| 1 | group: Compile |
| 2 | depends_on: |
| 3 | - image-build |
| 4 | steps: |
| 5 | - label: ":nvidia: (B200) Sequence Parallel Correctness" |
| 6 | key: sequence-parallel-correctness-tests-2xb200 |
| 7 | timeout_in_minutes: 45 |
| 8 | working_dir: "/vllm-workspace/" |
| 9 | device: b200-k8s |
| 10 | optional: true |
| 11 | num_devices: 2 |
| 12 | commands: |
| 13 | - export VLLM_TEST_CLEAN_GPU_MEMORY=1 |
| 14 | - pytest -v -s tests/compile/correctness_e2e/test_sequence_parallel.py |
| 15 | mirror: |
| 16 | amd: |
| 17 | label: ":amd: (MI355) Sequence Parallel Correctness" |
| 18 | dind: false |
| 19 | device: mi355_2 |
| 20 | num_devices: 2 |
| 21 | timeout_in_minutes: 180 |
| 22 | depends_on: |
| 23 | - image-build-amd |
| 24 | source_file_dependencies: |
| 25 | - .buildkite/test_areas/compile.yaml |
| 26 | - vllm/compilation/ |
| 27 | - vllm/config/ |
| 28 | - vllm/distributed/ |
| 29 | - vllm/model_executor/kernels/linear/ |
| 30 | - vllm/model_executor/layers/ |
| 31 | - vllm/model_executor/models/llama.py |
| 32 | - vllm/v1/attention/ |
| 33 | - vllm/v1/worker/ |
| 34 | - vllm/v1/cudagraph_dispatcher.py |
| 35 | - vllm/platforms/rocm.py |
| 36 | - tests/compile/correctness_e2e/test_sequence_parallel.py |
| 37 | - tests/models/registry.py |
| 38 | - tests/utils.py |
| 39 | |
| 40 | - label: ":nvidia: (B200) AsyncTP Correctness" |
| 41 | key: asynctp-correctness-tests-b200 |
| 42 | timeout_in_minutes: 30 |
| 43 | working_dir: "/vllm-workspace/" |
| 44 | device: b200-k8s |
| 45 | optional: true |
| 46 | num_devices: 2 |
| 47 | commands: |
| 48 | - export VLLM_TEST_CLEAN_GPU_MEMORY=1 |
| 49 | - pytest -v -s tests/compile/correctness_e2e/test_async_tp.py |
| 50 | mirror: |
| 51 | amd: |
| 52 | label: ":amd: (MI355) AsyncTP Correctness" |
| 53 | dind: false |
| 54 | device: mi355_2 |
| 55 | num_devices: 2 |
| 56 | timeout_in_minutes: 60 |
| 57 | working_dir: "/vllm-workspace/" |
| 58 | depends_on: |
| 59 | - image-build-amd |
| 60 | source_file_dependencies: |
| 61 | - .buildkite/test_areas/compile.yaml |
| 62 | - vllm/compilation/ |
| 63 | - vllm/config/ |
| 64 | - vllm/distributed/ |
| 65 | - vllm/model_executor/layers/ |
| 66 | - vllm/model_executor/kernels/linear/ |
| 67 | - vllm/model_executor/models/llama.py |
| 68 | - vllm/model_executor/models/qwen3.py |
| 69 | - vllm/v1/attention/ |
| 70 | - vllm/_aiter_ops.py |
| 71 | - vllm/platforms/rocm.py |
| 72 | - vllm/transformers_utils/repo_utils.py |
| 73 | - tests/compile/correctness_e2e/test_async_tp.py |
| 74 | - tests/compile/fusions_e2e/common.py |
| 75 | - tests/models/registry.py |
| 76 | - tests/utils.py |
| 77 | commands: |
| 78 | - export VLLM_TEST_CLEAN_GPU_MEMORY=1 |
| 79 | - pytest -v -s tests/compile/correctness_e2e/test_async_tp.py::test_rocm_async_tp_bf16_output_correctness |
| 80 | - pytest -v -s tests/compile/correctness_e2e/test_async_tp.py::test_async_tp_pass_correctness -k RedHatAI |
| 81 | - pytest -v -s tests/compile/correctness_e2e/test_async_tp.py::test_async_tp_pass_nvfp4_correctness |
