[TRTLLM-9965][test] add long-context disagg test for GB300/GB200 and remove config_index in yaml (#10225)

Signed-off-by: Ruodi Lu <ruodil@users.noreply.github.com>
Co-authored-by: Ruodi Lu <ruodil@users.noreply.github.com>

tests/integration/defs/perf/disagg/test_configs/disagg/perf/Qwen3-235B-A22B-FP4_1k1k_ctx1_gen1_dep16_bs64_eplb0_mtp3_ccb-NIXL.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 17
 slurm:
   script_file: disaggr_torch.slurm
   partition: <partition>

tests/integration/defs/perf/disagg/test_configs/disagg/perf/Qwen3-235B-A22B-FP4_1k1k_ctx1_gen1_dep16_bs64_eplb0_mtp3_ccb-UCX.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 21
 slurm:
   script_file: disaggr_torch.slurm
   partition: <partition>

tests/integration/defs/perf/disagg/test_configs/disagg/perf/Qwen3-235B-A22B-FP4_1k1k_ctx1_gen1_dep32_bs16_eplb0_mtp3_ccb-NIXL.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 19
 slurm:
   script_file: disaggr_torch.slurm
   partition: <partition>

tests/integration/defs/perf/disagg/test_configs/disagg/perf/Qwen3-235B-A22B-FP4_1k1k_ctx1_gen1_dep32_bs16_eplb0_mtp3_ccb-UCX.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 23
 slurm:
   script_file: disaggr_torch.slurm
   partition: <partition>

tests/integration/defs/perf/disagg/test_configs/disagg/perf/Qwen3-235B-A22B-FP4_1k1k_ctx1_gen4_tep8_bs32_eplb0_mtp0_ccb-NIXL.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 16
 slurm:
   script_file: disaggr_torch.slurm
   partition: <partition>

tests/integration/defs/perf/disagg/test_configs/disagg/perf/Qwen3-235B-A22B-FP4_1k1k_ctx1_gen4_tep8_bs32_eplb0_mtp0_ccb-UCX.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 20
 slurm:
   script_file: disaggr_torch.slurm
   partition: <partition>

tests/integration/defs/perf/disagg/test_configs/disagg/perf/Qwen3-235B-A22B-FP4_1k1k_ctx2_gen1_dep16_bs128_eplb0_mtp1_ccb-NIXL.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 18
 slurm:
   script_file: disaggr_torch.slurm
   partition: <partition>

tests/integration/defs/perf/disagg/test_configs/disagg/perf/Qwen3-235B-A22B-FP4_1k1k_ctx2_gen1_dep16_bs128_eplb0_mtp1_ccb-UCX.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 22
 slurm:
   script_file: disaggr_torch.slurm
   partition: <partition>

tests/integration/defs/perf/disagg/test_configs/disagg/perf/Qwen3-235B-A22B-FP8_1k1k_ctx1_gen1_tep8_bs32_eplb0_mtp0_ccb-NIXL.yaml

@@ -8,7 +8,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 21
 slurm:
   script_file: disaggr_torch.slurm
   partition: <partition>

tests/integration/defs/perf/disagg/test_configs/disagg/perf/Qwen3-235B-A22B-FP8_1k1k_ctx1_gen1_tep8_bs32_eplb0_mtp0_ccb-UCX.yaml

@@ -8,7 +8,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 21
 slurm:
   script_file: disaggr_torch.slurm
   partition: <partition>

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx1_pp4_gen13_tep4_bs1_eplb0_mtp0-Default.yaml

@@ -0,0 +1,99 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB300
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '1'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 1
+  num_gen_servers: 13
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 4
+    moe_expert_parallel_size: 4
+    enable_attention_dp: false
+    pipeline_parallel_size: 1
+    max_batch_size: 1
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+    allreduce_strategy: MNNVL
+  ctx:
+    max_batch_size: 1
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 4
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx1_pp4_gen5_tep4_bs4_eplb0_mtp0-Default.yaml

@@ -0,0 +1,99 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB300
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '4'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 1
+  num_gen_servers: 5
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 4
+    moe_expert_parallel_size: 4
+    enable_attention_dp: false
+    pipeline_parallel_size: 1
+    max_batch_size: 4
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+    allreduce_strategy: MNNVL
+  ctx:
+    max_batch_size: 4
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 4
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx1_pp4_gen6_tep8_bs1_eplb0_mtp3-Default.yaml

@@ -0,0 +1,105 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB300
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '1'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 1
+  num_gen_servers: 6
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 8
+    moe_expert_parallel_size: 8
+    enable_attention_dp: false
+    pipeline_parallel_size: 4
+    max_batch_size: 1
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+    speculative_config:
+      decoding_type: MTP
+      num_nextn_predict_layers: 3
+    allreduce_strategy: MNNVL
+  ctx:
+    max_batch_size: 1
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 4
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    speculative_config:
+      decoding_type: MTP
+      num_nextn_predict_layers: 3
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx1_pp4_gen7_tep8_bs1_eplb0_mtp0-Default.yaml

