| 1 | group: LM Eval |
| 2 | depends_on: |
| 3 | - image-build |
| 4 | steps: |
| 5 | - label: ":nvidia: (H200 MIG 35GB) LM Eval Small Models" |
| 6 | device: h200_35gb |
| 7 | key: lm-eval-small-models |
| 8 | timeout_in_minutes: 45 |
| 9 | source_file_dependencies: |
| 10 | - csrc/ |
| 11 | - vllm/model_executor/layers/quantization |
| 12 | autorun_on_main: true |
| 13 | commands: |
| 14 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small.txt |
| 15 | mirror: |
| 16 | amd: |
| 17 | label: ":amd: (MI300) LM Eval Small Models" |
| 18 | dind: false |
| 19 | device: mi300_1 |
| 20 | timeout_in_minutes: 45 |
| 21 | depends_on: |
| 22 | - image-build-amd |
| 23 | source_file_dependencies: |
| 24 | - csrc/ |
| 25 | - vllm/model_executor/layers/quantization |
| 26 | - vllm/model_executor/models/ |
| 27 | - vllm/model_executor/model_loader/ |
| 28 | - vllm/v1/attention/backends/ |
| 29 | - vllm/v1/attention/selector.py |
| 30 | - vllm/_aiter_ops.py |
| 31 | - vllm/platforms/rocm.py |
| 32 | |
| 33 | - label: ":amd: (MI300) LM Eval Small Models Harness" |
| 34 | key: amd-lm-eval-small-models-harness |
| 35 | dind: false |
| 36 | device: mi300_1 |
| 37 | timeout_in_minutes: 35 |
| 38 | depends_on: |
| 39 | - image-build-amd |
| 40 | working_dir: /vllm-workspace/.buildkite/lm-eval-harness |
| 41 | source_file_dependencies: |
| 42 | - csrc/ |
| 43 | - vllm/model_executor/layers/quantization |
| 44 | - vllm/model_executor/models/ |
| 45 | - vllm/model_executor/model_loader/ |
| 46 | - vllm/v1/attention/backends/ |
| 47 | - vllm/v1/attention/selector.py |
| 48 | - vllm/_aiter_ops.py |
| 49 | - vllm/platforms/rocm.py |
| 50 | - .buildkite/lm-eval-harness/ |
| 51 | commands: |
| 52 | - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-small-rocm.txt |
| 53 | |
| 54 | - label: ":nvidia: (H200 MIG 35GB) LM Eval Watermarking" |
| 55 | device: h200_35gb |
| 56 | key: lm-eval-watermarking |
| 57 | # Recorded runs took 80–140 seconds |
| 58 | timeout_in_minutes: 10 |
| 59 | source_file_dependencies: |
| 60 | - .buildkite/test_areas/lm_eval.yaml |
| 61 | - tests/evals/gsm8k/ |
| 62 | - vllm/config/watermarking.py |
| 63 | - vllm/sampling_params.py |
| 64 | - vllm/v1/watermarking/ |
| 65 | - vllm/v1/worker/gpu/ |
| 66 | autorun_on_main: true |
| 67 | commands: |
| 68 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-watermark.txt |
| 69 | mirror: |
| 70 | amd: |
| 71 | label: ":amd: (MI355 DPX) LM Eval Watermarking" |
| 72 | dind: false |
| 73 | device: mi355_dpx |
| 74 | num_devices: 1 |
| 75 | timeout_in_minutes: 45 |
| 76 | working_dir: /vllm-workspace/tests |
| 77 | depends_on: |
| 78 | - image-build-amd |
| 79 | source_file_dependencies: |
| 80 | - .buildkite/test_areas/lm_eval.yaml |
| 81 | - tests/evals/gsm8k/ |
| 82 | - vllm/config/watermarking.py |
| 83 | - vllm/sampling_params.py |
| 84 | - vllm/v1/watermarking/ |
| 85 | - vllm/v1/worker/gpu/ |
| 86 | - vllm/platforms/rocm.py |
| 87 | env: |
| 88 | VLLM_WORKER_MULTIPROC_METHOD: "spawn" |
| 89 | commands: |
| 90 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-watermark.txt |
| 91 | |
| 92 | # - label: LM Eval Large Models (4xA100) |
| 93 | # key: lm-eval-large-models-4xa100 |
| 94 | # device: a100 |
| 95 | # optional: true |
| 96 | # num_devices: 4 |
| 97 | # working_dir: "/vllm-workspace/.buildkite/lm-eval-harness" |
| 98 | # source_file_dependencies: |
| 99 | # - csrc/ |
| 100 | # - vllm/model_executor/layers/quantization |
| 101 | # commands: |
| 102 | # - export VLLM_WORKER_MULTIPROC_METHOD=spawn |
| 103 | # - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4 |
| 104 | |
| 105 | - label: ":nvidia: (H100) LM Eval Large Models" |
| 106 | key: lm-eval-large-models-4xh100 |
| 107 | device: h100 |
| 108 | optional: true |
| 109 | num_devices: 4 |
| 110 | working_dir: "/vllm-workspace/.buildkite/lm-eval-harness" |
| 111 | source_file_dependencies: |
| 112 | - csrc/ |
| 113 | - vllm/model_executor/layers/quantization |
| 114 | commands: |
| 115 | - export VLLM_USE_DEEP_GEMM=0 # We found Triton is faster than DeepGEMM for H100 |
| 116 | - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large-hopper.txt --tp-size=4 |
| 117 | mirror: |
| 118 | amd: |
| 119 | label: ":amd: (MI355) LM Eval Large Models FP8" |
| 120 | dind: false |
| 121 | device: mi355_4 |
| 122 | timeout_in_minutes: 45 |
| 123 | depends_on: |
| 124 | - image-build-amd |
| 125 | source_file_dependencies: |
| 126 | - csrc/ |
| 127 | - vllm/model_executor/layers/quantization |
| 128 | - vllm/model_executor/models/ |
| 129 | - vllm/model_executor/model_loader/ |
| 130 | - vllm/v1/attention/backends/ |
| 131 | - vllm/v1/attention/selector.py |
| 132 | - .buildkite/lm-eval-harness/ |
| 133 | - vllm/_aiter_ops.py |
| 134 | - vllm/platforms/rocm.py |
| 135 | commands: |
| 136 | - export VLLM_USE_DEEP_GEMM=0 |
| 137 | - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large-rocm-tp4.txt --tp-size=4 |
| 138 | |
| 139 | - label: ":nvidia: (B200) LM Eval Small Models" |
| 140 | key: lm-eval-small-models-1xb200 |
| 141 | timeout_in_minutes: 50 |
| 142 | device: b200-k8s |
| 143 | optional: true |
| 144 | source_file_dependencies: |
| 145 | - csrc/ |
| 146 | - vllm/model_executor/layers/quantization |
| 147 | commands: |
| 148 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-blackwell.txt |
| 149 | mirror: |
| 150 | amd: |
| 151 | label: ":amd: (MI355 DPX) LM Eval Small Models" |
| 152 | dind: false |
| 153 | device: mi355_dpx |
| 154 | num_devices: 1 |
| 155 | timeout_in_minutes: 45 |
| 156 | working_dir: /vllm-workspace/tests |
| 157 | depends_on: |
| 158 | - image-build-amd |
| 159 | source_file_dependencies: |
| 160 | - .buildkite/test_areas/lm_eval.yaml |
| 161 | - csrc/ |
| 162 | - tests/evals/gsm8k/ |
| 163 | - vllm/model_executor/layers/quantization |
| 164 | - vllm/model_executor/kernels/linear/ |
| 165 | - vllm/model_executor/layers/fused_moe/ |
| 166 | - vllm/model_executor/models/ |
| 167 | - vllm/model_executor/model_loader/ |
| 168 | - vllm/v1/attention/backends/ |
| 169 | - vllm/v1/attention/selector.py |
| 170 | - vllm/_aiter_ops.py |
| 171 | - vllm/platforms/rocm.py |
| 172 | commands: |
| 173 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-gfx950-small.txt |
| 174 | |
| 175 | - label: ":nvidia: (B200) LM Eval Small Models Distributed" |
| 176 | key: lm-eval-small-models-distributed-2xb200 |
| 177 | timeout_in_minutes: 120 |
| 178 | device: b200-k8s |
| 179 | num_devices: 2 |
| 180 | optional: true |
| 181 | source_file_dependencies: |
| 182 | - csrc/ |
| 183 | - vllm/model_executor/layers/quantization |
