| 1 | group: Disaggregated |
| 2 | depends_on: |
| 3 | - image-build |
| 4 | steps: |
| 5 | - label: ":nvidia: (L4) Distributed NixlConnector PD accuracy" |
| 6 | key: distributed-nixlconnector-pd-accuracy-4-gpus |
| 7 | timeout_in_minutes: 55 |
| 8 | working_dir: "/vllm-workspace/tests" |
| 9 | device: l4 |
| 10 | num_devices: 4 |
| 11 | source_file_dependencies: |
| 12 | - vllm/distributed/kv_transfer/kv_connector/v1/nixl/ |
| 13 | - tests/v1/kv_connector/nixl_integration/ |
| 14 | commands: |
| 15 | - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh |
| 16 | - bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh |
| 17 | mirror: |
| 18 | amd: |
| 19 | label: ":amd: (MI355) Distributed NixlConnector PD accuracy" |
| 20 | dind: false |
| 21 | device: mi355_4 |
| 22 | timeout_in_minutes: 60 |
| 23 | depends_on: |
| 24 | - image-build-amd |
| 25 | source_file_dependencies: |
| 26 | - vllm/distributed/kv_transfer/kv_connector/v1/nixl/ |
| 27 | - tests/v1/kv_connector/nixl_integration/ |
| 28 | - vllm/platforms/rocm.py |
| 29 | commands: |
| 30 | - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt |
| 31 | - ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh |
| 32 | |
| 33 | - label: ":nvidia: (L4) Distributed FlashInfer NixlConnector PD accuracy" |
| 34 | key: distributed-flashinfer-nixlconnector-pd-accuracy-4-gpus |
| 35 | timeout_in_minutes: 55 |
| 36 | working_dir: "/vllm-workspace/tests" |
| 37 | device: l4 |
| 38 | num_devices: 4 |
| 39 | source_file_dependencies: |
| 40 | - vllm/distributed/kv_transfer/kv_connector/v1/nixl/ |
| 41 | - tests/v1/kv_connector/nixl_integration/ |
| 42 | commands: |
| 43 | - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh |
| 44 | - FLASHINFER=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh |
| 45 | mirror: |
| 46 | amd: |
| 47 | label: ":amd: (MI355) Distributed AITER NixlConnector PD accuracy" |
| 48 | dind: false |
| 49 | device: mi355_4 |
| 50 | num_devices: 4 |
| 51 | timeout_in_minutes: 60 |
| 52 | working_dir: "/vllm-workspace/tests" |
| 53 | depends_on: |
| 54 | - image-build-amd |
| 55 | source_file_dependencies: |
| 56 | - vllm/distributed/kv_transfer/kv_connector/v1/nixl/ |
| 57 | - tests/v1/kv_connector/nixl_integration/ |
| 58 | - tests/v1/kv_connector/rocm_pd_accuracy_utils.py |
| 59 | - tests/utils.py |
| 60 | - vllm/v1/attention/backends/rocm_aiter_unified_attn.py |
| 61 | - vllm/platforms/rocm.py |
| 62 | - requirements/common.txt |
| 63 | - requirements/kv_connectors_rocm.txt |
| 64 | commands: |
| 65 | - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt |
| 66 | # Preserve all six model/TP pairs with native AITER attention. |
| 67 | - pytest -v -s v1/kv_connector/nixl_integration/test_accuracy_rocm.py |
| 68 | |
| 69 | - label: ":nvidia: (L4) Push NixlConnector PP prefill PD accuracy" |
| 70 | key: push-nixlconnector-pp-prefill-pd-accuracy-4-gpus |
| 71 | timeout_in_minutes: 30 |
| 72 | working_dir: "/vllm-workspace/tests" |
| 73 | device: l4 |
| 74 | num_devices: 4 |
| 75 | source_file_dependencies: |
| 76 | - vllm/distributed/kv_transfer/kv_connector/v1/nixl/ |
| 77 | - tests/v1/kv_connector/nixl_integration/ |
| 78 | - tests/v1/kv_connector/nixl_push_integration/ |
| 79 | commands: |