@@ -0,0 +1,99 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB300
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '1'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 1
+  num_gen_servers: 7
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 8
+    moe_expert_parallel_size: 8
+    enable_attention_dp: false
+    pipeline_parallel_size: 1
+    max_batch_size: 1
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+    allreduce_strategy: MNNVL
+  ctx:
+    max_batch_size: 1
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 4
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx1_pp4_gen8_tep4_bs2_eplb0_mtp0-Default.yaml

@@ -0,0 +1,99 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB300
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '2'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 1
+  num_gen_servers: 8
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: false
+    pipeline_parallel_size: 4
+    max_batch_size: 2
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+    allreduce_strategy: MNNVL
+  ctx:
+    max_batch_size: 2
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 4
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx1_pp4_gen8_tep8_bs1_eplb0_mtp0-Default.yaml

@@ -0,0 +1,99 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB300
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '1'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 1
+  num_gen_servers: 8
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 8
+    moe_expert_parallel_size: 8
+    enable_attention_dp: false
+    pipeline_parallel_size: 1
+    max_batch_size: 1
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+    allreduce_strategy: MNNVL
+  ctx:
+    max_batch_size: 1
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 4
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx1_pp8_gen11_tep4_bs2_eplb0_mtp0-Default.yaml

@@ -0,0 +1,99 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB200
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '2'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 1
+  num_gen_servers: 11
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 4
+    moe_expert_parallel_size: 4
+    enable_attention_dp: false
+    pipeline_parallel_size: 1
+    max_batch_size: 2
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+    allreduce_strategy: MNNVL
+  ctx:
+    max_batch_size: 2
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 8
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx1_pp8_gen14_tep4_bs1_eplb0_mtp0-Default.yaml

@@ -0,0 +1,99 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB200
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '1'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 1
+  num_gen_servers: 14
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 4
+    moe_expert_parallel_size: 4
+    enable_attention_dp: false
+    pipeline_parallel_size: 1
+    max_batch_size: 1
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+    allreduce_strategy: MNNVL
+  ctx:
+    max_batch_size: 1
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 8
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx1_pp8_gen1_dep16_bs1_eplb0_mtp3-Default.yaml

@@ -0,0 +1,102 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB200
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '1'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 1
+  num_gen_servers: 1
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 16
+    moe_expert_parallel_size: 16
+    enable_attention_dp: true
+    pipeline_parallel_size: 1
+    max_batch_size: 1
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+    speculative_config: &id001
+      decoding_type: MTP
+      num_nextn_predict_layers: 3
+  ctx:
+    max_batch_size: 1
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 8
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    speculative_config: *id001
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx1_pp8_gen1_dep8_bs4_eplb0_mtp2-Default.yaml

@@ -0,0 +1,102 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB200
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '4'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 1
+  num_gen_servers: 1
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 8
+    moe_expert_parallel_size: 8
+    enable_attention_dp: true
+    pipeline_parallel_size: 1
+    max_batch_size: 4
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+    speculative_config: &id001
+      decoding_type: MTP
+      num_nextn_predict_layers: 2
+  ctx:
+    max_batch_size: 4
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 8
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    speculative_config: *id001
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx1_pp8_gen1_tep8_bs1_eplb0_mtp0-Default.yaml

@@ -0,0 +1,99 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB200
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '1'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 1
+  num_gen_servers: 1
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 8
+    moe_expert_parallel_size: 8
+    enable_attention_dp: false
+    pipeline_parallel_size: 1
+    max_batch_size: 1
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+    allreduce_strategy: MNNVL
+  ctx:
+    max_batch_size: 1
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 8
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx1_pp8_gen1_tep8_bs1_eplb0_mtp3-Default.yaml

@@ -0,0 +1,103 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB200
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '1'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 1
+  num_gen_servers: 1
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 8
+    moe_expert_parallel_size: 8
+    enable_attention_dp: false
+    pipeline_parallel_size: 1
+    max_batch_size: 1
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+    allreduce_strategy: MNNVL
+    speculative_config: &id001
+      decoding_type: MTP
+      num_nextn_predict_layers: 3
+  ctx:
+    max_batch_size: 1
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 8
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    speculative_config: *id001
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx1_pp8_gen1_tep8_bs2_eplb0_mtp3-Default.yaml

@@ -0,0 +1,103 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB200
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '2'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 1
+  num_gen_servers: 1
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 8
+    moe_expert_parallel_size: 8
+    enable_attention_dp: false
+    pipeline_parallel_size: 1
+    max_batch_size: 2
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+    allreduce_strategy: MNNVL
+    speculative_config: &id001
+      decoding_type: MTP
+      num_nextn_predict_layers: 3
+  ctx:
+    max_batch_size: 2
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 8
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    speculative_config: *id001
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx1_pp8_gen5_tep8_bs2_eplb0_mtp3-Default.yaml

@@ -0,0 +1,103 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB200
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '2'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 1
+  num_gen_servers: 5
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 8
+    moe_expert_parallel_size: 8
+    enable_attention_dp: false
+    pipeline_parallel_size: 1
+    max_batch_size: 2
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+    allreduce_strategy: MNNVL
+    speculative_config: &id001
+      decoding_type: MTP
+      num_nextn_predict_layers: 3
+  ctx:
+    max_batch_size: 2
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 8
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    speculative_config: *id001
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx1_pp8_gen7_tep4_bs2_eplb0_mtp2-Default.yaml