| 82 | |
| 83 | - label: ":nvidia: (H100) Distributed Compile" |
| 84 | key: distributed-compile-unit-tests-2xh100 |
| 85 | timeout_in_minutes: 45 |
| 86 | working_dir: "/vllm-workspace/" |
| 87 | device: h100 |
| 88 | num_devices: 2 |
| 89 | source_file_dependencies: |
| 90 | - vllm/compilation/ |
| 91 | - vllm/model_executor/layers |
| 92 | - tests/compile/passes/distributed/ |
| 93 | commands: |
| 94 | - export VLLM_TEST_CLEAN_GPU_MEMORY=1 |
| 95 | - pytest -s -v tests/compile/passes/distributed |
| 96 | |
| 97 | - label: ":nvidia: (B200) Fusion and Compile" |
| 98 | key: fusion-and-compile-unit-tests-2xb200 |
| 99 | timeout_in_minutes: 30 |
| 100 | working_dir: "/vllm-workspace/" |
| 101 | device: b200-k8s |
| 102 | source_file_dependencies: |
| 103 | - csrc/quantization/fp4/ |
| 104 | - vllm/model_executor/layers/quantization/ |
| 105 | - vllm/model_executor/layers/fusion/ |
| 106 | - vllm/model_executor/layers/layernorm.py |
| 107 | - vllm/model_executor/layers/activation.py |
| 108 | - vllm/model_executor/layers/attention/attention.py |
| 109 | - vllm/v1/attention/backends/flashinfer.py |
| 110 | - vllm/compilation/ # TODO(luka) limit to vllm/compilation/passes |
| 111 | - tests/compile/passes/test_fusion_attn.py |
| 112 | - tests/compile/passes/test_mla_attn_quant_fusion.py |
| 113 | - tests/compile/passes/test_silu_mul_quant_fusion.py |
| 114 | - tests/compile/passes/test_silu_mul_quant_manual_fusion.py |
| 115 | - tests/compile/passes/distributed/test_fusion_all_reduce.py |
| 116 | - tests/compile/fullgraph/test_full_graph.py |
| 117 | commands: |
| 118 | # b200 runners are limited, so we limit the tests to the minimum set only supported on Blackwell |
| 119 | - nvidia-smi |
| 120 | - pytest -v -s tests/compile/passes/test_fusion_attn.py -k FLASHINFER |
| 121 | - pytest -v -s tests/compile/passes/test_mla_attn_quant_fusion.py |
| 122 | - pytest -v -s tests/compile/passes/test_silu_mul_quant_fusion.py |
| 123 | - pytest -v -s tests/compile/passes/test_silu_mul_quant_manual_fusion.py |
| 124 | # this runner has 2 GPUs available even though num_devices=2 is not set |
| 125 | - pytest -v -s tests/compile/passes/distributed/test_fusion_all_reduce.py |
| 126 | mirror: |
| 127 | amd: |
| 128 | label: ":amd: (MI355) Fusion and Compile FP8" |
| 129 | dind: false |
| 130 | device: mi355_2 |
| 131 | num_devices: 2 |
| 132 | working_dir: "/vllm-workspace/" |
| 133 | timeout_in_minutes: 45 |
| 134 | depends_on: |
| 135 | - image-build-amd |
| 136 | source_file_dependencies: |
| 137 | - vllm/compilation/ |
| 138 | - vllm/model_executor/layers/ |
| 139 | - vllm/v1/attention/ |
| 140 | - vllm/distributed/ |
| 141 | - vllm/_aiter_ops.py |
| 142 | - tests/compile/passes/ |
| 143 | - tests/compile/correctness_e2e/test_attn_quant.py |
| 144 | - vllm/model_executor/kernels/linear/ |
| 145 | - vllm/model_executor/models/llama.py |
| 146 | commands: |
| 147 | - amd-smi |
| 148 | # MLA static FP8 and distributed AR/RMS are owned by the existing ROCm jobs. |