| 80 | - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh |
| 81 | - bash v1/kv_connector/nixl_push_integration/config_sweep_accuracy_test.sh |
| 82 | mirror: |
| 83 | amd: |
| 84 | label: ":amd: (MI300) Push NixlConnector PP prefill PD accuracy" |
| 85 | dind: false |
| 86 | device: mi300_4 |
| 87 | timeout_in_minutes: 55 |
| 88 | depends_on: |
| 89 | - image-build-amd |
| 90 | source_file_dependencies: |
| 91 | - vllm/distributed/kv_transfer/kv_connector/v1/nixl/ |
| 92 | - tests/v1/kv_connector/nixl_integration/ |
| 93 | - tests/v1/kv_connector/nixl_push_integration/ |
| 94 | - vllm/platforms/rocm.py |
| 95 | commands: |
| 96 | - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt |
| 97 | - ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_push_integration/config_sweep_accuracy_test.sh |
| 98 | |
| 99 | - label: ":nvidia: (L4) DP EP Distributed NixlConnector PD accuracy" |
| 100 | key: dp-ep-distributed-nixlconnector-pd-accuracy-tests-4-gpus |
| 101 | timeout_in_minutes: 30 |
| 102 | working_dir: "/vllm-workspace/tests" |
| 103 | device: l4 |
| 104 | num_devices: 4 |
| 105 | source_file_dependencies: |
| 106 | - vllm/distributed/kv_transfer/kv_connector/v1/nixl/ |
| 107 | - tests/v1/kv_connector/nixl_integration/ |
| 108 | commands: |
| 109 | - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh |
| 110 | - DP_EP=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh |
| 111 | mirror: |
| 112 | amd: |
| 113 | label: ":amd: (MI355) DP EP Distributed NixlConnector PD accuracy" |
| 114 | dind: false |
| 115 | device: mi355_4 |
| 116 | timeout_in_minutes: 40 |
| 117 | depends_on: |
| 118 | - image-build-amd |
| 119 | source_file_dependencies: |
| 120 | - vllm/distributed/kv_transfer/kv_connector/v1/nixl/ |
| 121 | - tests/v1/kv_connector/nixl_integration/ |
| 122 | - vllm/platforms/rocm.py |
| 123 | commands: |
| 124 | - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt |
| 125 | - DP_EP=1 ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh |
| 126 | |
| 127 | - label: ":nvidia: (L4) CrossLayer KV layout Distributed NixlConnector PD accuracy" |
| 128 | key: crosslayer-kv-layout-distributed-nixlconnector-pd-accuracy-tests-4-gpus |
| 129 | timeout_in_minutes: 55 |
| 130 | working_dir: "/vllm-workspace/tests" |
| 131 | device: l4 |
| 132 | num_devices: 4 |
| 133 | source_file_dependencies: |
| 134 | - vllm/distributed/kv_transfer/kv_connector/v1/nixl/ |
| 135 | - tests/v1/kv_connector/nixl_integration/ |
| 136 | commands: |
| 137 | - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh |
| 138 | - CROSS_LAYERS_BLOCKS=True bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh |
| 139 | mirror: |
| 140 | amd: |
| 141 | label: ":amd: (MI300) CrossLayer KV layout Distributed NixlConnector PD accuracy" |
| 142 | dind: false |
| 143 | device: mi300_4 |
| 144 | timeout_in_minutes: 60 |
| 145 | depends_on: |
| 146 | - image-build-amd |
| 147 | source_file_dependencies: |
| 148 | - vllm/distributed/kv_transfer/kv_connector/v1/nixl/ |
| 149 | - tests/v1/kv_connector/nixl_integration/ |
| 150 | - vllm/platforms/rocm.py |
| 151 | commands: |
| 152 | - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt |
| 153 | - CROSS_LAYERS_BLOCKS=True ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh |
| 154 | |
| 155 | - label: ":nvidia: (L4) Hybrid SSM NixlConnector PD accuracy" |
| 156 | key: hybrid-ssm-nixlconnector-pd-accuracy-tests-4-gpus |
| 157 | timeout_in_minutes: 60 |
| 158 | working_dir: "/vllm-workspace/tests" |
| 159 | device: l4 |
| 160 | num_devices: 4 |
| 161 | source_file_dependencies: |
| 162 | - vllm/distributed/kv_transfer/kv_connector/v1/nixl/ |
| 163 | - tests/v1/kv_connector/nixl_integration/ |
| 164 | commands: |
| 165 | - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh |
| 166 | - HYBRID_SSM=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh |
| 167 | mirror: |
| 168 | amd: |
| 169 | label: ":amd: (MI300) Hybrid SSM NixlConnector PD accuracy" |
| 170 | dind: false |
| 171 | device: mi300_4 |
| 172 | timeout_in_minutes: 80 |
| 173 | depends_on: |
| 174 | - image-build-amd |
| 175 | source_file_dependencies: |
| 176 | - vllm/distributed/kv_transfer/kv_connector/v1/nixl/ |
| 177 | - tests/v1/kv_connector/nixl_integration/ |
| 178 | - vllm/platforms/rocm.py |
| 179 | commands: |
| 180 | - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt |
| 181 | - HYBRID_SSM=1 ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh |
| 182 | |
| 183 | - label: ":nvidia: (L4) NixlConnector PD edge case" |
| 184 | key: nixlconnector-pd-edge-cases-2-gpus |
| 185 | timeout_in_minutes: 40 |
| 186 | working_dir: "/vllm-workspace/tests" |
| 187 | device: l4 |
| 188 | num_devices: 2 |
| 189 | source_file_dependencies: |
| 190 | - vllm/distributed/kv_transfer/kv_connector/v1/nixl/ |
| 191 | - vllm/v1/core/sched/ |
| 192 | - tests/v1/kv_connector/nixl_integration/ |
| 193 | env: |
| 194 | PREFILL_GPU_ID: "0" |
| 195 | DECODE_GPU_ID: "1" |
| 196 | commands: |
| 197 | - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh |
| 198 | - bash v1/kv_connector/nixl_integration/run_edge_case_test.sh |
| 199 | mirror: |
| 200 | amd: |
| 201 | label: ":amd: (MI300) NixlConnector PD edge case" |
| 202 | dind: false |
| 203 | device: mi300_2 |
| 204 | timeout_in_minutes: 65 |
| 205 | depends_on: |
| 206 | - image-build-amd |
| 207 | source_file_dependencies: |
| 208 | - vllm/distributed/kv_transfer/kv_connector/v1/nixl/ |
| 209 | - vllm/v1/core/sched/ |
| 210 | - tests/v1/kv_connector/nixl_integration/ |
| 211 | - vllm/platforms/rocm.py |
| 212 | commands: |
| 213 | - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt |
| 214 | - PREFILL_GPU_ID=0 DECODE_GPU_ID=1 ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/run_edge_case_test.sh |
| 215 | |
| 216 | - label: ":nvidia: (L4) Hybrid SSM NixlConnector PD prefix cache" |
| 217 | key: hybrid-ssm-nixlconnector-pd-prefix-cache-2-gpus |
| 218 | timeout_in_minutes: 25 |
| 219 | working_dir: "/vllm-workspace/tests" |
| 220 | device: l4 |
| 221 | num_devices: 2 |
| 222 | source_file_dependencies: |
| 223 | - vllm/distributed/kv_transfer/kv_connector/v1/nixl/ |
| 224 | - vllm/v1/core/sched/ |
| 225 | - vllm/v1/core/kv_cache_coordinator.py |
| 226 | - tests/v1/kv_connector/nixl_integration/ |
| 227 | commands: |
| 228 | - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh |
| 229 | - bash v1/kv_connector/nixl_integration/run_mamba_prefix_cache_test.sh |
| 230 | mirror: |
| 231 | amd: |