@@ -0,0 +1,103 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB200
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '2'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 1
+  num_gen_servers: 7
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 4
+    moe_expert_parallel_size: 4
+    enable_attention_dp: false
+    pipeline_parallel_size: 1
+    max_batch_size: 2
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+    allreduce_strategy: MNNVL
+    speculative_config: &id001
+      decoding_type: MTP
+      num_nextn_predict_layers: 2
+  ctx:
+    max_batch_size: 2
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 8
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    speculative_config: *id001
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx1_pp8_gen7_tep8_bs1_eplb0_mtp0-Default.yaml

@@ -0,0 +1,99 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB200
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '1'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 1
+  num_gen_servers: 7
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 8
+    moe_expert_parallel_size: 8
+    enable_attention_dp: false
+    pipeline_parallel_size: 1
+    max_batch_size: 1
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+    allreduce_strategy: MNNVL
+  ctx:
+    max_batch_size: 1
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 8
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx1_pp8_gen8_tep4_bs4_eplb0_mtp0-Default.yaml

@@ -0,0 +1,99 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB200
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '4'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 1
+  num_gen_servers: 8
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 4
+    moe_expert_parallel_size: 4
+    enable_attention_dp: false
+    pipeline_parallel_size: 1
+    max_batch_size: 4
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+    allreduce_strategy: MNNVL
+  ctx:
+    max_batch_size: 4
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 8
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx2_pp4_gen7_tep8_bs2_eplb0_mtp3-Default.yaml

@@ -0,0 +1,105 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB300
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '2'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 2
+  num_gen_servers: 7
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 8
+    moe_expert_parallel_size: 8
+    enable_attention_dp: false
+    pipeline_parallel_size: 1
+    max_batch_size: 2
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+    speculative_config:
+      decoding_type: MTP
+      num_nextn_predict_layers: 3
+    allreduce_strategy: MNNVL
+  ctx:
+    max_batch_size: 2
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 4
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    speculative_config:
+      decoding_type: MTP
+      num_nextn_predict_layers: 3
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx2_pp8_gen1_dep16_bs8_eplb0_mtp0-Default.yaml

@@ -0,0 +1,98 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB200
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '8'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 2
+  num_gen_servers: 1
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 16
+    moe_expert_parallel_size: 16
+    enable_attention_dp: true
+    pipeline_parallel_size: 1
+    max_batch_size: 8
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+  ctx:
+    max_batch_size: 8
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 8
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx2_pp8_gen1_dep32_bs2_eplb0_mtp0-Default.yaml

@@ -0,0 +1,98 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB200
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '2'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 2
+  num_gen_servers: 1
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 32
+    moe_expert_parallel_size: 32
+    enable_attention_dp: true
+    pipeline_parallel_size: 1
+    max_batch_size: 2
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+  ctx:
+    max_batch_size: 2
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 8
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx3_pp4_gen1_dep8_bs16_eplb0_mtp1-Default.yaml

@@ -0,0 +1,104 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB300
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '128'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 3
+  num_gen_servers: 1
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 8
+    moe_expert_parallel_size: 8
+    enable_attention_dp: true
+    pipeline_parallel_size: 1
+    max_batch_size: 16
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+    speculative_config:
+      decoding_type: MTP
+      num_nextn_predict_layers: 1
+  ctx:
+    max_batch_size: 16
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 4
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    speculative_config:
+      decoding_type: MTP
+      num_nextn_predict_layers: 1
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx3_pp8_gen1_dep16_bs16_eplb0_mtp0-Default.yaml

@@ -0,0 +1,98 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB200
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '16'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 3
+  num_gen_servers: 1
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 16
+    moe_expert_parallel_size: 16
+    enable_attention_dp: true
+    pipeline_parallel_size: 1
+    max_batch_size: 16
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+  ctx:
+    max_batch_size: 16
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 8
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx3_pp8_gen1_dep16_bs8_eplb0_mtp2-Default.yaml

@@ -0,0 +1,102 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB200
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '8'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 3
+  num_gen_servers: 1
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 16
+    moe_expert_parallel_size: 16
+    enable_attention_dp: true
+    pipeline_parallel_size: 1
+    max_batch_size: 8
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+    speculative_config: &id001
+      decoding_type: MTP
+      num_nextn_predict_layers: 2
+  ctx:
+    max_batch_size: 8
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 8
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    speculative_config: *id001
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx3_pp8_gen1_dep32_bs2_eplb0_mtp3-Default.yaml

@@ -0,0 +1,102 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB200
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '2'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 3
+  num_gen_servers: 1
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 32
+    moe_expert_parallel_size: 32
+    enable_attention_dp: true
+    pipeline_parallel_size: 1
+    max_batch_size: 2
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+    speculative_config: &id001
+      decoding_type: MTP
+      num_nextn_predict_layers: 3
+  ctx:
+    max_batch_size: 2
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 8
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    speculative_config: *id001
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx3_pp8_gen1_dep32_bs4_eplb0_mtp0-Default.yaml