| 149 | - pytest -v -s tests/compile/passes/test_fusion_attn.py tests/compile/passes/test_silu_mul_quant_fusion.py tests/compile/passes/test_silu_mul_quant_manual_fusion.py |
| 150 | - pytest -v -s tests/compile/correctness_e2e/test_attn_quant.py |
| 151 | |
| 152 | - label: ":nvidia: (H100) Fusion E2E Quick" |
| 153 | key: fusion-e2e-quick-h100 |
| 154 | timeout_in_minutes: 25 |
| 155 | working_dir: "/vllm-workspace/" |
| 156 | device: h100 |
| 157 | num_devices: 1 |
| 158 | source_file_dependencies: |
| 159 | - csrc/quantization/ |
| 160 | - vllm/model_executor/ |
| 161 | - vllm/v1/attention/ |
| 162 | - vllm/compilation/ |
| 163 | - tests/compile/fusions_e2e/ |
| 164 | commands: |
| 165 | - nvidia-smi |
| 166 | # Run all models and attn backends but only Inductor partition and native custom ops |
| 167 | - pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k "inductor_partition and not +rms_norm and not +quant_fp8" |
| 168 | # Qwen/Deepseek requires +quant_fp8 as -quant_fp8 rms+quant fusion is not supported |
| 169 | - pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k "inductor_partition and not +rms_norm and +quant_fp8 and (qwen3 or deepseek)" |
| 170 | mirror: |
| 171 | amd: |
| 172 | label: ":amd: (MI355 DPX) Fusion E2E Quick" |
| 173 | dind: false |
| 174 | device: mi355_dpx |
| 175 | working_dir: "/vllm-workspace/" |
| 176 | timeout_in_minutes: 30 |
| 177 | depends_on: |
| 178 | - image-build-amd |
| 179 | source_file_dependencies: |
| 180 | - csrc/quantization/ |
| 181 | - vllm/model_executor/ |
| 182 | - vllm/v1/attention/ |
| 183 | - vllm/compilation/ |
| 184 | - tests/compile/fusions_e2e/ |
| 185 | - vllm/_aiter_ops.py |
| 186 | - vllm/platforms/rocm.py |
| 187 | commands: |
| 188 | - rocm-smi |
| 189 | - "pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k 'inductor_partition and not +rms_norm and not +quant_fp8'" |
| 190 | - "pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k 'inductor_partition and +rms_norm and +quant_fp8 and qwen3'" |
| 191 | |
| 192 | - label: ":nvidia: (H100) Fusion E2E Config Sweep" |
| 193 | key: fusion-e2e-config-sweep-h100 |
| 194 | timeout_in_minutes: 25 |
| 195 | working_dir: "/vllm-workspace/" |
| 196 | device: h100 |
| 197 | num_devices: 1 |
| 198 | source_file_dependencies: |
| 199 | - csrc/quantization/ |
| 200 | - vllm/compilation/ |
| 201 | # can affect pattern matching |
| 202 | - vllm/model_executor/layers/layernorm.py |
| 203 | - vllm/model_executor/layers/activation.py |
| 204 | - vllm/model_executor/layers/attention/attention.py |
| 205 | - vllm/model_executor/layers/quantization/input_quant_fp8.py |
| 206 | - tests/compile/fusions_e2e/ |
| 207 | commands: |
| 208 | - nvidia-smi |
| 209 | # Run just llama3 (fp8) for all config combinations |
| 210 | - pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k "llama-3" |
| 211 | mirror: |
| 212 | amd: |
| 213 | label: ":amd: (MI355 DPX) Fusion E2E Config Sweep" |
| 214 | dind: false |
| 215 | device: mi355_dpx |
| 216 | working_dir: "/vllm-workspace/" |
| 217 | timeout_in_minutes: 40 |