| 232 | label: ":amd: (MI355) Hybrid SSM NixlConnector PD prefix cache" |
| 233 | dind: false |
| 234 | device: mi355_2 |
| 235 | timeout_in_minutes: 25 |
| 236 | depends_on: |
| 237 | - image-build-amd |
| 238 | source_file_dependencies: |
| 239 | - vllm/distributed/kv_transfer/kv_connector/v1/nixl/ |
| 240 | - vllm/v1/core/sched/ |
| 241 | - vllm/v1/core/kv_cache_coordinator.py |
| 242 | - tests/v1/kv_connector/nixl_integration/ |
| 243 | - vllm/platforms/rocm.py |
| 244 | commands: |
| 245 | - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt |
| 246 | - ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/run_mamba_prefix_cache_test.sh |
| 247 | |
| 248 | - label: ":nvidia: (L4) MultiConnector (Nixl+Offloading) PD accuracy" |
| 249 | key: multiconnector-nixl-offloading-pd-accuracy-2-gpus |
| 250 | timeout_in_minutes: 40 |
| 251 | working_dir: "/vllm-workspace/tests" |
| 252 | device: l4 |
| 253 | num_devices: 2 |
| 254 | source_file_dependencies: |
| 255 | - vllm/distributed/kv_transfer/kv_connector/v1/nixl/ |
| 256 | - vllm/distributed/kv_transfer/kv_connector/v1/multi_connector.py |
| 257 | - vllm/distributed/kv_transfer/kv_connector/v1/offloading_connector.py |
| 258 | - vllm/distributed/kv_transfer/kv_connector/v1/offloading/ |
| 259 | - tests/v1/kv_connector/nixl_integration/ |
| 260 | commands: |
| 261 | - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh |
| 262 | - bash v1/kv_connector/nixl_integration/run_multi_connector_accuracy_test.sh |
| 263 | mirror: |
| 264 | amd: |
| 265 | label: ":amd: (MI355) MultiConnector (Nixl+Offloading) PD accuracy" |
| 266 | dind: false |
| 267 | device: mi355_2 |
| 268 | timeout_in_minutes: 40 |
| 269 | depends_on: |
| 270 | - image-build-amd |
| 271 | source_file_dependencies: |
| 272 | - vllm/distributed/kv_transfer/kv_connector/v1/nixl/ |
| 273 | - vllm/distributed/kv_transfer/kv_connector/v1/multi_connector.py |
| 274 | - vllm/distributed/kv_transfer/kv_connector/v1/offloading_connector.py |
| 275 | - vllm/distributed/kv_transfer/kv_connector/v1/offloading/ |
| 276 | - tests/v1/kv_connector/nixl_integration/ |
| 277 | - requirements/kv_connectors_rocm.txt |
| 278 | - vllm/platforms/rocm.py |
| 279 | commands: |
| 280 | - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt |
| 281 | - ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/run_multi_connector_accuracy_test.sh |
| 282 | |
| 283 | - label: ":nvidia: (A100) NixlConnector PD + Spec Decode acceptance" |
| 284 | key: nixlconnector-pd-spec-decode-acceptance-2-gpus |
| 285 | timeout_in_minutes: 45 |
| 286 | device: a100 |
| 287 | working_dir: "/vllm-workspace/tests" |
| 288 | num_devices: 2 |
| 289 | source_file_dependencies: |
| 290 | - vllm/distributed/kv_transfer/kv_connector/v1/nixl/ |
| 291 | - vllm/v1/worker/kv_connector_model_runner_mixin.py |
| 292 | - tests/v1/kv_connector/nixl_integration/ |
| 293 | commands: |
| 294 | - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh |
| 295 | - bash v1/kv_connector/nixl_integration/config_sweep_spec_decode_test.sh |
| 296 | mirror: |
| 297 | amd: |
| 298 | label: ":amd: (MI355) NixlConnector PD + Spec Decode acceptance" |
| 299 | dind: false |
| 300 | device: mi355_2 |
| 301 | timeout_in_minutes: 45 |
| 302 | depends_on: |
| 303 | - image-build-amd |