@@ -0,0 +1,98 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB200
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '4'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 3
+  num_gen_servers: 1
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 32
+    moe_expert_parallel_size: 32
+    enable_attention_dp: true
+    pipeline_parallel_size: 1
+    max_batch_size: 4
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+  ctx:
+    max_batch_size: 4
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 8
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx5_pp4_gen1_dep16_bs16_eplb0_mtp0-Default.yaml

@@ -0,0 +1,98 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB300
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '256'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 5
+  num_gen_servers: 1
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 16
+    moe_expert_parallel_size: 16
+    enable_attention_dp: true
+    pipeline_parallel_size: 1
+    max_batch_size: 16
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+  ctx:
+    max_batch_size: 16
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 4
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx5_pp4_gen1_dep16_bs8_eplb0_mtp3-Default.yaml

@@ -0,0 +1,104 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB300
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '128'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 5
+  num_gen_servers: 1
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 16
+    moe_expert_parallel_size: 16
+    enable_attention_dp: true
+    pipeline_parallel_size: 1
+    max_batch_size: 8
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+    speculative_config:
+      decoding_type: MTP
+      num_nextn_predict_layers: 3
+  ctx:
+    max_batch_size: 8
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 4
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    speculative_config:
+      decoding_type: MTP
+      num_nextn_predict_layers: 3
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx5_pp4_gen1_dep32_bs2_eplb0_mtp3-Default.yaml

@@ -0,0 +1,104 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB300
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '64'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 5
+  num_gen_servers: 1
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 32
+    moe_expert_parallel_size: 32
+    enable_attention_dp: true
+    pipeline_parallel_size: 1
+    max_batch_size: 2
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+    speculative_config:
+      decoding_type: MTP
+      num_nextn_predict_layers: 3
+  ctx:
+    max_batch_size: 2
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 4
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    speculative_config:
+      decoding_type: MTP
+      num_nextn_predict_layers: 3
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx5_pp4_gen1_dep32_bs4_eplb0_mtp0-Default.yaml

@@ -0,0 +1,98 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB300
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '128'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 5
+  num_gen_servers: 1
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 32
+    moe_expert_parallel_size: 32
+    enable_attention_dp: true
+    pipeline_parallel_size: 1
+    max_batch_size: 4
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+  ctx:
+    max_batch_size: 4
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 4
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx7_pp4_gen1_dep16_bs16_eplb0_mtp1-Default.yaml

@@ -0,0 +1,104 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB300
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '256'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 7
+  num_gen_servers: 1
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 16
+    moe_expert_parallel_size: 16
+    enable_attention_dp: true
+    pipeline_parallel_size: 1
+    max_batch_size: 16
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+    speculative_config:
+      decoding_type: MTP
+      num_nextn_predict_layers: 1
+  ctx:
+    max_batch_size: 16
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 4
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    speculative_config:
+      decoding_type: MTP
+      num_nextn_predict_layers: 1
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx7_pp4_gen1_dep16_bs32_eplb0_mtp0-Default.yaml

@@ -0,0 +1,98 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB300
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '512'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 7
+  num_gen_servers: 1
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 16
+    moe_expert_parallel_size: 16
+    enable_attention_dp: true
+    pipeline_parallel_size: 1
+    max_batch_size: 32
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+  ctx:
+    max_batch_size: 32
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 4
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx8_pp4_gen1_dep16_bs32_eplb0_mtp1-Default.yaml

@@ -0,0 +1,104 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB300
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '512'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 8
+  num_gen_servers: 1
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 16
+    moe_expert_parallel_size: 16
+    enable_attention_dp: true
+    pipeline_parallel_size: 1
+    max_batch_size: 32
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+    speculative_config:
+      decoding_type: MTP
+      num_nextn_predict_layers: 1
+  ctx:
+    max_batch_size: 32
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 4
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    speculative_config:
+      decoding_type: MTP
+      num_nextn_predict_layers: 1
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx8_pp4_gen1_dep32_bs4_eplb0_mtp3-Default.yaml

@@ -0,0 +1,104 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB300
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '128'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 8
+  num_gen_servers: 1
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 32
+    moe_expert_parallel_size: 32
+    enable_attention_dp: true
+    pipeline_parallel_size: 1
+    max_batch_size: 4
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+    speculative_config:
+      decoding_type: MTP
+      num_nextn_predict_layers: 3
+  ctx:
+    max_batch_size: 4
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 4
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    speculative_config:
+      decoding_type: MTP
+      num_nextn_predict_layers: 3
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx8_pp4_gen1_dep32_bs8_eplb0_mtp0-Default.yaml

@@ -0,0 +1,98 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB300
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '256'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 8
+  num_gen_servers: 1
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 32
+    moe_expert_parallel_size: 32
+    enable_attention_dp: true
+    pipeline_parallel_size: 1
+    max_batch_size: 8
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+  ctx:
+    max_batch_size: 8
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 4
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_128k8k_ctx8_pp4_gen1_dep32_bs8_eplb0_mtp3-Default.yaml