| 218 | depends_on: |
| 219 | - image-build-amd |
| 220 | source_file_dependencies: |
| 221 | - csrc/quantization/ |
| 222 | - vllm/compilation/ |
| 223 | - vllm/model_executor/layers/layernorm.py |
| 224 | - vllm/model_executor/layers/activation.py |
| 225 | - vllm/model_executor/layers/attention/attention.py |
| 226 | - vllm/model_executor/layers/quantization/input_quant_fp8.py |
| 227 | - tests/compile/fusions_e2e/ |
| 228 | - vllm/_aiter_ops.py |
| 229 | - vllm/platforms/rocm.py |
| 230 | commands: |
| 231 | - rocm-smi |
| 232 | - pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k "llama-3" |
| 233 | |
| 234 | - label: ":nvidia: (B200) Fusion E2E Config Sweep" |
| 235 | key: fusion-e2e-config-sweep-b200 |
| 236 | timeout_in_minutes: 30 |
| 237 | working_dir: "/vllm-workspace/" |
| 238 | device: b200-k8s |
| 239 | num_devices: 1 |
| 240 | optional: true |
| 241 | commands: |
| 242 | - nvidia-smi |
| 243 | # Run all models but only FLASHINFER, Inductor partition and native custom ops |
| 244 | # Qwen/Deepseek requires +quant_fp8 as -quant_fp8 rms+quant fusion is not supported |
| 245 | # Run just llama3 (fp8 & fp4) for all config combinations (only inductor partition) |
| 246 | - pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k "inductor_partition and (FLASHINFER and not +rms_norm and (not +quant_fp8 or +quant_fp8 and (qwen3 or deepseek)) or llama-3)" |
| 247 | mirror: |
| 248 | amd: |
| 249 | label: ":amd: (MI355 DPX) Fusion E2E FP8 Config Sweep" |
| 250 | dind: false |
| 251 | device: mi355_dpx |
| 252 | timeout_in_minutes: 45 |
| 253 | working_dir: "/vllm-workspace/" |
| 254 | depends_on: |
| 255 | - image-build-amd |
| 256 | source_file_dependencies: |
| 257 | - .buildkite/test_areas/compile.yaml |
| 258 | - csrc/quantization/ |
| 259 | - vllm/compilation/ |
| 260 | - vllm/model_executor/ |
| 261 | - vllm/v1/attention/ |
| 262 | - vllm/_aiter_ops.py |
| 263 | - vllm/platforms/rocm.py |
| 264 | - tests/compile/fusions_e2e/ |
| 265 | commands: |
| 266 | - rocm-smi |
| 267 | - pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k "test_tp1_fp8_fusions and inductor_partition and not +rms_norm and not True and ((Llama-3 and TRITON_ATTN) or (Qwen3-30B and +quant_fp8 and TRITON_ATTN) or (DeepSeek-Coder and +quant_fp8 and TRITON_MLA))" |
| 268 | |
| 269 | - label: ":nvidia: (H100) Fusion E2E TP2 Quick" |
| 270 | key: fusion-e2e-tp2-quick-h100 |
| 271 | timeout_in_minutes: 35 |
| 272 | working_dir: "/vllm-workspace/" |
| 273 | device: h100 |
| 274 | num_devices: 2 |
| 275 | source_file_dependencies: |
| 276 | - csrc/quantization/ |
| 277 | - vllm/model_executor/ |
| 278 | - vllm/v1/attention/ |
| 279 | - vllm/compilation/ |
| 280 | - tests/compile/fusions_e2e/ |
| 281 | commands: |
| 282 | - nvidia-smi |
| 283 | # Run all models and attn backends but only Inductor partition and native custom ops |
| 284 | - pytest -v -s tests/compile/fusions_e2e/test_tp2_ar_rms.py -k "inductor_partition and not +rms_norm and (not +quant_fp8 or +quant_fp8 and (qwen3 or deepseek))" |