| 184 | autorun_on_main: true |
| 185 | commands: |
| 186 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small-tp.txt |
| 187 | mirror: |
| 188 | amd: |
| 189 | label: ":amd: (MI355) LM Eval Small Models Distributed" |
| 190 | dind: false |
| 191 | device: mi355_2 |
| 192 | timeout_in_minutes: 145 |
| 193 | depends_on: |
| 194 | - image-build-amd |
| 195 | source_file_dependencies: |
| 196 | - csrc/ |
| 197 | - vllm/model_executor/layers/quantization |
| 198 | - vllm/model_executor/models/ |
| 199 | - vllm/model_executor/model_loader/ |
| 200 | - vllm/v1/attention/backends/ |
| 201 | - vllm/v1/attention/selector.py |
| 202 | - vllm/v1/worker/ |
| 203 | - vllm/v1/core/ |
| 204 | - vllm/config/ |
| 205 | - tests/evals/gsm8k/configs/models-mi3xx-fp8-and-mixed.txt |
| 206 | - vllm/_aiter_ops.py |
| 207 | - vllm/platforms/rocm.py |
| 208 | - tests/evals/gsm8k/ |
| 209 | - tests/utils.py |
| 210 | - vllm/distributed/ |
| 211 | - vllm/transformers_utils/configs/diffusion_gemma.py |
| 212 | - vllm/v1/attention/ops/ |
| 213 | commands: |
| 214 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi3xx-fp8-and-mixed.txt |
| 215 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small-tp.txt |
| 216 | # Dense MLA PCP-only accuracy; DP1/DCP1. |
| 217 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small-tp-mi355.txt |
| 218 | |
| 219 | - label: ":nvidia: (B200) LM Eval HiSparse Nightly" |
| 220 | key: lm-eval-hisparse-nightly-4xb200 |
| 221 | timeout_in_minutes: 60 |
| 222 | device: b200-k8s |
| 223 | num_devices: 4 |
| 224 | optional: true |
| 225 | source_file_dependencies: |
| 226 | - .buildkite/test_areas/lm_eval.yaml |
| 227 | - tests/evals/gsm8k/ |
| 228 | - vllm/v1/hisparse/ |
| 229 | - vllm/distributed/kv_transfer/kv_connector/v1/hisparse/ |
| 230 | - vllm/v1/attention/backends/mla/ |
| 231 | - vllm/model_executor/layers/attention/sparse_mla_attention.py |
| 232 | - vllm/v1/worker/gpu/ |
| 233 | - csrc/libtorch_stable/hisparse_kernels.cu |
| 234 | commands: |
| 235 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-hisparse.txt |
| 236 | |
| 237 | - label: ":nvidia: (B200) LM Eval PCP" |
| 238 | key: lm-eval-pcp-4xb200 |
| 239 | timeout_in_minutes: 360 |
| 240 | device: b200-k8s |
| 241 | num_devices: 4 |
| 242 | optional: true |
| 243 | source_file_dependencies: |
| 244 | - csrc/ |
| 245 | - tests/evals/gsm8k/configs/GLM-5.2-NVFP4-TP2-PCP2-EP.yaml |
| 246 | - tests/evals/gsm8k/configs/GLM-5.2-NVFP4-TP1-PCP4-EP.yaml |
| 247 | - tests/evals/gsm8k/configs/GLM-5.2-NVFP4-TP1-PCP4-DCP4-EP.yaml |
| 248 | - tests/evals/gsm8k/configs/models-pcp.txt |
| 249 | - vllm/model_executor/layers/quantization |
| 250 | - vllm/config/parallel.py |
| 251 | - vllm/distributed/parallel_state.py |
| 252 | - vllm/model_executor/layers/attention/mla_attention.py |
| 253 | - vllm/model_executor/layers/attention/pcp.py |
| 254 | - vllm/model_executor/layers/attention/sparse_mla_attention.py |
| 255 | - vllm/model_executor/layers/sparse_attn_indexer.py |
| 256 | - vllm/v1/attention/backends/mla/flashmla_sparse.py |
| 257 | - vllm/v1/attention/backends/mla/indexer.py |
| 258 | - vllm/v1/worker/gpu/model_runner.py |
| 259 | - vllm/v1/worker/gpu/pcp_manager.py |
| 260 | autorun_on_main: true |
| 261 | commands: |
| 262 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-pcp.txt |
| 263 | |
| 264 | - label: ":nvidia: (H100) LM Eval Watermark Feature Combination" |
| 265 | key: lm-eval-dspark-watermark-2xh100 |
| 266 | device: h100 |
| 267 | num_devices: 2 |
| 268 | # Local TP2 runs took 84–379 seconds, including startup. |
| 269 | timeout_in_minutes: 15 |
| 270 | source_file_dependencies: |
| 271 | - .buildkite/test_areas/lm_eval.yaml |
| 272 | - tests/evals/gsm8k/ |
| 273 | - vllm/config/speculative.py |
| 274 | - vllm/config/vllm.py |
| 275 | - vllm/config/watermarking.py |
| 276 | - vllm/sampling_params.py |
| 277 | - vllm/model_executor/models/qwen3_dspark.py |
| 278 | - vllm/model_executor/models/qwen3_dflash.py |
| 279 | - vllm/model_executor/warmup/ |
| 280 | - vllm/v1/watermarking/ |
| 281 | - vllm/v1/worker/gpu/ |
| 282 | autorun_on_main: true |
| 283 | commands: |
| 284 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-watermark-dspark-tp2.txt |
| 285 | mirror: |
| 286 | amd: |
| 287 | label: ":amd: (MI355) LM Eval Watermark Feature Combination" |
| 288 | dind: false |
| 289 | device: mi355_2 |
| 290 | num_devices: 2 |
| 291 | timeout_in_minutes: 60 |
| 292 | working_dir: /vllm-workspace/tests |
| 293 | depends_on: |
| 294 | - image-build-amd |
| 295 | source_file_dependencies: |
| 296 | - .buildkite/test_areas/lm_eval.yaml |
| 297 | - tests/evals/gsm8k/ |
| 298 | - vllm/config/speculative.py |
| 299 | - vllm/config/vllm.py |
| 300 | - vllm/config/watermarking.py |
| 301 | - vllm/sampling_params.py |
| 302 | - vllm/model_executor/models/qwen3_dspark.py |
| 303 | - vllm/model_executor/models/qwen3_dflash.py |
| 304 | - vllm/model_executor/warmup/ |
| 305 | - vllm/v1/watermarking/ |
| 306 | - vllm/v1/worker/gpu/ |
| 307 | - vllm/platforms/rocm.py |
| 308 | env: |
| 309 | VLLM_WORKER_MULTIPROC_METHOD: "spawn" |
| 310 | commands: |
| 311 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-watermark-dspark-tp2.txt |
| 312 | |
| 313 | - label: ":nvidia: (B200) LM Eval Spec Decode" |
| 314 | key: lm-eval-spec-decode-4xb200 |
| 315 | timeout_in_minutes: 120 |
| 316 | device: b200-k8s |
| 317 | num_devices: 4 |
| 318 | optional: true |
| 319 | source_file_dependencies: |
| 320 | - tests/evals/gsm8k/configs/DeepSeek-V4-Flash-DSpark-confidence-TEP4.yaml |
| 321 | - tests/evals/gsm8k/configs/Kimi-K3-pruned75-DSpark-TP4.yaml |
| 322 | - tests/evals/gsm8k/configs/models-spec-decode.txt |
| 323 | - vllm/models/deepseek_v4/ |
| 324 | - vllm/models/kimi_k3/ |
| 325 | - vllm/model_executor/models/qwen3_dspark.py |
| 326 | - vllm/v1/spec_decode/metrics.py |
| 327 | - vllm/v1/worker/gpu/spec_decode/adaptive_verification.py |
| 328 | - vllm/v1/worker/gpu/spec_decode/dflash/ |
| 329 | - vllm/v1/worker/gpu/spec_decode/dspark/ |
| 330 | - vllm/v1/worker/gpu/spec_decode/rejection_sampler.py |
| 331 | - vllm/v1/worker/gpu/spec_decode/speculator.py |
| 332 | commands: |
| 333 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-spec-decode.txt |
| 334 | mirror: |
| 335 | amd: |
| 336 | label: ":amd: (MI355) LM Eval Spec Decode (Fixed-length FP8)" |
| 337 | dind: false |
| 338 | device: mi355_4 |
| 339 | timeout_in_minutes: 145 |
| 340 | depends_on: |
| 341 | - image-build-amd |
| 342 | source_file_dependencies: |
| 343 | - csrc/ |
| 344 | - tests/evals/gsm8k/configs/Kimi-K3-pruned75-DSpark-AITER-TP4.yaml |