@@ -0,0 +1,104 @@
+metadata:
+  model_name: deepseek-r1-fp4
+  precision: fp4
+  model_dir_name: DeepSeek-R1-0528-FP4-v2
+  supported_gpus:
+  - GB300
+  script_file: disaggr_torch.slurm
+  benchmark_type: 128k8k
+slurm:
+  script_file: disaggr_torch.slurm
+  partition: <partition>
+  account: <account>
+  job_time: 02:00:00
+  job_name: unified-benchmark
+  extra_args: --gres=gpu:4
+  numa_bind: true
+benchmark:
+  mode: e2e
+  use_nv_sa_benchmark: true
+  multi_round: 1
+  benchmark_ratio: 0.8
+  streaming: true
+  concurrency_list: '256'
+  input_length: 131072
+  output_length: 8192
+  dataset_file: <dataset_file>
+hardware:
+  gpus_per_node: 4
+  num_ctx_servers: 8
+  num_gen_servers: 1
+environment:
+  container_mount: <container_mount>
+  container_image: <container_image>
+  model_path: <model_path>
+  trtllm_repo: ''
+  build_wheel: false
+  work_dir: <full_path_to_work_dir>
+  worker_env_var: TLLM_LOG_LEVEL=INFO TRTLLM_SERVER_DISABLE_GC=1 TRTLLM_WORKER_DISABLE_GC=1
+    TRTLLM_ENABLE_PDL=1 ENROOT_ALLOW_DEV=yes
+  server_env_var: TRTLLM_SERVER_DISABLE_GC=1
+profiling:
+  nsys_on: false
+accuracy:
+  enable_accuracy_test: false
+worker_config:
+  gen:
+    tensor_parallel_size: 32
+    moe_expert_parallel_size: 32
+    enable_attention_dp: true
+    pipeline_parallel_size: 1
+    max_batch_size: 8
+    max_num_tokens: 128
+    max_seq_len: 139296
+    cuda_graph_config:
+      enable_padding: true
+      batch_sizes:
+      - 1
+      - 2
+      - 4
+      - 8
+      - 16
+      - 32
+      - 64
+      - 128
+      - 256
+      - 512
+    print_iter_log: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    moe_config:
+      backend: TRTLLM
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    stream_interval: 20
+    num_postprocess_workers: 4
+    speculative_config:
+      decoding_type: MTP
+      num_nextn_predict_layers: 3
+  ctx:
+    max_batch_size: 8
+    max_num_tokens: 131104
+    max_seq_len: 131104
+    tensor_parallel_size: 1
+    moe_expert_parallel_size: 1
+    enable_attention_dp: true
+    pipeline_parallel_size: 4
+    print_iter_log: true
+    cuda_graph_config: null
+    disable_overlap_scheduler: true
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.4
+      dtype: fp8
+    cache_transceiver_config:
+      max_tokens_in_buffer: 131104
+      backend: DEFAULT
+    speculative_config:
+      decoding_type: MTP
+      num_nextn_predict_layers: 3
+    moe_config:
+      backend: TRTLLM

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_1k1k_ctx1_gen1_dep32_bs32_eplb0_mtp0_ccb-NIXL.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 3
 slurm:
   script_file: disaggr_torch.slurm
   partition: <partition>

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_1k1k_ctx1_gen1_dep32_bs32_eplb0_mtp0_ccb-UCX.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 7
 slurm:
   script_file: disaggr_torch.slurm
   partition: <partition>

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_1k1k_ctx1_gen4_tep8_bs32_eplb0_mtp0_ccb-NIXL.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 1
 slurm:
   script_file: disaggr_torch.slurm
   partition: <partition>

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_1k1k_ctx1_gen4_tep8_bs32_eplb0_mtp0_ccb-UCX.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 5
 slurm:
   script_file: disaggr_torch.slurm
   partition: <partition>

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_1k1k_ctx1_gen4_tep8_bs32_eplb0_mtp3_ccb-NIXL.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 0
 slurm:
   script_file: disaggr_torch.slurm
   partition: <partition>

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_1k1k_ctx1_gen4_tep8_bs32_eplb0_mtp3_ccb-UCX.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 4
 slurm:
   script_file: disaggr_torch.slurm
   partition: <partition>

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_1k1k_ctx2_gen1_dep16_bs128_eplb0_mtp3_ccb-NIXL.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 2
 slurm:
   script_file: disaggr_torch.slurm
   partition: <partition>

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_1k1k_ctx2_gen1_dep16_bs128_eplb0_mtp3_ccb-UCX.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 6
 slurm:
   script_file: disaggr_torch.slurm
   partition: <partition>

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_8k1k_ctx1_gen3_tep8_bs16_eplb0_mtp3_ccb-NIXL.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 8k1k
-  config_index: 8
 slurm:
   script_file: disaggr_torch.slurm
   partition: <partition>

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_8k1k_ctx1_gen3_tep8_bs16_eplb0_mtp3_ccb-UCX.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 8k1k
-  config_index: 12
 slurm:
   script_file: disaggr_torch.slurm
   partition: <partition>

