+        iterations: 10
+        isl: 8192
+        osl: 1024
+        random_range_ratio: 0.2
+        backend: "openai"
+
+  # 1k1k configs - TEP8 with TRTLLM, MTP3
  - name: "r1_fp4_v2_tep8_mtp3"
    model_name: "deepseek_r1_0528_fp4_v2"
    trust_remote_code: true
--- a/tests/scripts/perf-sanity/deepseek_r1_fp4_v2_blackwell.yaml
+++ b/tests/scripts/perf-sanity/deepseek_r1_fp4_v2_blackwell.yaml
@ -2,72 +2,8 @@ metadata:
  model_name: deepseek_r1_0528_fp4_v2
  supported_gpus:
  - B200
-  - B300
 server_configs:
-  - name: "r1_fp4_v2_dep4_mtp1_1k1k"
-    model_name: "deepseek_r1_0528_fp4_v2"
-    tensor_parallel_size: 4
-    moe_expert_parallel_size: 4
-    pipeline_parallel_size: 1
-    max_batch_size: 512
-    max_num_tokens: 8192
-    attn_backend: "TRTLLM"
-    enable_attention_dp: true
-    attention_dp_config:
-      batching_wait_iters: 0
-      enable_balance: true
-      timeout_iters: 60
-    moe_config:
-      backend: 'CUTLASS'
-    cuda_graph_config:
-      enable_padding: true
-      max_batch_size: 512
-    kv_cache_config:
-      dtype: 'fp8'
-      enable_block_reuse: false
-      free_gpu_memory_fraction: 0.8
-    speculative_config:
-      decoding_type: 'MTP'
-      num_nextn_predict_layers: 1
-    client_configs:
-      - name: "con2048_iter5_1k1k"
-        concurrency: 2048
-        iterations: 5
-        isl: 1024
-        osl: 1024
-        random_range_ratio: 0.8
-        backend: "openai"
-
-  - name: "r1_fp4_v2_tep4_mtp3_1k1k"
-    model_name: "deepseek_r1_0528_fp4_v2"
-    tensor_parallel_size: 4
-    moe_expert_parallel_size: 4
-    pipeline_parallel_size: 1
-    max_batch_size: 32
-    max_num_tokens: 8192
-    attn_backend: "TRTLLM"
-    enable_attention_dp: false
-    moe_config:
-      backend: 'TRTLLM'
-    cuda_graph_config:
-      enable_padding: true
-      max_batch_size: 32
-    kv_cache_config:
-      dtype: 'fp8'
-      enable_block_reuse: false
-      free_gpu_memory_fraction: 0.8
-    speculative_config:
-      decoding_type: 'MTP'
-      num_nextn_predict_layers: 3
-    client_configs:
-      - name: "con32_iter10_1k1k"
-        concurrency: 32
-        iterations: 10
-        isl: 1024
-        osl: 1024
-        random_range_ratio: 0.8
-        backend: "openai"
-
+  # 1k1k configs - TP4 with TRTLLM, MTP3
  - name: "r1_fp4_v2_tp4_mtp3_1k1k"
    model_name: "deepseek_r1_0528_fp4_v2"
    tensor_parallel_size: 4
@ -90,10 +26,111 @@ server_configs:
      decoding_type: 'MTP'
      num_nextn_predict_layers: 3
    client_configs:
-      - name: "con4_iter10_1k1k"
-        concurrency: 4
+      - name: "con2_iter10_1k1k"
+        concurrency: 2
        iterations: 10
        isl: 1024
        osl: 1024
-        random_range_ratio: 0.8
+        random_range_ratio: 0.2
+        backend: "openai"
+
+  # 1k1k configs - DEP8 with CUTLASS, MTP1
+  - name: "r1_fp4_v2_dep8_mtp1_1k1k"
+    model_name: "deepseek_r1_0528_fp4_v2"
+    tensor_parallel_size: 8
+    moe_expert_parallel_size: 8
+    pipeline_parallel_size: 1
+    max_batch_size: 512
+    max_num_tokens: 8192
+    attn_backend: "TRTLLM"
+    enable_attention_dp: true
+    attention_dp_config:
+      batching_wait_iters: 0
+      enable_balance: true
+      timeout_iters: 60
+    moe_config:
+      backend: 'CUTLASS'
+    cuda_graph_config:
+      enable_padding: true
+      max_batch_size: 512
+    kv_cache_config:
+      dtype: 'fp8'
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.8
+    speculative_config:
+      decoding_type: 'MTP'
+      num_nextn_predict_layers: 1
+    client_configs:
+      - name: "con1024_iter10_1k1k"
+        concurrency: 1024
+        iterations: 10
+        isl: 1024
+        osl: 1024
+        random_range_ratio: 0.2
+        backend: "openai"
+
+  # 8k1k configs - TP4 with TRTLLM, MTP3
+  - name: "r1_fp4_v2_tp4_mtp3_8k1k"
+    model_name: "deepseek_r1_0528_fp4_v2"
+    tensor_parallel_size: 4
+    moe_expert_parallel_size: 1
+    pipeline_parallel_size: 1
+    max_batch_size: 4
+    max_num_tokens: 12288
+    attn_backend: "TRTLLM"
+    enable_attention_dp: false
+    moe_config:
+      backend: 'TRTLLM'
+    cuda_graph_config:
+      enable_padding: true
+      max_batch_size: 4
+    kv_cache_config:
+      dtype: 'fp8'
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.8
+    speculative_config:
+      decoding_type: 'MTP'
+      num_nextn_predict_layers: 3
+    client_configs:
+      - name: "con4_iter10_8k1k"
+        concurrency: 4
+        iterations: 10
+        isl: 8192
+        osl: 1024
+        random_range_ratio: 0.2
+        backend: "openai"
+
+  # 8k1k configs - DEP8 with CUTLASS, MTP1
+  - name: "r1_fp4_v2_dep8_mtp1_8k1k"
+    model_name: "deepseek_r1_0528_fp4_v2"
+    tensor_parallel_size: 8
+    moe_expert_parallel_size: 8
+    pipeline_parallel_size: 1
+    max_batch_size: 512
+    max_num_tokens: 12288
+    attn_backend: "TRTLLM"
+    enable_attention_dp: true
+    attention_dp_config:
+      batching_wait_iters: 0
+      enable_balance: true
+      timeout_iters: 60
+    moe_config:
+      backend: 'CUTLASS'
+    cuda_graph_config:
+      enable_padding: true
+      max_batch_size: 512
+    kv_cache_config:
+      dtype: 'fp8'
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.8
+    speculative_config:
+      decoding_type: 'MTP'
+      num_nextn_predict_layers: 1
+    client_configs:
+      - name: "con256_iter10_8k1k"
+        concurrency: 256
+        iterations: 10
+        isl: 8192
+        osl: 1024
+        random_range_ratio: 0.2
        backend: "openai"
--- a/tests/scripts/perf-sanity/deepseek_r1_fp4_v2_grace_blackwell.yaml
+++ b/tests/scripts/perf-sanity/deepseek_r1_fp4_v2_grace_blackwell.yaml
@ -4,7 +4,7 @@ metadata:
  - GB200
  - GB300
 server_configs:
-  # 1k1k configs
+  # 1k1k configs - DEP4 with CUTLASS, MTP1
  - name: "r1_fp4_v2_dep4_mtp1_1k1k"
    model_name: "deepseek_r1_0528_fp4_v2"
    tensor_parallel_size: 4
@ -38,7 +38,15 @@ server_configs:
        osl: 1024
        random_range_ratio: 0.8
        backend: "openai"
+      - name: "con1024_iter10_1k1k"
+        concurrency: 1024
+        iterations: 10
+        isl: 1024
+        osl: 1024
+        random_range_ratio: 0.2
+        backend: "openai"

+  # 1k1k configs - TEP4 with TRTLLM, MTP3
  - name: "r1_fp4_v2_tep4_mtp3_1k1k"
    model_name: "deepseek_r1_0528_fp4_v2"
    tensor_parallel_size: 4
@ -69,6 +77,7 @@ server_configs:
        random_range_ratio: 0.8
        backend: "openai"

+  # 1k1k configs - TP4 with TRTLLM, MTP3