| 345 | - tests/evals/gsm8k/configs/DeepSeek-V4-Flash-DSpark-AITER-TEP4.yaml |
| 346 | - tests/evals/gsm8k/configs/DeepSeek-V4-Flash-DSpark-FP8-TP4-ROCm.yaml |
| 347 | - tests/evals/gsm8k/configs/models-spec-decode-rocm.txt |
| 348 | - tests/evals/gsm8k/test_gsm8k_correctness.py |
| 349 | - tests/evals/gsm8k/gsm8k_eval.py |
| 350 | - tests/evals/gsm8k/conftest.py |
| 351 | - vllm/config/ |
| 352 | - vllm/envs.py |
| 353 | - vllm/models/kimi_k3/ |
| 354 | - vllm/models/deepseek_v4/ |
| 355 | - vllm/model_executor/kernels/mhc/ |
| 356 | - vllm/model_executor/layers/mhc.py |
| 357 | - vllm/model_executor/layers/quantization/ |
| 358 | - vllm/model_executor/layers/sparse_attn_indexer.py |
| 359 | - vllm/model_executor/models/config.py |
| 360 | - vllm/model_executor/models/registry.py |
| 361 | - vllm/model_executor/models/qwen3_dspark.py |
| 362 | - vllm/model_executor/model_loader/ |
| 363 | - vllm/model_executor/warmup/ |
| 364 | - vllm/model_executor/layers/fused_moe/ |
| 365 | - vllm/tokenizers/ |
| 366 | - vllm/v1/attention/ |
| 367 | - vllm/v1/core/kv_cache_utils.py |
| 368 | - vllm/v1/spec_decode/metrics.py |
| 369 | - vllm/v1/worker/ |
| 370 | - vllm/_aiter_ops.py |
| 371 | - vllm/platforms/rocm.py |
| 372 | - tests/evals/gsm8k/ |
| 373 | - tests/utils.py |
| 374 | - vllm/distributed/ |
| 375 | commands: |
| 376 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-spec-decode-rocm.txt |
| 377 | |
| 378 | - label: ":nvidia: (B200) LM Eval Large Models EP" |
| 379 | key: lm-eval-large-models-ep-2xb200 |
| 380 | timeout_in_minutes: 60 |
| 381 | device: b200-k8s |
| 382 | optional: true |
| 383 | num_devices: 2 |
| 384 | source_file_dependencies: |
| 385 | - csrc/ |
| 386 | - vllm/model_executor/layers/quantization |
| 387 | commands: |
| 388 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-blackwell-ep.txt |
| 389 | mirror: |
| 390 | amd: |
| 391 | label: ":amd: (MI355) LM Eval Large Models EP" |
| 392 | dind: false |
| 393 | device: mi355_2 |
| 394 | num_devices: 2 |
| 395 | timeout_in_minutes: 120 |
| 396 | working_dir: "/vllm-workspace/tests" |
| 397 | depends_on: |
| 398 | - image-build-amd |
| 399 | source_file_dependencies: |
| 400 | - .buildkite/test_areas/lm_eval.yaml |
| 401 | - tests/evals/gsm8k/ |
| 402 | - tests/utils.py |
| 403 | - tests/conftest.py |
| 404 | - csrc/ |
| 405 | - vllm/config/ |
| 406 | - vllm/distributed/ |
| 407 | - vllm/model_executor/models/config.py |
| 408 | - vllm/model_executor/models/registry.py |
| 409 | - vllm/model_executor/models/utils.py |
| 410 | - vllm/model_executor/models/qwen3_next.py |
| 411 | - vllm/model_executor/models/qwen3_next_mtp.py |
| 412 | - vllm/model_executor/models/nemotron_h.py |
| 413 | - vllm/model_executor/models/nemotron_h_mtp.py |
| 414 | - vllm/transformers_utils/configs/qwen3_next.py |
| 415 | - vllm/transformers_utils/configs/nemotron_h.py |
| 416 | - vllm/model_executor/model_loader/ |
| 417 | - vllm/model_executor/layers/quantization/ |
| 418 | - vllm/model_executor/layers/fused_moe/ |
| 419 | - vllm/model_executor/layers/mamba/ |
| 420 | - vllm/model_executor/layers/layernorm.py |
| 421 | - vllm/model_executor/layers/linear.py |
| 422 | - vllm/model_executor/kernels/linear/ |
| 423 | - vllm/third_party/flash_linear_attention/ops/ |
| 424 | - vllm/v1/attention/ |
| 425 | - vllm/v1/spec_decode/ |
| 426 | - vllm/v1/worker/ |
| 427 | - vllm/_aiter_ops.py |
| 428 | - vllm/platforms/rocm.py |
| 429 | commands: |
| 430 | - export VLLM_WORKER_MULTIPROC_METHOD=spawn |
| 431 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-blackwell-ep-mi355.txt |
| 432 | |
| 433 | - label: ":nvidia: (B200) LM Eval Qwen3.5 Models" |
| 434 | key: lm-eval-qwen3-5-models-2xb200 |
| 435 | timeout_in_minutes: 45 |
| 436 | device: b200-k8s |
| 437 | optional: true |
| 438 | num_devices: 2 |
| 439 | source_file_dependencies: |
| 440 | - vllm/model_executor/models/qwen3_5.py |
| 441 | - vllm/model_executor/models/qwen3_5_mtp.py |
| 442 | - vllm/transformers_utils/configs/qwen3_5.py |
| 443 | - vllm/transformers_utils/configs/qwen3_5_moe.py |
| 444 | - vllm/model_executor/models/qwen3_next.py |
| 445 | - vllm/model_executor/models/qwen3_next_mtp.py |
| 446 | - vllm/third_party/flash_linear_attention/ops/ |
| 447 | commands: |
| 448 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-qwen35-blackwell.txt |
| 449 | mirror: |
| 450 | amd: |
| 451 | label: ":amd: (MI355) LM Eval Qwen3.5 Models" |
| 452 | dind: false |
| 453 | device: mi355_2 |
| 454 | timeout_in_minutes: 75 |
| 455 | depends_on: |
| 456 | - image-build-amd |
| 457 | source_file_dependencies: |
| 458 | - vllm/model_executor/models/qwen3_5.py |
| 459 | - vllm/model_executor/models/qwen3_5_mtp.py |
| 460 | - vllm/transformers_utils/configs/qwen3_5.py |
| 461 | - vllm/transformers_utils/configs/qwen3_5_moe.py |
| 462 | - vllm/model_executor/models/qwen2.py |
| 463 | - vllm/model_executor/models/qwen3.py |
| 464 | - vllm/model_executor/models/qwen3_next.py |
| 465 | - vllm/model_executor/models/qwen3_next_mtp.py |
| 466 | - vllm/third_party/flash_linear_attention/ops/ |
| 467 | - vllm/distributed/ |
| 468 | - vllm/model_executor/layers/fused_moe/ |
| 469 | - vllm/model_executor/layers/quantization/modelopt.py |
| 470 | - vllm/model_executor/kernels/linear/nvfp4/ |
| 471 | - vllm/v1/attention/ |
| 472 | - vllm/v1/worker/ |
| 473 | - tests/evals/gsm8k/configs/models-qwen35-mi355.txt |
| 474 | - tests/evals/gsm8k/configs/Qwen3.5-35B-A3B-DEP2.yaml |
| 475 | - tests/evals/gsm8k/configs/Qwen3.5-35B-A3B-FP8-DEP2.yaml |
| 476 | - tests/evals/gsm8k/configs/Qwen3.5-35B-A3B-MXFP4-AITER-TP2.yaml |
| 477 | - tests/evals/gsm8k/configs/Qwen3.5-35B-A3B-MXFP4-EMU-TP2.yaml |
| 478 | - tests/evals/gsm8k/configs/Qwen3.5-397B-A17B-NVFP4-DEP2.yaml |
| 479 | - tests/evals/gsm8k/test_gsm8k_correctness.py |
| 480 | - vllm/_aiter_ops.py |
| 481 | - vllm/platforms/rocm.py |
| 482 | commands: |
| 483 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-qwen35-mi355.txt |
| 484 | |
| 485 | - label: ":nvidia: (H200) LM Eval DeepSeek R1 TP" |
| 486 | key: lm-eval-deepseek-r1-tp-8xh200 |
| 487 | timeout_in_minutes: 35 |
| 488 | device: h200 |
| 489 | optional: true |
| 490 | num_devices: 8 |
| 491 | commands: |
| 492 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-h200-deepseek-r1-tp.txt |
| 493 | mirror: |
| 494 | amd: |
| 495 | label: ":amd: (MI355) LM Eval DeepSeek R1 TP" |
| 496 | dind: false |
| 497 | device: mi355_4 |
| 498 | num_devices: 4 |
| 499 | timeout_in_minutes: 65 |
| 500 | working_dir: /vllm-workspace/tests |
| 501 | depends_on: |