| 304 | source_file_dependencies: |
| 305 | - vllm/distributed/kv_transfer/kv_connector/v1/nixl/ |
| 306 | - vllm/v1/worker/kv_connector_model_runner_mixin.py |
| 307 | - tests/v1/kv_connector/nixl_integration/ |
| 308 | - vllm/platforms/rocm.py |
| 309 | commands: |
| 310 | - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt |
| 311 | - KV_CACHE_MEMORY_BYTES=8G ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_spec_decode_test.sh |
| 312 | |
| 313 | - label: ":nvidia: (L4) MultiConnector (Nixl+Offloading) PD edge cases" |
| 314 | key: multiconnector-nixl-offloading-pd-edge-cases-2-gpus |
| 315 | timeout_in_minutes: 25 |
| 316 | working_dir: "/vllm-workspace/tests" |
| 317 | device: l4 |
| 318 | num_devices: 2 |
| 319 | source_file_dependencies: |
| 320 | - vllm/distributed/kv_transfer/kv_connector/v1/nixl/ |
| 321 | - vllm/distributed/kv_transfer/kv_connector/v1/multi_connector.py |
| 322 | - vllm/distributed/kv_transfer/kv_connector/v1/offloading_connector.py |
| 323 | - vllm/distributed/kv_transfer/kv_connector/v1/offloading/ |
| 324 | - tests/v1/kv_connector/nixl_integration/ |
| 325 | commands: |
| 326 | - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh |
| 327 | - bash v1/kv_connector/nixl_integration/run_multi_connector_edge_case_test.sh |
| 328 | mirror: |
| 329 | amd: |
| 330 | label: ":amd: (MI355) MultiConnector (Nixl+Offloading) PD edge cases" |
| 331 | dind: false |
| 332 | device: mi355_2 |
| 333 | timeout_in_minutes: 25 |
| 334 | depends_on: |
| 335 | - image-build-amd |
| 336 | source_file_dependencies: |
| 337 | - vllm/distributed/kv_transfer/kv_connector/v1/nixl/ |
| 338 | - vllm/distributed/kv_transfer/kv_connector/v1/multi_connector.py |
| 339 | - vllm/distributed/kv_transfer/kv_connector/v1/offloading_connector.py |
| 340 | - vllm/distributed/kv_transfer/kv_connector/v1/offloading/ |
| 341 | - tests/v1/kv_connector/nixl_integration/ |
| 342 | - vllm/platforms/rocm.py |
| 343 | commands: |
| 344 | - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt |
| 345 | - ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/run_multi_connector_edge_case_test.sh |
| 346 | |
| 347 | # P TP 4 - D DPEP 4 test case for DSv4-Flash |
| 348 | - label: ":nvidia: (H200) DSv4-Flash Disaggregated DP EP" |
| 349 | key: dsv4-flash-disaggregated |
| 350 | timeout_in_minutes: 60 |
| 351 | device: h200 |
| 352 | optional: true |
| 353 | working_dir: "/vllm-workspace/tests" |
| 354 | num_devices: 8 |
| 355 | env: |
| 356 | ENABLE_HMA_FLAG: "1" |
| 357 | DP_EP: "1" |
| 358 | GPU_MEMORY_UTILIZATION: "0.85" |
| 359 | PREFILLER_TP_SIZE: "4" |
| 360 | DECODER_TP_SIZE: "4" |
| 361 | PREFILL_BLOCK_SIZE: "256" |
| 362 | DECODE_BLOCK_SIZE: "256" |
| 363 | MODEL_NAMES: "deepseek-ai/DeepSeek-V4-Flash" |
| 364 | VLLM_ENGINE_READY_TIMEOUT_S: "1800" |
| 365 | VLLM_SERVE_EXTRA_ARGS: "--trust-remote-code,--kv-cache-dtype,fp8" |
| 366 | # The V4 indexer only supports block-outermost layouts; let it pick. |
| 367 | PREFILLER_KV_LAYOUT: "" |
| 368 | DECODER_KV_LAYOUT: "" |
| 369 | commands: |
| 370 | - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh |
| 371 | - bash v1/kv_connector/nixl_integration/run_accuracy_test.sh |
| 372 | mirror: |
| 373 | amd: |