  - name: "r1_fp4_v2_tp4_mtp3_1k1k"
    model_name: "deepseek_r1_0528_fp4_v2"
    tensor_parallel_size: 4
@ -98,8 +107,15 @@ server_configs:
        osl: 1024
        random_range_ratio: 0.8
        backend: "openai"
+      - name: "con2_iter10_1k1k"
+        concurrency: 2
+        iterations: 10
+        isl: 1024
+        osl: 1024
+        random_range_ratio: 0.2
+        backend: "openai"

-  # 8k1k configs
+  # 8k1k configs - DEP4 with CUTLASS, MTP1
  - name: "r1_fp4_v2_dep4_mtp1_8k1k"
    model_name: "deepseek_r1_0528_fp4_v2"
    tensor_parallel_size: 4
@ -133,7 +149,15 @@ server_configs:
        osl: 1024
        random_range_ratio: 0.2
        backend: "openai"
+      - name: "con256_iter10_8k1k"
+        concurrency: 256
+        iterations: 10
+        isl: 8192
+        osl: 1024
+        random_range_ratio: 0.2
+        backend: "openai"

+  # 8k1k configs - TEP4 with TRTLLM, MTP3
  - name: "r1_fp4_v2_tep4_mtp3_8k1k"
    model_name: "deepseek_r1_0528_fp4_v2"
    tensor_parallel_size: 4
@ -164,6 +188,7 @@ server_configs:
        random_range_ratio: 0.2
        backend: "openai"

+  # 8k1k configs - TP4 with TRTLLM, MTP3
  - name: "r1_fp4_v2_tp4_mtp3_8k1k"
    model_name: "deepseek_r1_0528_fp4_v2"
    tensor_parallel_size: 4
@ -193,8 +218,15 @@ server_configs:
        osl: 1024
        random_range_ratio: 0.2
        backend: "openai"
+      - name: "con2_iter10_8k1k"
+        concurrency: 2
+        iterations: 10
+        isl: 8192
+        osl: 1024
+        random_range_ratio: 0.2
+        backend: "openai"

-  # 1k8k configs
+  # 1k8k configs - DEP4 with CUTLASS, MTP1
  - name: "r1_fp4_v2_dep4_mtp1_1k8k"
    model_name: "deepseek_r1_0528_fp4_v2"
    tensor_parallel_size: 4
@ -229,6 +261,7 @@ server_configs:
        random_range_ratio: 0.8
        backend: "openai"

+  # 1k8k configs - TEP4 with TRTLLM, MTP3
  - name: "r1_fp4_v2_tep4_mtp3_1k8k"
    model_name: "deepseek_r1_0528_fp4_v2"
    tensor_parallel_size: 4
@ -259,6 +292,7 @@ server_configs:
        random_range_ratio: 0.8
        backend: "openai"

+  # 1k8k configs - TP4 with TRTLLM, MTP3
  - name: "r1_fp4_v2_tp4_mtp3_1k8k"
    model_name: "deepseek_r1_0528_fp4_v2"
    tensor_parallel_size: 4
--- a/tests/scripts/perf-sanity/deepseek_r1_fp8_blackwell.yaml
+++ b/tests/scripts/perf-sanity/deepseek_r1_fp8_blackwell.yaml
@ -2,8 +2,39 @@ metadata:
  model_name: deepseek_r1_0528_fp8
  supported_gpus:
  - B200
-  - B300
 server_configs:
+  # 1k1k configs - TP8 with TRTLLM, MTP3
+  - name: "r1_fp8_tp8_mtp3_1k1k"
+    model_name: "deepseek_r1_0528_fp8"
+    tensor_parallel_size: 8
+    moe_expert_parallel_size: 1
+    pipeline_parallel_size: 1
+    max_batch_size: 8
+    max_num_tokens: 8192
+    attn_backend: "TRTLLM"
+    enable_attention_dp: false
+    moe_config:
+      backend: 'TRTLLM'
+    cuda_graph_config:
+      enable_padding: true
+      max_batch_size: 8
+    kv_cache_config:
+      dtype: 'fp8'
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.8
+    speculative_config:
+      decoding_type: 'MTP'
+      num_nextn_predict_layers: 3
+    client_configs:
+      - name: "con4_iter10_1k1k"
+        concurrency: 4
+        iterations: 10
+        isl: 1024
+        osl: 1024
+        random_range_ratio: 0.2