| 285 | - pytest -v -s tests/compile/fusions_e2e/test_tp2_async_tp.py -k "inductor_partition and not +rms_norm and (not +quant_fp8 or +quant_fp8 and (qwen3 or deepseek))" |
| 286 | mirror: |
| 287 | amd: |
| 288 | label: ":amd: (MI355) Fusion E2E TP2 Quick" |
| 289 | dind: false |
| 290 | device: mi355_2 |
| 291 | timeout_in_minutes: 60 |
| 292 | depends_on: |
| 293 | - image-build-amd |
| 294 | working_dir: /vllm-workspace/ |
| 295 | source_file_dependencies: |
| 296 | - csrc/quantization/ |
| 297 | - vllm/compilation/ |
| 298 | - vllm/distributed/ |
| 299 | - vllm/model_executor/ |
| 300 | - vllm/v1/attention/ |
| 301 | - tests/compile/fusions_e2e/ |
| 302 | - vllm/_aiter_ops.py |
| 303 | - vllm/platforms/rocm.py |
| 304 | commands: |
| 305 | - export VLLM_TEST_CLEAN_GPU_MEMORY=1 |
| 306 | - >- |
| 307 | VLLM_ROCM_USE_AITER=1 |
| 308 | VLLM_ROCM_USE_AITER_CUSTOM_AR=1 |
| 309 | pytest -v -s tests/compile/fusions_e2e/test_tp2_ar_rms.py |
| 310 | -k 'test_tp2_ar_rms_fp8_fusions and inductor_partition and |
| 311 | not +quant_fp8 and not +rms_norm and Llama-3 and |
| 312 | ROCM_AITER_UNIFIED_ATTN' |
| 313 | - >- |
| 314 | VLLM_ROCM_USE_AITER=1 |
| 315 | VLLM_ROCM_USE_AITER_CUSTOM_AR=1 |
| 316 | pytest -v -s tests/compile/fusions_e2e/test_tp2_ar_rms.py |
| 317 | -k 'test_tp2_ar_rms_fp8_fusions and inductor_partition and |
| 318 | +quant_fp8 and +rms_norm and qwen3 and |
| 319 | ROCM_AITER_UNIFIED_ATTN' |
| 320 | - >- |
| 321 | VLLM_ROCM_USE_AITER=1 |
| 322 | VLLM_ROCM_USE_AITER_CUSTOM_AR=1 |
| 323 | pytest -v -s tests/compile/fusions_e2e/test_tp2_ar_rms.py |
| 324 | -k 'test_tp2_ar_rms_fp8_fusions and inductor_partition and |
| 325 | not +quant_fp8 and not +rms_norm and DeepSeek-Coder and |
| 326 | ROCM_AITER_MLA' |
| 327 | |
| 328 | - label: ":nvidia: (H100) Fusion E2E TP2 AR-RMS Config Sweep" |
| 329 | key: fusion-e2e-tp2-ar-rms-config-sweep-h100 |
| 330 | timeout_in_minutes: 30 |
| 331 | working_dir: "/vllm-workspace/" |
| 332 | device: h100 |
| 333 | num_devices: 2 |
| 334 | source_file_dependencies: |
| 335 | - csrc/quantization/ |
| 336 | - vllm/compilation/ |
| 337 | # can affect pattern matching |
| 338 | - vllm/model_executor/layers/layernorm.py |
| 339 | - vllm/model_executor/layers/activation.py |
| 340 | - vllm/model_executor/layers/attention/attention.py |
| 341 | - vllm/model_executor/layers/quantization/input_quant_fp8.py |
| 342 | - tests/compile/fusions_e2e/ |
| 343 | commands: |
| 344 | - nvidia-smi |
| 345 | # Run just llama3 (fp8 & bf16) for all config combinations |
| 346 | - pytest -v -s tests/compile/fusions_e2e/test_tp2_ar_rms.py -k "llama-3" |
| 347 | mirror: |
| 348 | amd: |
| 349 | label: ":amd: (MI355) Fusion E2E TP2 AR-RMS FP8 Config Sweep" |
| 350 | dind: false |
| 351 | device: mi355_2 |
| 352 | num_devices: 2 |
| 353 | timeout_in_minutes: 45 |
| 354 | working_dir: "/vllm-workspace/" |
| 355 | depends_on: |
| 356 | - image-build-amd |
| 357 | source_file_dependencies: |
| 358 | - vllm/compilation/ |
| 359 | - vllm/model_executor/ |