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_8k1k_ctx1_gen3_tep8_bs32_eplb0_mtp0_ccb-NIXL.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 8k1k
-  config_index: 9
 slurm:
   script_file: disaggr_torch.slurm
   partition: <partition>

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_8k1k_ctx1_gen3_tep8_bs32_eplb0_mtp0_ccb-UCX.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 8k1k
-  config_index: 13
 slurm:
   script_file: disaggr_torch.slurm
   partition: <partition>

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_8k1k_ctx6_gen1_dep16_bs64_eplb0_mtp0_ccb-NIXL.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 8k1k
-  config_index: 11
 slurm:
   script_file: disaggr_torch.slurm
   partition: <partition>

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_8k1k_ctx6_gen1_dep16_bs64_eplb0_mtp0_ccb-UCX.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 8k1k
-  config_index: 15
 slurm:
   script_file: disaggr_torch.slurm
   partition: <partition>

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_8k1k_ctx8_gen1_dep32_bs16_eplb0_mtp3_ccb-NIXL.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 8k1k
-  config_index: 10
 slurm:
   script_file: disaggr_torch.slurm
   partition: <partition>

tests/integration/defs/perf/disagg/test_configs/disagg/perf/deepseek-r1-fp4_8k1k_ctx8_gen1_dep32_bs16_eplb0_mtp3_ccb-UCX.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 8k1k
-  config_index: 14
 slurm:
   script_file: disaggr_torch.slurm
   partition: <partition>

tests/integration/defs/perf/disagg/test_configs/wideep/accuracy/deepseek-r1-fp4_1k1k_ctx2_gen1_dep16_bs128_eplb288_mtp3_ccb-NIXL.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 0
   dataset_file: disagg_datasets/deepseek-r1-1024-1024-100000-ratio-1_for_serve.json
   accuracy:
     datasets:

tests/integration/defs/perf/disagg/test_configs/wideep/accuracy/kimi-k2-thinking-fp4_1k1k_ctx3_gen1_dep32_bs1024_eplb384_mtp0_ccb-NIXL.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 6
   dataset_file: disagg_datasets/kimi-k2-1024-1024-100000-ratio-1_for_serve.json
   accuracy:
     datasets:

tests/integration/defs/perf/disagg/test_configs/wideep/perf/Qwen3-235B-A22B-FP4_1k1k_ctx1_gen1_dep16_bs64_eplb288_mtp3_ccb-NIXL.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 8
   dataset_file: disagg_datasets/deepseek-r1-1024-1024-100000-ratio-1_for_serve.json
 slurm:
   script_file: disaggr_torch.slurm

tests/integration/defs/perf/disagg/test_configs/wideep/perf/Qwen3-235B-A22B-FP4_1k1k_ctx1_gen1_dep16_bs64_eplb288_mtp3_ccb-UCX.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 11
   dataset_file: disagg_datasets/deepseek-r1-1024-1024-100000-ratio-1_for_serve.json
 slurm:
   script_file: disaggr_torch.slurm

tests/integration/defs/perf/disagg/test_configs/wideep/perf/Qwen3-235B-A22B-FP4_1k1k_ctx1_gen1_dep32_bs16_eplb288_mtp3_ccb-NIXL.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 10
   dataset_file: disagg_datasets/deepseek-r1-1024-1024-100000-ratio-1_for_serve.json
 slurm:
   script_file: disaggr_torch.slurm

tests/integration/defs/perf/disagg/test_configs/wideep/perf/Qwen3-235B-A22B-FP4_1k1k_ctx1_gen1_dep32_bs16_eplb288_mtp3_ccb-UCX.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 13
   dataset_file: disagg_datasets/deepseek-r1-1024-1024-100000-ratio-1_for_serve.json
 slurm:
   script_file: disaggr_torch.slurm

tests/integration/defs/perf/disagg/test_configs/wideep/perf/Qwen3-235B-A22B-FP4_1k1k_ctx2_gen1_dep16_bs128_eplb288_mtp1_ccb-NIXL.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 9
   dataset_file: disagg_datasets/deepseek-r1-1024-1024-100000-ratio-1_for_serve.json
 slurm:
   script_file: disaggr_torch.slurm

tests/integration/defs/perf/disagg/test_configs/wideep/perf/Qwen3-235B-A22B-FP4_1k1k_ctx2_gen1_dep16_bs128_eplb288_mtp1_ccb-UCX.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 12
   dataset_file: disagg_datasets/deepseek-r1-1024-1024-100000-ratio-1_for_serve.json
 slurm:
   script_file: disaggr_torch.slurm

tests/integration/defs/perf/disagg/test_configs/wideep/perf/deepseek-r1-fp4_1k1k_ctx1_gen1_dep32_bs32_eplb288_mtp0_ccb-NIXL.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 1
   dataset_file: disagg_datasets/deepseek-r1-1024-1024-100000-ratio-1_for_serve.json
 slurm:
   script_file: disaggr_torch.slurm

tests/integration/defs/perf/disagg/test_configs/wideep/perf/deepseek-r1-fp4_1k1k_ctx1_gen1_dep32_bs32_eplb288_mtp0_ccb-UCX.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 3
   dataset_file: disagg_datasets/deepseek-r1-1024-1024-100000-ratio-1_for_serve.json
 slurm:
   script_file: disaggr_torch.slurm