+        backend: "openai"
+
+  # 1k1k configs - DEP8 with DEEPGEMM, MTP1
  - name: "r1_fp8_dep8_mtp1_1k1k"
    model_name: "deepseek_r1_0528_fp8"
    tensor_parallel_size: 8
@ -30,51 +61,22 @@ server_configs:
      decoding_type: 'MTP'
      num_nextn_predict_layers: 1
    client_configs:
-      - name: "con4096_iter5_1k1k"
-        concurrency: 4096
-        iterations: 5
-        isl: 1024
-        osl: 1024
-        random_range_ratio: 0.8
-        backend: "openai"
-
-  - name: "r1_fp8_tep8_mtp3_1k1k"
-    model_name: "deepseek_r1_0528_fp8"
-    tensor_parallel_size: 8
-    moe_expert_parallel_size: 8
-    pipeline_parallel_size: 1
-    max_batch_size: 64
-    max_num_tokens: 8192
-    attn_backend: "TRTLLM"
-    enable_attention_dp: false
-    moe_config:
-      backend: 'DEEPGEMM'
-    cuda_graph_config:
-      enable_padding: true
-      max_batch_size: 64
-    kv_cache_config:
-      dtype: 'fp8'
-      enable_block_reuse: false
-      free_gpu_memory_fraction: 0.8
-    speculative_config:
-      decoding_type: 'MTP'
-      num_nextn_predict_layers: 3
-    client_configs:
-      - name: "con64_iter10_1k1k"
-        concurrency: 64
+      - name: "con1024_iter10_1k1k"
+        concurrency: 1024
        iterations: 10
        isl: 1024
        osl: 1024
-        random_range_ratio: 0.8
+        random_range_ratio: 0.2
        backend: "openai"

-  - name: "r1_fp8_tp8_mtp3_1k1k"
+  # 8k1k configs - TP8 with TRTLLM, MTP3
+  - name: "r1_fp8_tp8_mtp3_8k1k"
    model_name: "deepseek_r1_0528_fp8"
    tensor_parallel_size: 8
    moe_expert_parallel_size: 1
    pipeline_parallel_size: 1
    max_batch_size: 8
-    max_num_tokens: 8192
+    max_num_tokens: 12288
    attn_backend: "TRTLLM"
    enable_attention_dp: false
    moe_config:
@ -90,10 +92,45 @@ server_configs:
      decoding_type: 'MTP'
      num_nextn_predict_layers: 3
    client_configs:
-      - name: "con8_iter10_1k1k"
-        concurrency: 8
+      - name: "con4_iter10_8k1k"
+        concurrency: 4
        iterations: 10
-        isl: 1024
+        isl: 8192
        osl: 1024
-        random_range_ratio: 0.8
+        random_range_ratio: 0.2
+        backend: "openai"
+
+  # 8k1k configs - DEP8 with DEEPGEMM, MTP1
+  - name: "r1_fp8_dep8_mtp1_8k1k"
+    model_name: "deepseek_r1_0528_fp8"
+    tensor_parallel_size: 8
+    moe_expert_parallel_size: 8
+    pipeline_parallel_size: 1
+    max_batch_size: 512
+    max_num_tokens: 12288
+    attn_backend: "TRTLLM"
+    enable_attention_dp: true
+    attention_dp_config:
+      batching_wait_iters: 0
+      enable_balance: true
+      timeout_iters: 60
+    moe_config:
+      backend: 'DEEPGEMM'
+    cuda_graph_config:
+      enable_padding: true
+      max_batch_size: 512
+    kv_cache_config:
+      dtype: 'fp8'
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.8
+    speculative_config:
+      decoding_type: 'MTP'
+      num_nextn_predict_layers: 1
+    client_configs:
+      - name: "con256_iter10_8k1k"
+        concurrency: 256
+        iterations: 10
+        isl: 8192
+        osl: 1024
+        random_range_ratio: 0.2
        backend: "openai"
--- a/tests/scripts/perf-sanity/deepseek_v32_fp4_blackwell.yaml
+++ b/tests/scripts/perf-sanity/deepseek_v32_fp4_blackwell.yaml