| 502 | - image-build-amd |
| 503 | source_file_dependencies: |
| 504 | - .buildkite/test_areas/lm_eval.yaml |
| 505 | - tests/evals/gsm8k/ |
| 506 | - tests/utils.py |
| 507 | - tests/conftest.py |
| 508 | - vllm/config/ |
| 509 | - vllm/distributed/ |
| 510 | - vllm/model_executor/models/ |
| 511 | - vllm/models/ |
| 512 | - vllm/model_executor/model_loader/ |
| 513 | - vllm/model_executor/layers/quantization/ |
| 514 | - vllm/model_executor/layers/fused_moe/ |
| 515 | - vllm/v1/attention/ |
| 516 | - vllm/v1/spec_decode/ |
| 517 | - vllm/v1/worker/gpu/spec_decode/ |
| 518 | - vllm/v1/core/kv_cache_utils.py |
| 519 | - vllm/model_executor/layers/layernorm.py |
| 520 | - csrc/ |
| 521 | - vllm/_aiter_ops.py |
| 522 | - vllm/platforms/rocm.py |
| 523 | - vllm/model_executor/layers/sparse_attn_indexer.py |
| 524 | commands: |
| 525 | - export VLLM_WORKER_MULTIPROC_METHOD=spawn |
| 526 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi355-deepseek-r1-tp.txt |
| 527 | |
| 528 | - label: ":nvidia: (H200) LM Eval DeepSeek R1 DP" |
| 529 | key: lm-eval-deepseek-r1-dp-8xh200 |
| 530 | timeout_in_minutes: 35 |
| 531 | device: h200 |
| 532 | optional: true |
| 533 | num_devices: 8 |
| 534 | commands: |
| 535 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-h200-deepseek-r1-dp.txt |
| 536 | mirror: |
| 537 | amd: |
| 538 | label: ":amd: (MI355) LM Eval DeepSeek R1 DP" |
| 539 | dind: false |
| 540 | device: mi355_4 |
| 541 | num_devices: 4 |
| 542 | timeout_in_minutes: 65 |
| 543 | working_dir: /vllm-workspace/tests |
| 544 | depends_on: |
| 545 | - image-build-amd |
| 546 | source_file_dependencies: |
| 547 | - .buildkite/test_areas/lm_eval.yaml |
| 548 | - tests/evals/gsm8k/ |
| 549 | - tests/utils.py |
| 550 | - tests/conftest.py |
| 551 | - vllm/config/ |
| 552 | - vllm/distributed/ |
| 553 | - vllm/model_executor/models/ |
| 554 | - vllm/models/ |
| 555 | - vllm/model_executor/model_loader/ |
| 556 | - vllm/model_executor/layers/quantization/ |
| 557 | - vllm/model_executor/layers/fused_moe/ |
| 558 | - vllm/v1/attention/ |
| 559 | - vllm/v1/spec_decode/ |
| 560 | - vllm/v1/worker/gpu/spec_decode/ |
| 561 | - vllm/v1/core/kv_cache_utils.py |
| 562 | - vllm/model_executor/layers/layernorm.py |
| 563 | - csrc/ |
| 564 | - vllm/_aiter_ops.py |
| 565 | - vllm/platforms/rocm.py |
| 566 | - vllm/model_executor/layers/sparse_attn_indexer.py |
| 567 | commands: |
| 568 | - export VLLM_WORKER_MULTIPROC_METHOD=spawn |
| 569 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi355-deepseek-r1-dp.txt |
| 570 | |
| 571 | - label: ":nvidia: (H200) LM Eval DeepSeek V3.2 TP" |
| 572 | key: lm-eval-deepseek-v32-tp-8xh200 |
| 573 | timeout_in_minutes: 35 |
| 574 | device: h200 |
| 575 | optional: true |
| 576 | num_devices: 8 |
| 577 | commands: |
| 578 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-h200-deepseek-v32-tp.txt |
| 579 | mirror: |
| 580 | amd: |
| 581 | label: ":amd: (MI355) LM Eval DeepSeek V3-2 TP" |
| 582 | dind: false |
| 583 | device: mi355_4 |
| 584 | num_devices: 4 |
| 585 | timeout_in_minutes: 65 |
| 586 | working_dir: /vllm-workspace/tests |
| 587 | depends_on: |
| 588 | - image-build-amd |
| 589 | source_file_dependencies: |
| 590 | - .buildkite/test_areas/lm_eval.yaml |
| 591 | - tests/evals/gsm8k/ |
| 592 | - tests/utils.py |
| 593 | - tests/conftest.py |
| 594 | - vllm/config/ |
| 595 | - vllm/distributed/ |
| 596 | - vllm/model_executor/models/ |
| 597 | - vllm/models/ |
| 598 | - vllm/model_executor/model_loader/ |
| 599 | - vllm/model_executor/layers/quantization/ |
| 600 | - vllm/model_executor/layers/fused_moe/ |
| 601 | - vllm/v1/attention/ |
| 602 | - vllm/v1/spec_decode/ |
| 603 | - vllm/v1/worker/gpu/spec_decode/ |
| 604 | - vllm/v1/core/kv_cache_utils.py |
| 605 | - vllm/model_executor/layers/layernorm.py |
| 606 | - csrc/ |
| 607 | - vllm/_aiter_ops.py |
| 608 | - vllm/platforms/rocm.py |
| 609 | - vllm/model_executor/layers/sparse_attn_indexer.py |
| 610 | commands: |
| 611 | - export VLLM_WORKER_MULTIPROC_METHOD=spawn |
| 612 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi355-deepseek-v32-tp.txt |
| 613 | |
| 614 | - label: ":nvidia: (H200) LM Eval DeepSeek V3.2 DP" |
| 615 | key: lm-eval-deepseek-v32-dp-8xh200 |
| 616 | timeout_in_minutes: 35 |
| 617 | device: h200 |
| 618 | optional: true |
| 619 | num_devices: 8 |
| 620 | commands: |
| 621 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-h200-deepseek-v32-dp.txt |
| 622 | mirror: |
| 623 | amd: |
| 624 | label: ":amd: (MI355) LM Eval DeepSeek V3-2 DP" |
| 625 | dind: false |
| 626 | device: mi355_4 |
| 627 | num_devices: 4 |
| 628 | timeout_in_minutes: 65 |
| 629 | working_dir: /vllm-workspace/tests |
| 630 | depends_on: |
| 631 | - image-build-amd |
| 632 | source_file_dependencies: |
| 633 | - .buildkite/test_areas/lm_eval.yaml |
| 634 | - tests/evals/gsm8k/ |
| 635 | - tests/utils.py |
| 636 | - tests/conftest.py |
| 637 | - vllm/config/ |
| 638 | - vllm/distributed/ |
| 639 | - vllm/model_executor/models/ |
| 640 | - vllm/models/ |
| 641 | - vllm/model_executor/model_loader/ |
| 642 | - vllm/model_executor/layers/quantization/ |
| 643 | - vllm/model_executor/layers/fused_moe/ |
| 644 | - vllm/v1/attention/ |
| 645 | - vllm/v1/spec_decode/ |
| 646 | - vllm/v1/worker/gpu/spec_decode/ |
| 647 | - vllm/v1/core/kv_cache_utils.py |
| 648 | - vllm/model_executor/layers/layernorm.py |
| 649 | - csrc/ |
| 650 | - vllm/_aiter_ops.py |
| 651 | - vllm/platforms/rocm.py |
| 652 | - vllm/model_executor/layers/sparse_attn_indexer.py |
| 653 | commands: |
| 654 | - export VLLM_WORKER_MULTIPROC_METHOD=spawn |
| 655 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi355-deepseek-v32-dp.txt |
| 656 | |
| 657 | - label: ":nvidia: (H200) LM Eval Nemotron 3 Super" |
| 658 | key: lm-eval-nemotron-3-super-8xh200 |
| 659 | timeout_in_minutes: 35 |
| 660 | device: h200 |
| 661 | optional: true |
| 662 | num_devices: 8 |
| 663 | commands: |
| 664 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-h200-nemotron-3-super.txt |
| 665 | mirror: |
| 666 | amd: |
| 667 | label: ":amd: (MI355) LM Eval Nemotron 3 Super" |
| 668 | dind: false |
| 669 | device: mi355_4 |
| 670 | num_devices: 4 |
| 671 | timeout_in_minutes: 65 |
| 672 | working_dir: /vllm-workspace/tests |
| 673 | depends_on: |
| 674 | - image-build-amd |
| 675 | source_file_dependencies: |
| 676 | - .buildkite/test_areas/lm_eval.yaml |
| 677 | - tests/evals/gsm8k/ |
| 678 | - tests/utils.py |
| 679 | - tests/conftest.py |
| 680 | - vllm/config/ |
| 681 | - vllm/distributed/ |
| 682 | - vllm/model_executor/models/ |