| 374 | label: ":amd: (MI355) DSv4-Flash Disaggregated DP EP" |
| 375 | dind: false |
| 376 | device: mi355_8 |
| 377 | timeout_in_minutes: 85 |
| 378 | depends_on: |
| 379 | - image-build-amd |
| 380 | source_file_dependencies: |
| 381 | - requirements/kv_connectors_rocm.txt |
| 382 | - vllm/distributed/kv_transfer/kv_connector/v1/nixl/ |
| 383 | - vllm/model_executor/layers/fused_moe/ |
| 384 | - vllm/model_executor/layers/quantization/ |
| 385 | - vllm/models/deepseek_v4/ |
| 386 | - vllm/v1/attention/ |
| 387 | - tests/v1/kv_connector/nixl_integration/ |
| 388 | - vllm/_aiter_ops.py |
| 389 | - vllm/platforms/rocm.py |
| 390 | commands: |
| 391 | - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt |
| 392 | # A mirror does not inherit the parent env, so it is repeated here in |
| 393 | # the parent's order. The layouts must stay empty: unset, the script |
| 394 | # defaults to HND, which the V4 indexer rejects. |
| 395 | - >- |
| 396 | VLLM_ROCM_USE_AITER=1 |
| 397 | ENABLE_HMA_FLAG=1 |
| 398 | DP_EP=1 |
| 399 | GPU_MEMORY_UTILIZATION=0.85 |
| 400 | PREFILLER_TP_SIZE=4 |
| 401 | DECODER_TP_SIZE=4 |
| 402 | PREFILL_BLOCK_SIZE=256 |
| 403 | DECODE_BLOCK_SIZE=256 |
| 404 | MODEL_NAMES=deepseek-ai/DeepSeek-V4-Flash |
| 405 | VLLM_ENGINE_READY_TIMEOUT_S=1800 |
| 406 | VLLM_SERVE_EXTRA_ARGS=--trust-remote-code,--kv-cache-dtype,fp8 |
| 407 | PREFILLER_KV_LAYOUT= |
| 408 | DECODER_KV_LAYOUT= |
| 409 | bash v1/kv_connector/nixl_integration/run_accuracy_test.sh |
| 410 | |
| 411 | - label: ":nvidia: (H200) Kimi-Linear-48B-A3B Disaggregated DP EP" |
| 412 | key: kimi-linear-disaggregated |
| 413 | timeout_in_minutes: 60 |
| 414 | device: h200 |
| 415 | optional: true |
| 416 | working_dir: "/vllm-workspace/tests" |
| 417 | num_devices: 4 |
| 418 | env: |
| 419 | ENABLE_HMA_FLAG: "1" |
| 420 | DP_EP: "1" |
| 421 | GPU_MEMORY_UTILIZATION: "0.9" |
| 422 | VLLM_SSM_CONV_STATE_LAYOUT: "DS" |
| 423 | PREFILLER_TP_SIZE: "2" |
| 424 | DECODER_TP_SIZE: "2" |
| 425 | MODEL_NAMES: "moonshotai/Kimi-Linear-48B-A3B-Instruct" |
| 426 | # Keep TP=2 prefill and TP=1 decode above both hybrid page-size minima. |
| 427 | PREFILL_BLOCK_SIZE: "2048" |
| 428 | DECODE_BLOCK_SIZE: "2048" |
| 429 | VLLM_SERVE_EXTRA_ARGS: "--trust-remote-code,--mamba-cache-mode align,--max-model-len 32768" |
| 430 | commands: |
| 431 | - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh |
| 432 | - bash v1/kv_connector/nixl_integration/run_accuracy_test.sh |
| 433 | mirror: |
| 434 | amd: |
| 435 | label: ":amd: (MI355) Kimi-Linear-48B-A3B Disaggregated DP EP" |
| 436 | dind: false |
| 437 | device: mi355_4 |
| 438 | num_devices: 4 |
| 439 | timeout_in_minutes: 105 |
| 440 | working_dir: /vllm-workspace/tests |
| 441 | depends_on: |
| 442 | - image-build-amd |
| 443 | source_file_dependencies: |
| 444 | - vllm/platforms/rocm.py |
| 445 | - vllm/models/kimi_k3/ |
| 446 | - vllm/distributed/kv_transfer/kv_connector/v1/nixl/ |
| 447 | - tests/v1/kv_connector/nixl_integration/ |
| 448 | - requirements/kv_connectors_rocm.txt |
| 449 | env: |
| 450 | VLLM_WORKER_MULTIPROC_METHOD: "spawn" |
| 451 | commands: |
| 452 | - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt |
| 453 | - bash v1/kv_connector/nixl_integration/run_accuracy_test.sh |