@ -0,0 +1,70 @@
+metadata:
+  model_name: deepseek_v32_fp4
+  supported_gpus:
+  - B200
+server_configs:
+  # 8k1k configs - TEP8 with TRTLLM, MTP3
+  - name: "v32_fp4_tep8_mtp3_8k1k"
+    model_name: "deepseek_v32_fp4"
+    tensor_parallel_size: 8
+    moe_expert_parallel_size: 8
+    pipeline_parallel_size: 1
+    max_batch_size: 32
+    max_num_tokens: 12288
+    attn_backend: "TRTLLM"
+    enable_attention_dp: false
+    moe_config:
+      backend: 'TRTLLM'
+    cuda_graph_config:
+      enable_padding: true
+      max_batch_size: 32
+    kv_cache_config:
+      dtype: 'fp8'
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.8
+    speculative_config:
+      decoding_type: 'MTP'
+      num_nextn_predict_layers: 3
+    client_configs:
+      - name: "con2_iter10_8k1k"
+        concurrency: 2
+        iterations: 10
+        isl: 8192
+        osl: 1024
+        random_range_ratio: 0.2
+        backend: "openai"
+
+  # 8k1k configs - DEP8 with CUTLASS, MTP1
+  - name: "v32_fp4_dep8_mtp1_8k1k"
+    model_name: "deepseek_v32_fp4"
+    tensor_parallel_size: 8
+    moe_expert_parallel_size: 8
+    pipeline_parallel_size: 1
+    max_batch_size: 512
+    max_num_tokens: 12288
+    attn_backend: "TRTLLM"
+    enable_attention_dp: true
+    attention_dp_config:
+      batching_wait_iters: 0
+      enable_balance: true
+      timeout_iters: 60
+    moe_config:
+      backend: 'CUTLASS'
+    cuda_graph_config:
+      enable_padding: true
+      max_batch_size: 512
+    kv_cache_config:
+      dtype: 'fp8'
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.8
+    speculative_config:
+      decoding_type: 'MTP'
+      num_nextn_predict_layers: 1
+    client_configs:
+      - name: "con256_iter10_8k1k"
+        concurrency: 256
+        iterations: 10
+        isl: 8192
+        osl: 1024
+        random_range_ratio: 0.2
+        backend: "openai"
--- a/tests/scripts/perf-sanity/deepseek_v32_fp4_grace_blackwell.yaml
+++ b/tests/scripts/perf-sanity/deepseek_v32_fp4_grace_blackwell.yaml
@ -0,0 +1,136 @@
+metadata:
+  model_name: deepseek_v32_fp4
+  supported_gpus:
+  - GB200
+server_configs:
+  # 1k1k configs - TEP4 with TRTLLM, MTP3
+  - name: "v32_fp4_tep4_mtp3_1k1k"
+    model_name: "deepseek_v32_fp4"
+    tensor_parallel_size: 4
+    moe_expert_parallel_size: 4
+    pipeline_parallel_size: 1
+    max_batch_size: 32
+    max_num_tokens: 8192
+    attn_backend: "TRTLLM"
+    enable_attention_dp: false
+    moe_config:
+      backend: 'TRTLLM'
+    cuda_graph_config:
+      enable_padding: true
+      max_batch_size: 32
+    kv_cache_config:
+      dtype: 'fp8'
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.8
+    speculative_config:
+      decoding_type: 'MTP'
+      num_nextn_predict_layers: 3
+    client_configs:
+      - name: "con2_iter10_1k1k"
+        concurrency: 2
+        iterations: 10
+        isl: 1024
+        osl: 1024
+        random_range_ratio: 0.2
+        backend: "openai"
+
+  # 1k1k configs - DEP4 with CUTLASS, MTP1
+  - name: "v32_fp4_dep4_mtp1_1k1k"
+    model_name: "deepseek_v32_fp4"
+    tensor_parallel_size: 4
+    moe_expert_parallel_size: 4
+    pipeline_parallel_size: 1
+    max_batch_size: 512
+    max_num_tokens: 8192