| 683 | - vllm/models/ |
| 684 | - vllm/model_executor/model_loader/ |
| 685 | - vllm/model_executor/layers/quantization/ |
| 686 | - vllm/model_executor/layers/fused_moe/ |
| 687 | - vllm/v1/attention/ |
| 688 | - vllm/v1/spec_decode/ |
| 689 | - vllm/v1/worker/gpu/spec_decode/ |
| 690 | - vllm/v1/core/kv_cache_utils.py |
| 691 | - vllm/model_executor/layers/layernorm.py |
| 692 | - csrc/ |
| 693 | - vllm/_aiter_ops.py |
| 694 | - vllm/platforms/rocm.py |
| 695 | - vllm/model_executor/layers/mamba/ |
| 696 | commands: |
| 697 | - export VLLM_WORKER_MULTIPROC_METHOD=spawn |
| 698 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi355-nemotron-3-super.txt |
| 699 | |
| 700 | - label: ":amd: (MI300) LM Eval Large Models" |
| 701 | key: amd-lm-eval-large-models |
| 702 | dind: false |
| 703 | device: mi300_8 |
| 704 | optional: true |
| 705 | timeout_in_minutes: 65 |
| 706 | depends_on: |
| 707 | - image-build-amd |
| 708 | commands: |
| 709 | - export VLLM_WORKER_MULTIPROC_METHOD=spawn |
| 710 | - export PYTORCH_ROCM_ARCH=gfx942 # Limit Quark compilation to save time |
| 711 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi3xx.txt |
| 712 | source_file_dependencies: |
| 713 | - vllm/_aiter_ops.py |
| 714 | - vllm/platforms/rocm.py |
| 715 | |
| 716 | - label: ":nvidia: (H100) MoE Refactor Integration TEMPORARY" |
| 717 | key: moe-refactor-integration-test-h100-temporary |
| 718 | device: h100 |
| 719 | optional: true |
| 720 | num_devices: 2 |
| 721 | parallelism: 4 |
| 722 | commands: |
| 723 | # Timing-balanced static split of the former config-h100.txt (measured on |
| 724 | # build 83851): shards ~12.4/13.6/12.4/14.5m of eval time. New configs must |
| 725 | # be added to exactly one config-h100-shard-*.txt. |
| 726 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/moe-refactor/config-h100-shard-$$BUILDKITE_PARALLEL_JOB.txt |
| 727 | mirror: |
| 728 | amd: |
| 729 | label: ":amd: (MI355) MoE Refactor Integration Shard %N" |
| 730 | dind: false |
| 731 | device: mi355_2 |
| 732 | num_devices: 2 |
| 733 | timeout_in_minutes: 105 |
| 734 | working_dir: /vllm-workspace/tests |
| 735 | depends_on: |
| 736 | - image-build-amd |
| 737 | source_file_dependencies: |
| 738 | - vllm/platforms/rocm.py |
| 739 | - tests/evals/gsm8k/ |
| 740 | - vllm/model_executor/layers/fused_moe/ |
| 741 | - vllm/model_executor/layers/quantization/ |
| 742 | - vllm/_aiter_ops.py |
| 743 | env: |
| 744 | VLLM_WORKER_MULTIPROC_METHOD: "spawn" |
| 745 | commands: |
| 746 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/moe-refactor/config-rocm-shard-$$BUILDKITE_PARALLEL_JOB.txt |
| 747 | |
| 748 | - label: ":nvidia: (B200) MoE Refactor Integration TEMPORARY Shard %N" |
| 749 | key: moe-refactor-integration-test-b200-temporary |
| 750 | device: b200-k8s |
| 751 | optional: true |
| 752 | num_devices: 2 |
| 753 | parallelism: 4 |
| 754 | commands: |
| 755 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/moe-refactor/config-b200-shard-$$BUILDKITE_PARALLEL_JOB.txt |
| 756 | |
| 757 | - label: ":nvidia: (B200) MoE Refactor DP Integration TEMPORARY" |
| 758 | key: moe-refactor-integration-test-b200-dp-temporary |
| 759 | device: b200-k8s |
| 760 | optional: true |
| 761 | num_devices: 2 |
| 762 | commands: |
| 763 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/moe-refactor-dp-ep/config-b200.txt |
| 764 | |
| 765 | - label: ":nvidia: (L4) LM Eval Humming TEMPORARY" |
| 766 | key: lm-eval-humming-l4 |
| 767 | optional: true |
| 768 | timeout_in_minutes: 90 |
| 769 | device: l4 |
| 770 | num_devices: 1 |
| 771 | source_file_dependencies: |
| 772 | - .buildkite/test_areas/lm_eval.yaml |
| 773 | - tests/evals/gsm8k/configs/humming/ |
| 774 | - requirements/cuda.txt |
| 775 | - vllm/utils/humming.py |
| 776 | - vllm/model_executor/layers/quantization/humming.py |
| 777 | - vllm/model_executor/layers/quantization/utils/humming/ |
| 778 | - vllm/model_executor/layers/fused_moe/experts/fused_humming_moe.py |
| 779 | - vllm/model_executor/layers/fused_moe/oracle/ |
| 780 | - vllm/model_executor/kernels/linear/ |
| 781 | commands: |
| 782 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/humming/config-l4.txt |
| 783 | |
| 784 | - label: ":nvidia: (A100) LM Eval Humming Default TEMPORARY Shard %N" |
| 785 | key: lm-eval-humming-a100-default |
| 786 | optional: true |
| 787 | timeout_in_minutes: 75 |
| 788 | device: a100 |
| 789 | num_devices: 1 |
| 790 | parallelism: 3 |
| 791 | source_file_dependencies: |
| 792 | - vllm/model_executor/layers/quantization/humming.py |
| 793 | - vllm/model_executor/layers/quantization/utils/humming/ |
| 794 | - vllm/model_executor/layers/fused_moe/experts/fused_humming_moe.py |
| 795 | - vllm/model_executor/layers/fused_moe/oracle/ |
| 796 | - vllm/model_executor/kernels/linear/ |
| 797 | commands: |
| 798 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/humming/config-a100-shard-$$BUILDKITE_PARALLEL_JOB.txt |
| 799 | |
| 800 | - label: ":nvidia: (A100) LM Eval Humming Activation INT8 TEMPORARY Shard %N" |
| 801 | key: lm-eval-humming-a100-act |
| 802 | optional: true |
| 803 | timeout_in_minutes: 45 |
| 804 | device: a100 |
| 805 | num_devices: 1 |
| 806 | parallelism: 2 |
| 807 | source_file_dependencies: |
| 808 | - vllm/model_executor/layers/quantization/humming.py |
| 809 | - vllm/model_executor/layers/quantization/utils/humming/ |
| 810 | - vllm/model_executor/layers/fused_moe/experts/fused_humming_moe.py |
| 811 | - vllm/model_executor/layers/fused_moe/oracle/ |
| 812 | - vllm/model_executor/kernels/linear/ |
| 813 | commands: |
| 814 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/humming/config-a100-act-int8-shard-$$BUILDKITE_PARALLEL_JOB.txt |
| 815 | |
| 816 | - label: ":nvidia: (H100) LM Eval Humming Default TEMPORARY Shard %N" |
| 817 | key: lm-eval-humming-h100-default |
| 818 | optional: true |
| 819 | timeout_in_minutes: 70 |
| 820 | device: h100 |
| 821 | num_devices: 1 |
| 822 | parallelism: 3 |
| 823 | source_file_dependencies: |
| 824 | - vllm/model_executor/layers/quantization/humming.py |
| 825 | - vllm/model_executor/layers/quantization/utils/humming/ |
| 826 | - vllm/model_executor/layers/fused_moe/experts/fused_humming_moe.py |
| 827 | - vllm/model_executor/layers/fused_moe/oracle/ |
| 828 | - vllm/model_executor/kernels/linear/ |
| 829 | commands: |
| 830 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/humming/config-h100-shard-$$BUILDKITE_PARALLEL_JOB.txt |
| 831 | |
| 832 | - label: ":nvidia: (H100) LM Eval Humming Activation FP8/INT8 TEMPORARY" |
| 833 | key: lm-eval-humming-h100-act |