tests/integration/defs/perf/disagg/test_configs/wideep/perf/deepseek-r1-fp4_1k1k_ctx2_gen1_dep16_bs128_eplb288_mtp3_ccb-NIXL.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 0
   dataset_file: disagg_datasets/deepseek-r1-1024-1024-100000-ratio-1_for_serve.json
 slurm:
   script_file: disaggr_torch.slurm

tests/integration/defs/perf/disagg/test_configs/wideep/perf/deepseek-r1-fp4_1k1k_ctx2_gen1_dep16_bs128_eplb288_mtp3_ccb-UCX.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 2
   dataset_file: disagg_datasets/deepseek-r1-1024-1024-100000-ratio-1_for_serve.json
 slurm:
   script_file: disaggr_torch.slurm

tests/integration/defs/perf/disagg/test_configs/wideep/perf/deepseek-r1-fp4_1k1k_ctx2_gen1_dep48_bs16_eplb288_mtp3_ccb-DEFAULT.yaml

@@ -8,7 +8,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 8k1k
-  config_index: 7
   dataset_file: disagg_datasets/deepseek-r1-8192-1024-200000-ratio-1_for_serve.json
 slurm:
   script_file: disaggr_torch.slurm

tests/integration/defs/perf/disagg/test_configs/wideep/perf/deepseek-r1-fp4_8k1k_ctx2_gen1_dep32_bs128_eplb288_mtp3_ccb-DEFAULT.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 8k1k
-  config_index: 14
   dataset_file: disagg_datasets/deepseek-r1-8192-1024-200000-ratio-1_for_serve.json
 slurm:
   script_file: disaggr_torch.slurm

tests/integration/defs/perf/disagg/test_configs/wideep/perf/deepseek-r1-fp4_8k1k_ctx6_gen1_dep16_bs64_eplb288_mtp0_ccb-NIXL.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 8k1k
-  config_index: 5
   dataset_file: disagg_datasets/deepseek-r1-8192-1024-200000-ratio-1_for_serve.json
 slurm:
   script_file: disaggr_torch.slurm

tests/integration/defs/perf/disagg/test_configs/wideep/perf/deepseek-r1-fp4_8k1k_ctx6_gen1_dep16_bs64_eplb288_mtp0_ccb-UCX.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 8k1k
-  config_index: 7
   dataset_file: disagg_datasets/deepseek-r1-8192-1024-200000-ratio-1_for_serve.json
 slurm:
   script_file: disaggr_torch.slurm

tests/integration/defs/perf/disagg/test_configs/wideep/perf/deepseek-r1-fp4_8k1k_ctx8_gen1_dep32_bs16_eplb288_mtp3_ccb-NIXL.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 8k1k
-  config_index: 4
   dataset_file: disagg_datasets/deepseek-r1-8192-1024-200000-ratio-1_for_serve.json
 slurm:
   script_file: disaggr_torch.slurm

tests/integration/defs/perf/disagg/test_configs/wideep/perf/deepseek-r1-fp4_8k1k_ctx8_gen1_dep32_bs16_eplb288_mtp3_ccb-UCX.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 8k1k
-  config_index: 6
   dataset_file: disagg_datasets/deepseek-r1-8192-1024-200000-ratio-1_for_serve.json
 slurm:
   script_file: disaggr_torch.slurm

tests/integration/defs/perf/disagg/test_configs/wideep/perf/kimi-k2-thinking-fp4_1k1k_ctx3_gen1_dep32_bs1024_eplb384_mtp0_ccb-NIXL.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 1k1k
-  config_index: 6
   dataset_file: disagg_datasets/kimi-k2-1024-1024-100000-ratio-1_for_serve.json
 slurm:
   script_file: disaggr_torch.slurm

tests/integration/defs/perf/disagg/test_configs/wideep/perf/kimi-k2-thinking-fp4_8k1k_ctx8_gen1_dep32_bs256_eplb416_mtp0_ccb-NIXL.yaml

@@ -7,7 +7,6 @@ metadata:
   - GB300
   script_file: disaggr_torch.slurm
   benchmark_type: 8k1k
-  config_index: 6
   dataset_file: disagg_datasets/kimi-k2-8192-1024-20000-ratio-1_for_serve.json
 slurm:
   script_file: disaggr_torch.slurm

tests/integration/defs/perf/disagg/testlist/all.txt

@@ -26,6 +26,45 @@ test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_Qwen3-235B-A22B-
 # test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_Qwen3-235B-A22B-FP4_1k1k_ctx1_gen1_dep16_bs64_eplb0_mtp3_ccb-UCX]
 # test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_Qwen3-235B-A22B-FP4_1k1k_ctx1_gen1_dep32_bs16_eplb0_mtp3_ccb-NIXL]
 