+    attn_backend: "TRTLLM"
+    enable_attention_dp: true
+    attention_dp_config:
+      batching_wait_iters: 0
+      enable_balance: true
+      timeout_iters: 60
+    moe_config:
+      backend: 'CUTLASS'
+    cuda_graph_config:
+      enable_padding: true
+      max_batch_size: 512
+    kv_cache_config:
+      dtype: 'fp8'
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.8
+    speculative_config:
+      decoding_type: 'MTP'
+      num_nextn_predict_layers: 1
+    client_configs:
+      - name: "con1024_iter10_1k1k"
+        concurrency: 1024
+        iterations: 10
+        isl: 1024
+        osl: 1024
+        random_range_ratio: 0.2
+        backend: "openai"
+
+  # 8k1k configs - TEP4 with TRTLLM, MTP3
+  - name: "v32_fp4_tep4_mtp3_8k1k"
+    model_name: "deepseek_v32_fp4"
+    tensor_parallel_size: 4
+    moe_expert_parallel_size: 4
+    pipeline_parallel_size: 1
+    max_batch_size: 32
+    max_num_tokens: 12288
+    attn_backend: "TRTLLM"
+    enable_attention_dp: false
+    moe_config:
+      backend: 'TRTLLM'
+    cuda_graph_config:
+      enable_padding: true
+      max_batch_size: 32
+    kv_cache_config:
+      dtype: 'fp8'
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.8
+    speculative_config:
+      decoding_type: 'MTP'
+      num_nextn_predict_layers: 3
+    client_configs:
+      - name: "con2_iter10_8k1k"
+        concurrency: 2
+        iterations: 10
+        isl: 8192
+        osl: 1024
+        random_range_ratio: 0.2
+        backend: "openai"
+
+  # 8k1k configs - DEP4 with CUTLASS, MTP1
+  - name: "v32_fp4_dep4_mtp1_8k1k"
+    model_name: "deepseek_v32_fp4"
+    tensor_parallel_size: 4
+    moe_expert_parallel_size: 4
+    pipeline_parallel_size: 1
+    max_batch_size: 512
+    max_num_tokens: 12288
+    attn_backend: "TRTLLM"
+    enable_attention_dp: true
+    attention_dp_config:
+      batching_wait_iters: 0
+      enable_balance: true
+      timeout_iters: 60
+    moe_config:
+      backend: 'CUTLASS'
+    cuda_graph_config:
+      enable_padding: true
+      max_batch_size: 512
+    kv_cache_config:
+      dtype: 'fp8'
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.8
+    speculative_config:
+      decoding_type: 'MTP'
+      num_nextn_predict_layers: 1
+    client_configs:
+      - name: "con256_iter10_8k1k"
+        concurrency: 256
+        iterations: 10
+        isl: 8192
+        osl: 1024
+        random_range_ratio: 0.2
+        backend: "openai"
--- a/tests/scripts/perf-sanity/k2_thinking_fp4_2_nodes_grace_blackwell.yaml
+++ b/tests/scripts/perf-sanity/k2_thinking_fp4_2_nodes_grace_blackwell.yaml
@ -0,0 +1,68 @@
+metadata:
+  model_name: k2_thinking_fp4
+  supported_gpus:
+  - GB200
+hardware:
+  gpus_per_node: 4
+server_configs:
+  # 32k8k configs - TEP8 with TRTLLM
+  - name: "k2_thinking_fp4_tep8_32k8k"
+    model_name: "k2_thinking_fp4"
+    tensor_parallel_size: 8
+    moe_expert_parallel_size: 8
+    pipeline_parallel_size: 1
+    max_batch_size: 32
+    max_num_tokens: 8192
+    attn_backend: "TRTLLM"
+    enable_chunked_prefill: true
+    enable_attention_dp: false
+    moe_config:
+      backend: 'TRTLLM'
+    cuda_graph_config:
+      enable_padding: true
+      max_batch_size: 32
+    kv_cache_config:
+      dtype: 'fp8'
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.8
+    client_configs:
+      - name: "con2_iter10_32k8k"
+        concurrency: 2
+        iterations: 10
+        isl: 32768
+        osl: 8192
+        random_range_ratio: 0.2
+        backend: "openai"
+
+  # 32k8k configs - DEP8 with CUTLASS
+  - name: "k2_thinking_fp4_dep8_32k8k"
+    model_name: "k2_thinking_fp4"
+    tensor_parallel_size: 8
+    moe_expert_parallel_size: 8
+    pipeline_parallel_size: 1
+    max_batch_size: 256
+    max_num_tokens: 8192
+    attn_backend: "TRTLLM"
+    enable_chunked_prefill: true
+    enable_attention_dp: true
+    attention_dp_config:
+      batching_wait_iters: 0
+      enable_balance: true
+      timeout_iters: 60
+    moe_config:
+      backend: 'CUTLASS'
+    cuda_graph_config:
+      enable_padding: true
+      max_batch_size: 256
+    kv_cache_config:
+      dtype: 'fp8'
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.8
+    client_configs:
+      - name: "con128_iter10_32k8k"
+        concurrency: 128
+        iterations: 10
+        isl: 32768
+        osl: 8192
+        random_range_ratio: 0.2
+        backend: "openai"
--- a/tests/scripts/perf-sanity/k2_thinking_fp4_blackwell.yaml
+++ b/tests/scripts/perf-sanity/k2_thinking_fp4_blackwell.yaml
@ -0,0 +1,126 @@
+metadata:
+  model_name: k2_thinking_fp4
+  supported_gpus:
+  - B200
+server_configs:
+  # 8k1k configs - TEP8 with TRTLLM
+  - name: "k2_thinking_fp4_tep8_8k1k"
+    model_name: "k2_thinking_fp4"
+    tensor_parallel_size: 8
+    moe_expert_parallel_size: 8
+    pipeline_parallel_size: 1
+    max_batch_size: 32
+    max_num_tokens: 12288
+    attn_backend: "TRTLLM"
+    enable_attention_dp: false
+    moe_config:
+      backend: 'TRTLLM'
+    cuda_graph_config:
+      enable_padding: true
+      max_batch_size: 32
+    kv_cache_config:
+      dtype: 'fp8'
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.8
+    client_configs:
+      - name: "con2_iter10_8k1k"
+        concurrency: 2
+        iterations: 10
+        isl: 8192
+        osl: 1024
+        random_range_ratio: 0.2
+        backend: "openai"
+
+  # 8k1k configs - DEP8 with CUTLASS
+  - name: "k2_thinking_fp4_dep8_8k1k"
+    model_name: "k2_thinking_fp4"
+    tensor_parallel_size: 8
+    moe_expert_parallel_size: 8
+    pipeline_parallel_size: 1
+    max_batch_size: 512
+    max_num_tokens: 12288
+    attn_backend: "TRTLLM"
+    enable_attention_dp: true
+    attention_dp_config:
+      batching_wait_iters: 0
+      enable_balance: true
+      timeout_iters: 60
+    moe_config:
+      backend: 'CUTLASS'
+    cuda_graph_config:
+      enable_padding: true
+      max_batch_size: 512
+    kv_cache_config:
+      dtype: 'fp8'
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.8
+    client_configs:
+      - name: "con512_iter10_8k1k"
+        concurrency: 512
+        iterations: 10
+        isl: 8192
+        osl: 1024
+        random_range_ratio: 0.2
+        backend: "openai"
+
+  # 32k8k configs - TEP8 with TRTLLM
+  - name: "k2_thinking_fp4_tep8_32k8k"
+    model_name: "k2_thinking_fp4"
+    tensor_parallel_size: 8
+    moe_expert_parallel_size: 8
+    pipeline_parallel_size: 1
+    max_batch_size: 32
+    max_num_tokens: 8192