| 834 | optional: true |
| 835 | timeout_in_minutes: 90 |
| 836 | device: h100 |
| 837 | num_devices: 1 |
| 838 | source_file_dependencies: |
| 839 | - vllm/model_executor/layers/quantization/humming.py |
| 840 | - vllm/model_executor/layers/quantization/utils/humming/ |
| 841 | - vllm/model_executor/layers/fused_moe/experts/fused_humming_moe.py |
| 842 | - vllm/model_executor/layers/fused_moe/oracle/ |
| 843 | - vllm/model_executor/kernels/linear/ |
| 844 | commands: |
| 845 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/humming/config-act-fp8.txt |
| 846 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/humming/config-act-int8.txt |
| 847 | |
| 848 | - label: ":nvidia: (H100) LM Eval Humming DeepEP TEMPORARY" |
| 849 | key: lm-eval-humming-h100-deepep |
| 850 | optional: true |
| 851 | timeout_in_minutes: 60 |
| 852 | device: h100 |
| 853 | num_devices: 2 |
| 854 | source_file_dependencies: |
| 855 | - .buildkite/test_areas/lm_eval.yaml |
| 856 | - tests/evals/gsm8k/configs/humming/ |
| 857 | - requirements/cuda.txt |
| 858 | - vllm/utils/humming.py |
| 859 | - vllm/model_executor/layers/quantization/humming.py |
| 860 | - vllm/model_executor/layers/quantization/utils/humming/ |
| 861 | - vllm/model_executor/layers/fused_moe/ |
| 862 | - vllm/model_executor/kernels/linear/ |
| 863 | - vllm/distributed/ |
| 864 | - csrc/moe/ |
| 865 | commands: |
| 866 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/humming/config-deepep.txt |
| 867 | |
| 868 | - label: ":nvidia: (B200) LM Eval Humming Default TEMPORARY" |
| 869 | key: lm-eval-humming-b200-default |
| 870 | optional: true |
| 871 | timeout_in_minutes: 50 |
| 872 | device: b200-k8s |
| 873 | num_devices: 1 |
| 874 | source_file_dependencies: |
| 875 | - vllm/model_executor/layers/quantization/humming.py |
| 876 | - vllm/model_executor/layers/quantization/utils/humming/ |
| 877 | - vllm/model_executor/layers/fused_moe/experts/fused_humming_moe.py |
| 878 | - vllm/model_executor/layers/fused_moe/oracle/ |
| 879 | - vllm/model_executor/kernels/linear/ |
| 880 | commands: |
| 881 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/humming/config.txt |
| 882 | |
| 883 | - label: ":nvidia: (B200) LM Eval Humming Activation FP8/INT8 TEMPORARY" |
| 884 | key: lm-eval-humming-b200-act |
| 885 | optional: true |
| 886 | timeout_in_minutes: 50 |
| 887 | device: b200-k8s |
| 888 | num_devices: 1 |
| 889 | source_file_dependencies: |
| 890 | - vllm/model_executor/layers/quantization/humming.py |
| 891 | - vllm/model_executor/layers/quantization/utils/humming/ |
| 892 | - vllm/model_executor/layers/fused_moe/experts/fused_humming_moe.py |
| 893 | - vllm/model_executor/layers/fused_moe/oracle/ |
| 894 | - vllm/model_executor/kernels/linear/ |
| 895 | commands: |
| 896 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/humming/config-act-fp8.txt |
| 897 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/humming/config-act-int8.txt |
| 898 | |
| 899 | - label: ":nvidia: (H200 MIG 18GB) LM Eval TurboQuant k8v4" |
| 900 | key: lm-eval-turboquant-k8v4 |
| 901 | timeout_in_minutes: 30 |
| 902 | device: h200_18gb |
| 903 | source_file_dependencies: |
| 904 | - vllm/model_executor/layers/quantization/turboquant/ |
| 905 | - vllm/v1/attention/backends/turboquant_attn.py |
| 906 | - vllm/v1/attention/ops/triton_turboquant_decode.py |
| 907 | - vllm/v1/attention/ops/triton_turboquant_store.py |
| 908 | commands: |
| 909 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/models-turboquant-k8v4.txt |
| 910 | mirror: |
| 911 | amd: |
| 912 | label: ":amd: (MI355 DPX) LM Eval TurboQuant k8v4" |
| 913 | dind: false |
| 914 | device: mi355_dpx |
| 915 | num_devices: 1 |
| 916 | timeout_in_minutes: 30 |
| 917 | working_dir: /vllm-workspace/tests |
| 918 | depends_on: |
| 919 | - image-build-amd |
| 920 | source_file_dependencies: |
| 921 | - .buildkite/test_areas/lm_eval.yaml |
| 922 | - tests/evals/gsm8k/ |
| 923 | - tests/utils.py |
| 924 | - tests/conftest.py |
| 925 | - vllm/model_executor/layers/quantization/turboquant/ |
| 926 | - vllm/v1/attention/backends/turboquant_attn.py |
| 927 | - vllm/v1/attention/backends/triton_attn.py |
| 928 | - vllm/v1/attention/ops/triton_unified_attention.py |
| 929 | - vllm/v1/attention/selector.py |
| 930 | - vllm/v1/attention/ops/triton_turboquant_decode.py |
| 931 | - vllm/v1/attention/ops/triton_turboquant_store.py |
| 932 | - vllm/v1/attention/ops/flydsl_turboquant_decode.py |
| 933 | - vllm/v1/attention/ops/flydsl_kernels/ |
| 934 | - vllm/v1/attention/ops/turboquant_soa/ |
| 935 | - vllm/platforms/rocm.py |
| 936 | commands: |
| 937 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/models-turboquant-k8v4-rocm.txt |
| 938 | |
| 939 | - label: ":nvidia: (H200 MIG 18GB) LM Eval TurboQuant t4nc" |
| 940 | key: lm-eval-turboquant-t4nc |
| 941 | timeout_in_minutes: 30 |
| 942 | device: h200_18gb |
| 943 | source_file_dependencies: |
| 944 | - vllm/model_executor/layers/quantization/turboquant/ |
| 945 | - vllm/v1/attention/backends/turboquant_attn.py |
| 946 | - vllm/v1/attention/ops/triton_turboquant_decode.py |
| 947 | - vllm/v1/attention/ops/triton_turboquant_store.py |
| 948 | commands: |
| 949 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/models-turboquant-t4nc.txt |
| 950 | mirror: |
| 951 | amd: |
| 952 | label: ":amd: (MI355 DPX) LM Eval TurboQuant t4nc" |
| 953 | dind: false |
| 954 | device: mi355_dpx |
| 955 | num_devices: 1 |
| 956 | timeout_in_minutes: 30 |
| 957 | working_dir: /vllm-workspace/tests |
| 958 | depends_on: |
| 959 | - image-build-amd |
| 960 | source_file_dependencies: |
| 961 | - .buildkite/test_areas/lm_eval.yaml |
| 962 | - tests/evals/gsm8k/ |
| 963 | - tests/utils.py |
| 964 | - tests/conftest.py |
| 965 | - vllm/model_executor/layers/quantization/turboquant/ |
| 966 | - vllm/v1/attention/backends/turboquant_attn.py |
| 967 | - vllm/v1/attention/backends/triton_attn.py |
| 968 | - vllm/v1/attention/ops/triton_unified_attention.py |
| 969 | - vllm/v1/attention/selector.py |
| 970 | - vllm/v1/attention/ops/triton_turboquant_decode.py |
| 971 | - vllm/v1/attention/ops/triton_turboquant_store.py |
| 972 | - vllm/v1/attention/ops/flydsl_turboquant_decode.py |
| 973 | - vllm/v1/attention/ops/flydsl_kernels/ |
| 974 | - vllm/v1/attention/ops/turboquant_soa/ |
| 975 | - vllm/platforms/rocm.py |
| 976 | commands: |
| 977 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/models-turboquant-t4nc-rocm.txt |
| 978 | |
| 979 | - label: ":nvidia: (H200 MIG 18GB) LM Eval TurboQuant k3v4nc" |
| 980 | key: lm-eval-turboquant-k3v4nc |