+# 128k8k GB300 (pp4) cases
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp4_gen13_tep4_bs1_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp4_gen5_tep4_bs4_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp4_gen6_tep8_bs1_eplb0_mtp3-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp4_gen7_tep8_bs1_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp4_gen8_tep4_bs2_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp4_gen8_tep8_bs1_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx2_pp4_gen7_tep8_bs2_eplb0_mtp3-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx3_pp4_gen1_dep8_bs16_eplb0_mtp1-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx5_pp4_gen1_dep16_bs16_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx5_pp4_gen1_dep16_bs8_eplb0_mtp3-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx5_pp4_gen1_dep32_bs2_eplb0_mtp3-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx5_pp4_gen1_dep32_bs4_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx7_pp4_gen1_dep16_bs16_eplb0_mtp1-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx7_pp4_gen1_dep16_bs32_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx8_pp4_gen1_dep16_bs32_eplb0_mtp1-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx8_pp4_gen1_dep32_bs4_eplb0_mtp3-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx8_pp4_gen1_dep32_bs8_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx8_pp4_gen1_dep32_bs8_eplb0_mtp3-Default]
+
+# 128k8k GB200 (pp8) cases
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp8_gen11_tep4_bs2_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp8_gen14_tep4_bs1_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp8_gen1_dep16_bs1_eplb0_mtp3-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp8_gen1_dep8_bs4_eplb0_mtp2-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp8_gen1_tep8_bs1_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp8_gen1_tep8_bs1_eplb0_mtp3-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp8_gen1_tep8_bs2_eplb0_mtp3-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp8_gen5_tep8_bs2_eplb0_mtp3-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp8_gen7_tep4_bs2_eplb0_mtp2-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp8_gen7_tep8_bs1_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp8_gen8_tep4_bs4_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx2_pp8_gen1_dep16_bs8_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx2_pp8_gen1_dep32_bs2_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx3_pp8_gen1_dep16_bs16_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx3_pp8_gen1_dep16_bs8_eplb0_mtp2-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx3_pp8_gen1_dep32_bs2_eplb0_mtp3-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx3_pp8_gen1_dep32_bs4_eplb0_mtp0-Default]
+
 
 # WIDEEP cases
 test_disagg.py::TestDisaggBenchmark::test_benchmark[wideep_perf_deepseek-r1-fp4_8k1k_ctx2_gen1_dep32_bs128_eplb288_mtp3_ccb-DEFAULT]

tests/integration/defs/perf/disagg/testlist/disagg.txt

@@ -24,3 +24,42 @@ test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_Qwen3-235B-A22B-
 # test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_Qwen3-235B-A22B-FP4_1k1k_ctx1_gen1_dep32_bs16_eplb0_mtp3_ccb-UCX]
 # test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_Qwen3-235B-A22B-FP4_1k1k_ctx1_gen1_dep16_bs64_eplb0_mtp3_ccb-UCX]
 # test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_Qwen3-235B-A22B-FP4_1k1k_ctx1_gen1_dep32_bs16_eplb0_mtp3_ccb-NIXL]
+
+# 128k8k GB300 (pp4) cases
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp4_gen13_tep4_bs1_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp4_gen5_tep4_bs4_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp4_gen6_tep8_bs1_eplb0_mtp3-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp4_gen7_tep8_bs1_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp4_gen8_tep4_bs2_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp4_gen8_tep8_bs1_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx2_pp4_gen7_tep8_bs2_eplb0_mtp3-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx3_pp4_gen1_dep8_bs16_eplb0_mtp1-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx5_pp4_gen1_dep16_bs16_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx5_pp4_gen1_dep16_bs8_eplb0_mtp3-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx5_pp4_gen1_dep32_bs2_eplb0_mtp3-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx5_pp4_gen1_dep32_bs4_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx7_pp4_gen1_dep16_bs16_eplb0_mtp1-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx7_pp4_gen1_dep16_bs32_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx8_pp4_gen1_dep16_bs32_eplb0_mtp1-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx8_pp4_gen1_dep32_bs4_eplb0_mtp3-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx8_pp4_gen1_dep32_bs8_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx8_pp4_gen1_dep32_bs8_eplb0_mtp3-Default]
+
+# 128k8k GB200 (pp8) cases
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp8_gen11_tep4_bs2_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp8_gen14_tep4_bs1_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp8_gen1_dep16_bs1_eplb0_mtp3-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp8_gen1_dep8_bs4_eplb0_mtp2-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp8_gen1_tep8_bs1_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp8_gen1_tep8_bs1_eplb0_mtp3-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp8_gen1_tep8_bs2_eplb0_mtp3-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp8_gen5_tep8_bs2_eplb0_mtp3-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp8_gen7_tep4_bs2_eplb0_mtp2-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp8_gen7_tep8_bs1_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx1_pp8_gen8_tep4_bs4_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx2_pp8_gen1_dep16_bs8_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx2_pp8_gen1_dep32_bs2_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx3_pp8_gen1_dep16_bs16_eplb0_mtp0-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx3_pp8_gen1_dep16_bs8_eplb0_mtp2-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx3_pp8_gen1_dep32_bs2_eplb0_mtp3-Default]
+test_disagg.py::TestDisaggBenchmark::test_benchmark[disagg_perf_deepseek-r1-fp4_128k8k_ctx3_pp8_gen1_dep32_bs4_eplb0_mtp0-Default]