+    attn_backend: "TRTLLM"
+    enable_chunked_prefill: true
+    enable_attention_dp: false
+    moe_config:
+      backend: 'TRTLLM'
+    cuda_graph_config:
+      enable_padding: true
+      max_batch_size: 32
+    kv_cache_config:
+      dtype: 'fp8'
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.8
+    client_configs:
+      - name: "con2_iter10_32k8k"
+        concurrency: 2
+        iterations: 10
+        isl: 32768
+        osl: 8192
+        random_range_ratio: 0.2
+        backend: "openai"
+
+  # 32k8k configs - DEP8 with CUTLASS
+  - name: "k2_thinking_fp4_dep8_32k8k"
+    model_name: "k2_thinking_fp4"
+    tensor_parallel_size: 8
+    moe_expert_parallel_size: 8
+    pipeline_parallel_size: 1
+    max_batch_size: 256
+    max_num_tokens: 8192
+    attn_backend: "TRTLLM"
+    enable_chunked_prefill: true
+    enable_attention_dp: true
+    attention_dp_config:
+      batching_wait_iters: 0
+      enable_balance: true
+      timeout_iters: 60
+    moe_config:
+      backend: 'CUTLASS'
+    cuda_graph_config:
+      enable_padding: true
+      max_batch_size: 256
+    kv_cache_config:
+      dtype: 'fp8'
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.8
+    client_configs:
+      - name: "con128_iter10_32k8k"
+        concurrency: 128
+        iterations: 10
+        isl: 32768
+        osl: 8192
+        random_range_ratio: 0.2
+        backend: "openai"
--- a/tests/scripts/perf-sanity/k2_thinking_fp4_grace_blackwell.yaml
+++ b/tests/scripts/perf-sanity/k2_thinking_fp4_grace_blackwell.yaml
@ -0,0 +1,64 @@
+metadata:
+  model_name: k2_thinking_fp4
+  supported_gpus:
+  - GB200
+server_configs:
+  # 8k1k configs - TEP4 with TRTLLM
+  - name: "k2_thinking_fp4_tep4_8k1k"
+    model_name: "k2_thinking_fp4"
+    tensor_parallel_size: 4
+    moe_expert_parallel_size: 4
+    pipeline_parallel_size: 1
+    max_batch_size: 32
+    max_num_tokens: 12288
+    attn_backend: "TRTLLM"
+    enable_attention_dp: false
+    moe_config:
+      backend: 'TRTLLM'
+    cuda_graph_config:
+      enable_padding: true
+      max_batch_size: 32
+    kv_cache_config:
+      dtype: 'fp8'
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.8
+    client_configs:
+      - name: "con2_iter10_8k1k"
+        concurrency: 2
+        iterations: 10
+        isl: 8192
+        osl: 1024
+        random_range_ratio: 0.2
+        backend: "openai"
+
+  # 8k1k configs - DEP4 with CUTLASS (hang)
+  - name: "k2_thinking_fp4_dep4_8k1k"
+    model_name: "k2_thinking_fp4"
+    tensor_parallel_size: 4
+    moe_expert_parallel_size: 4
+    pipeline_parallel_size: 1
+    max_batch_size: 512
+    max_num_tokens: 12288
+    attn_backend: "TRTLLM"
+    enable_attention_dp: true
+    attention_dp_config:
+      batching_wait_iters: 0
+      enable_balance: true
+      timeout_iters: 60
+    moe_config:
+      backend: 'CUTLASS'
+    cuda_graph_config:
+      enable_padding: true
+      max_batch_size: 512
+    kv_cache_config:
+      dtype: 'fp8'
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.8
+    client_configs:
+      - name: "con256_iter10_8k1k"
+        concurrency: 256
+        iterations: 10
+        isl: 8192
+        osl: 1024
+        random_range_ratio: 0.2
+        backend: "openai"