| 981 | timeout_in_minutes: 30 |
| 982 | device: h200_18gb |
| 983 | source_file_dependencies: |
| 984 | - vllm/model_executor/layers/quantization/turboquant/ |
| 985 | - vllm/v1/attention/backends/turboquant_attn.py |
| 986 | - vllm/v1/attention/ops/triton_turboquant_decode.py |
| 987 | - vllm/v1/attention/ops/triton_turboquant_store.py |
| 988 | commands: |
| 989 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/models-turboquant-k3v4nc.txt |
| 990 | mirror: |
| 991 | amd: |
| 992 | label: ":amd: (MI355 DPX) LM Eval TurboQuant k3v4nc" |
| 993 | dind: false |
| 994 | device: mi355_dpx |
| 995 | num_devices: 1 |
| 996 | timeout_in_minutes: 30 |
| 997 | working_dir: /vllm-workspace/tests |
| 998 | depends_on: |
| 999 | - image-build-amd |
| 1000 | source_file_dependencies: |
| 1001 | - .buildkite/test_areas/lm_eval.yaml |
| 1002 | - tests/evals/gsm8k/ |
| 1003 | - tests/utils.py |
| 1004 | - tests/conftest.py |
| 1005 | - vllm/model_executor/layers/quantization/turboquant/ |
| 1006 | - vllm/v1/attention/backends/turboquant_attn.py |
| 1007 | - vllm/v1/attention/backends/triton_attn.py |
| 1008 | - vllm/v1/attention/ops/triton_unified_attention.py |
| 1009 | - vllm/v1/attention/selector.py |
| 1010 | - vllm/v1/attention/ops/triton_turboquant_decode.py |
| 1011 | - vllm/v1/attention/ops/triton_turboquant_store.py |
| 1012 | - vllm/v1/attention/ops/flydsl_turboquant_decode.py |
| 1013 | - vllm/v1/attention/ops/flydsl_kernels/ |
| 1014 | - vllm/v1/attention/ops/turboquant_soa/ |
| 1015 | - vllm/platforms/rocm.py |
| 1016 | commands: |
| 1017 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/models-turboquant-k3v4nc-rocm.txt |
| 1018 | |
| 1019 | - label: ":nvidia: (H200 MIG 18GB) LM Eval TurboQuant t3nc" |
| 1020 | key: lm-eval-turboquant-t3nc |
| 1021 | timeout_in_minutes: 30 |
| 1022 | device: h200_18gb |
| 1023 | source_file_dependencies: |
| 1024 | - vllm/model_executor/layers/quantization/turboquant/ |
| 1025 | - vllm/v1/attention/backends/turboquant_attn.py |
| 1026 | - vllm/v1/attention/ops/triton_turboquant_decode.py |
| 1027 | - vllm/v1/attention/ops/triton_turboquant_store.py |
| 1028 | commands: |
| 1029 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/models-turboquant-t3nc.txt |
| 1030 | mirror: |
| 1031 | amd: |
| 1032 | label: ":amd: (MI355 DPX) LM Eval TurboQuant t3nc" |
| 1033 | dind: false |
| 1034 | device: mi355_dpx |
| 1035 | num_devices: 1 |
| 1036 | timeout_in_minutes: 30 |
| 1037 | working_dir: /vllm-workspace/tests |
| 1038 | depends_on: |
| 1039 | - image-build-amd |
| 1040 | source_file_dependencies: |
| 1041 | - .buildkite/test_areas/lm_eval.yaml |
| 1042 | - tests/evals/gsm8k/ |
| 1043 | - tests/utils.py |
| 1044 | - tests/conftest.py |
| 1045 | - vllm/model_executor/layers/quantization/turboquant/ |
| 1046 | - vllm/v1/attention/backends/turboquant_attn.py |
| 1047 | - vllm/v1/attention/backends/triton_attn.py |
| 1048 | - vllm/v1/attention/ops/triton_unified_attention.py |
| 1049 | - vllm/v1/attention/selector.py |
| 1050 | - vllm/v1/attention/ops/triton_turboquant_decode.py |
| 1051 | - vllm/v1/attention/ops/triton_turboquant_store.py |
| 1052 | - vllm/v1/attention/ops/flydsl_turboquant_decode.py |
| 1053 | - vllm/v1/attention/ops/flydsl_kernels/ |
| 1054 | - vllm/v1/attention/ops/turboquant_soa/ |
| 1055 | - vllm/platforms/rocm.py |
| 1056 | commands: |
| 1057 | - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/models-turboquant-t3nc-rocm.txt |
| 1058 | |
| 1059 | - label: ":nvidia: (H100) GPQA Eval (GPT-OSS)" |
| 1060 | key: gpqa-eval-gpt-oss-2xh100 |
| 1061 | timeout_in_minutes: 35 |
| 1062 | device: h100 |
| 1063 | optional: true |
| 1064 | num_devices: 2 |
| 1065 | source_file_dependencies: |
| 1066 | - csrc/ |
| 1067 | - vllm/model_executor/layers/quantization |
| 1068 | - tests/evals/gpt_oss/ |
| 1069 | commands: |
| 1070 | - pytest -s -v evals/gpt_oss/test_gpqa_correctness.py --config-list-file=configs/models-h100.txt |
| 1071 | |
| 1072 | - label: ":nvidia: (B200) GPQA Eval (GPT-OSS)" |
| 1073 | key: gpqa-eval-gpt-oss-2xb200 |
| 1074 | timeout_in_minutes: 30 |
| 1075 | device: b200-k8s |
| 1076 | optional: true |
| 1077 | num_devices: 2 |
| 1078 | source_file_dependencies: |
| 1079 | - csrc/ |
| 1080 | - vllm/model_executor/layers/quantization |
| 1081 | - tests/evals/gpt_oss/ |
| 1082 | commands: |
| 1083 | - pytest -s -v evals/gpt_oss/test_gpqa_correctness.py --config-list-file=configs/models-b200.txt |
| 1084 | mirror: |
| 1085 | amd: |
| 1086 | label: ":amd: (MI355) GPQA Eval (GPT-OSS)" |
| 1087 | dind: false |
| 1088 | device: mi355_2 |
| 1089 | timeout_in_minutes: 40 |
| 1090 | depends_on: |
| 1091 | - image-build-amd |
| 1092 | source_file_dependencies: |
| 1093 | - csrc/ |
| 1094 | - vllm/model_executor/layers/quantization |
| 1095 | - vllm/model_executor/models/ |
| 1096 | - vllm/model_executor/model_loader/ |
| 1097 | - vllm/v1/attention/backends/ |
| 1098 | - vllm/v1/attention/selector.py |
| 1099 | - vllm/model_executor/layers/fused_moe/ |
| 1100 | - tests/evals/gpt_oss/ |
| 1101 | - vllm/_aiter_ops.py |
| 1102 | - vllm/platforms/rocm.py |
| 1103 | commands: |
| 1104 | - pytest -s -v evals/gpt_oss/test_gpqa_correctness.py --config-list-file=configs/models-gfx950.txt |
| 1105 | |
| 1106 | - label: ":nvidia: (B200) Qwen3.8-Flash-Next-FP8 Accuracy Eval" |
| 1107 | key: accuracy-eval-qwen3-8-flash-next-fp8-b200 |
| 1108 | timeout_in_minutes: 60 |
| 1109 | device: b200-k8s |
| 1110 | optional: true |
| 1111 | num_devices: 4 |
| 1112 | source_file_dependencies: |
| 1113 | - csrc/libtorch_stable/gdn/ |
| 1114 | - tests/evals/qwen4_exp/ |
| 1115 | - vllm/model_executor/layers/mamba/ |
| 1116 | - vllm/models/qwen4_exp/ |
| 1117 | - vllm/transformers_utils/configs/qwen4_exp.py |
| 1118 | - vllm/v1/attention/backends/short_conv_attn.py |
| 1119 | - vllm/v1/spec_decode/qwen4_exp.py |
| 1120 | commands: |
| 1121 | - uv pip install --system 'evalscope==1.10.0' |
| 1122 | - pytest -s -v evals/qwen4_exp/test_accuracy.py --config-list-file=configs/models-b200.txt |
| 1123 | |
| 1124 | - label: ":nvidia: (H200) Qwen3.8-Flash-Next-FP8 Accuracy Eval" |
| 1125 | key: accuracy-eval-qwen3-8-flash-next-fp8-h200 |
| 1126 | timeout_in_minutes: 60 |
| 1127 | device: h200 |
| 1128 | optional: true |
| 1129 | num_devices: 4 |
| 1130 | source_file_dependencies: |
| 1131 | - csrc/libtorch_stable/gdn/ |
| 1132 | - tests/evals/qwen4_exp/ |
| 1133 | - vllm/model_executor/layers/mamba/ |
| 1134 | - vllm/models/qwen4_exp/ |
| 1135 | - vllm/transformers_utils/configs/qwen4_exp.py |
| 1136 | - vllm/v1/attention/backends/short_conv_attn.py |
| 1137 | - vllm/v1/spec_decode/qwen4_exp.py |
| 1138 | commands: |
| 1139 | - uv pip install --system 'evalscope==1.10.0' |
| 1140 | - pytest -s -v evals/qwen4_exp/test_accuracy.py --config-list-file=configs/models-h200.txt |
| 1141 | mirror: |
| 1142 | amd: |
| 1143 | label: ":amd: (MI355) Qwen3.8-Flash-Next-FP8 Accuracy Eval" |
| 1144 | dind: false |
| 1145 | device: mi355_4 |
| 1146 | num_devices: 4 |
| 1147 | timeout_in_minutes: 105 |
| 1148 | working_dir: /vllm-workspace/tests |
| 1149 | depends_on: |
| 1150 | - image-build-amd |
| 1151 | source_file_dependencies: |
| 1152 | - csrc/libtorch_stable/gdn/ |
| 1153 | - tests/evals/qwen4_exp/ |
| 1154 | - vllm/model_executor/layers/mamba/ |
| 1155 | - vllm/models/qwen4_exp/ |
| 1156 | - vllm/transformers_utils/configs/qwen4_exp.py |
| 1157 | - vllm/v1/attention/backends/short_conv_attn.py |
| 1158 | - vllm/platforms/rocm.py |
| 1159 | env: |
| 1160 | VLLM_WORKER_MULTIPROC_METHOD: "spawn" |
| 1161 | commands: |
| 1162 | - uv pip install --system 'evalscope==1.10.0' |
| 1163 | - pytest -s -v evals/qwen4_exp/test_accuracy.py --config-list-file=configs/models-rocm.txt |
| 1164 | |
| 1165 | - label: ":nvidia: (DGX) Spark GPQA Eval (GPT-OSS)" |
| 1166 | key: gpqa-eval-gpt-oss-spark |
| 1167 | timeout_in_minutes: 35 |
| 1168 | device: dgx-spark |
| 1169 | optional: true |
| 1170 | num_devices: 1 |
| 1171 | depends_on: |
| 1172 | - arm64-image-build |
| 1173 | source_file_dependencies: |
| 1174 | - csrc/ |
| 1175 | - vllm/model_executor/layers/quantization |
| 1176 | - tests/evals/gpt_oss/ |
| 1177 | commands: |
| 1178 | - pytest -s -v evals/gpt_oss/test_gpqa_correctness.py --config-list-file=configs/models-spark.txt |
| 1179 | |
| 1180 | - label: ":nvidia: (H200 MIG 35GB) LM Eval KV-Offload" |
| 1181 | key: kv-offload-small |
| 1182 | timeout_in_minutes: 30 |
| 1183 | device: h200_35gb |
| 1184 | source_file_dependencies: |
| 1185 | - vllm/distributed/kv_transfer/kv_connector/v1/offloading/ |
| 1186 | - vllm/distributed/kv_transfer/kv_connector/v1/simple_cpu_offload_connector.py |
| 1187 | - vllm/v1/kv_offload/ |
| 1188 | - vllm/v1/simple_kv_offload/ |
| 1189 | - tests/evals/gsm8k/test_gsm8k_offloading.py |
| 1190 | commands: |
| 1191 | - pytest -s -v evals/gsm8k/test_gsm8k_offloading.py -k "nemotron-h-8b or gemma-4-e4b-it" |
| 1192 | mirror: |
| 1193 | amd: |
| 1194 | label: ":amd: (MI300) LM Eval KV-Offload" |
| 1195 | dind: false |
| 1196 | device: mi300_1 |
| 1197 | timeout_in_minutes: 55 |
| 1198 | depends_on: |
| 1199 | - image-build-amd |
| 1200 | working_dir: /vllm-workspace/tests |
| 1201 | source_file_dependencies: |
| 1202 | - vllm/distributed/kv_transfer/kv_connector/v1/offloading/ |
| 1203 | - vllm/distributed/kv_transfer/kv_connector/v1/simple_cpu_offload_connector.py |
| 1204 | - vllm/v1/kv_offload/ |
| 1205 | - vllm/v1/simple_kv_offload/ |
| 1206 | - tests/evals/gsm8k/test_gsm8k_offloading.py |
| 1207 | - vllm/platforms/rocm.py |
| 1208 | |
| 1209 | - label: ":nvidia: (H100) LM Eval KV-Offload Medium" |
| 1210 | key: kv-offload-medium |
| 1211 | timeout_in_minutes: 45 |
| 1212 | device: h100 |
| 1213 | num_devices: 2 |
| 1214 | source_file_dependencies: |
| 1215 | - vllm/distributed/kv_transfer/kv_connector/v1/offloading/ |
| 1216 | - vllm/distributed/kv_transfer/kv_connector/v1/simple_cpu_offload_connector.py |
| 1217 | - vllm/v1/kv_offload/ |
| 1218 | - vllm/v1/simple_kv_offload/ |
| 1219 | - tests/evals/gsm8k/test_gsm8k_offloading.py |
| 1220 | commands: |
| 1221 | - pytest -s -v evals/gsm8k/test_gsm8k_offloading.py -k "qwen3.5-35b or deepseek-v2-lite" |
| 1222 | mirror: |
| 1223 | amd: |
| 1224 | label: ":amd: (MI300) LM Eval KV-Offload Medium" |
| 1225 | dind: false |
| 1226 | device: mi300_2 |
| 1227 | timeout_in_minutes: 70 |
| 1228 | depends_on: |
| 1229 | - image-build-amd |
| 1230 | working_dir: /vllm-workspace/tests |
| 1231 | source_file_dependencies: |
| 1232 | - vllm/distributed/kv_transfer/kv_connector/v1/offloading/ |
| 1233 | - vllm/distributed/kv_transfer/kv_connector/v1/simple_cpu_offload_connector.py |
| 1234 | - vllm/v1/kv_offload/ |
| 1235 | - vllm/v1/simple_kv_offload/ |
| 1236 | - tests/evals/gsm8k/test_gsm8k_offloading.py |
| 1237 | - vllm/platforms/rocm.py |
| 1238 | |
| 1239 | - label: ":nvidia: (H100) LM Eval KV-Offload Large" |
| 1240 | key: kv-offload-large |
| 1241 | timeout_in_minutes: 40 |
| 1242 | device: h100 |
| 1243 | num_devices: 4 |
| 1244 | source_file_dependencies: |
| 1245 | - vllm/distributed/kv_transfer/kv_connector/v1/offloading/ |
| 1246 | - vllm/distributed/kv_transfer/kv_connector/v1/simple_cpu_offload_connector.py |
| 1247 | - vllm/v1/kv_offload/ |
| 1248 | - vllm/v1/simple_kv_offload/ |
| 1249 | - tests/evals/gsm8k/test_gsm8k_offloading.py |
| 1250 | commands: |
| 1251 | - pytest -s -v evals/gsm8k/test_gsm8k_offloading.py -k "deepseek-v4-flash" |
| 1252 | mirror: |
| 1253 | amd: |
| 1254 | label: ":amd: (MI300) LM Eval KV-Offload Large" |
| 1255 | dind: false |
| 1256 | device: mi300_4 |
| 1257 | timeout_in_minutes: 65 |
| 1258 | depends_on: |
| 1259 | - image-build-amd |
| 1260 | working_dir: /vllm-workspace/tests |
| 1261 | source_file_dependencies: |
| 1262 | - vllm/distributed/kv_transfer/kv_connector/v1/offloading/ |
| 1263 | - vllm/distributed/kv_transfer/kv_connector/v1/simple_cpu_offload_connector.py |
| 1264 | - vllm/v1/kv_offload/ |
| 1265 | - vllm/v1/simple_kv_offload/ |
| 1266 | - tests/evals/gsm8k/test_gsm8k_offloading.py |
| 1267 | - vllm/platforms/rocm.py |
| 1268 | commands: |
| 1269 | - VLLM_ROCM_USE_AITER=1 pytest -s -v evals/gsm8k/test_gsm8k_offloading.py -k "deepseek-v4-flash" |
| 1270 | |
| 1271 | - label: ":nvidia: (H200 MIG 35GB) MRCR Eval Small Models" |
| 1272 | key: mrcr-eval-small-models |
| 1273 | device: h200_35gb |
| 1274 | timeout_in_minutes: 25 |
| 1275 | source_file_dependencies: |
| 1276 | - tests/evals/mrcr/ |
| 1277 | commands: |
| 1278 | - pytest -s -v evals/mrcr/test_mrcr_correctness.py --config-list-file=evals/mrcr/configs/models-small.txt |
| 1279 | mirror: |
| 1280 | amd: |
| 1281 | label: ":amd: (MI355 DPX) MRCR Eval Small Models" |
| 1282 | dind: false |
| 1283 | device: mi355_dpx |
| 1284 | timeout_in_minutes: 40 |
| 1285 | depends_on: |
| 1286 | - image-build-amd |
| 1287 | source_file_dependencies: |
| 1288 | - vllm/model_executor/models/ |
| 1289 | - vllm/model_executor/model_loader/ |
| 1290 | - vllm/v1/attention/backends/ |
| 1291 | - vllm/v1/attention/selector.py |
| 1292 | - vllm/_aiter_ops.py |
| 1293 | - vllm/platforms/rocm.py |
| 1294 | - tests/evals/mrcr/ |