| 360 | - vllm/distributed/ |
| 361 | - vllm/v1/attention/ |
| 362 | - vllm/_aiter_ops.py |
| 363 | - vllm/platforms/rocm.py |
| 364 | - tests/compile/fusions_e2e/ |
| 365 | - tests/utils.py |
| 366 | - tests/conftest.py |
| 367 | - csrc/quantization/ |
| 368 | commands: |
| 369 | # Existing MI300 jobs own BF16; TP2 Quick owns the excluded Llama cell. |
| 370 | - amd-smi |
| 371 | - VLLM_ROCM_USE_AITER=1 VLLM_ROCM_USE_AITER_CUSTOM_AR=1 pytest -v -s tests/compile/fusions_e2e/test_tp2_ar_rms.py -k 'test_tp2_ar_rms_fp8_fusions and Llama-3 and ROCM_AITER_UNIFIED_ATTN and (dynamo_partition or +quant_fp8 or +rms_norm)' |
| 372 | |
| 373 | - label: ":nvidia: (H100) Fusion E2E TP2 AsyncTP Config Sweep" |
| 374 | key: fusion-e2e-tp2-asynctp-config-sweep-h100 |
| 375 | timeout_in_minutes: 40 |
| 376 | working_dir: "/vllm-workspace/" |
| 377 | device: h100 |
| 378 | num_devices: 2 |
| 379 | source_file_dependencies: |
| 380 | - csrc/quantization/ |
| 381 | - vllm/compilation/ |
| 382 | # can affect pattern matching |
| 383 | - vllm/model_executor/layers/layernorm.py |
| 384 | - vllm/model_executor/layers/activation.py |
| 385 | - vllm/model_executor/layers/attention/attention.py |
| 386 | - vllm/model_executor/layers/quantization/input_quant_fp8.py |
| 387 | - tests/compile/fusions_e2e/ |
| 388 | commands: |
| 389 | - nvidia-smi |
| 390 | # Run just llama3 (fp8 & bf16) for all config combinations |
| 391 | - pytest -v -s tests/compile/fusions_e2e/test_tp2_async_tp.py -k "llama-3" |
| 392 | mirror: |
| 393 | amd: |
| 394 | label: ':amd: (MI355) Fusion E2E TP2 AsyncTP BF16 Config Sweep' |
| 395 | dind: false |
| 396 | device: mi355_2 |
| 397 | num_devices: 2 |
| 398 | working_dir: /vllm-workspace/ |
| 399 | timeout_in_minutes: 30 |
| 400 | depends_on: |
| 401 | - image-build-amd |
| 402 | source_file_dependencies: |
| 403 | - vllm/compilation/ |
| 404 | - vllm/model_executor/layers/linear.py |
| 405 | - vllm/model_executor/layers/utils.py |
| 406 | - vllm/distributed/ |
| 407 | - vllm/v1/attention/backends/triton_attn.py |
| 408 | - tests/compile/fusions_e2e/ |
| 409 | - tests/compile/passes/distributed/test_async_tp.py |
| 410 | - tests/compile/rocm/test_async_tp.py |
| 411 | - tests/compile/correctness_e2e/test_async_tp.py |
| 412 | - vllm/model_executor/kernels/linear/ |
| 413 | - vllm/model_executor/models/qwen3.py |
| 414 | commands: |
| 415 | - rocm-smi |
| 416 | - pytest -v -s tests/compile/rocm/test_async_tp.py |
| 417 | - pytest -v -s tests/compile/fusions_e2e/test_tp2_rocm_async_tp.py |
| 418 | - pytest -v -s tests/compile/correctness_e2e/test_async_tp.py::test_rocm_async_tp_bf16_output_correctness |
| 419 | |
| 420 | - label: ":nvidia: (B200) Fusion E2E TP2" |
| 421 | key: fusion-e2e-tp2-b200 |
| 422 | timeout_in_minutes: 45 |
| 423 | working_dir: "/vllm-workspace/" |
| 424 | device: b200-k8s |
| 425 | num_devices: 2 |
| 426 | source_file_dependencies: |
| 427 | - csrc/quantization/ |
| 428 | - vllm/model_executor/ |
| 429 | - vllm/v1/attention/ |
| 430 | - vllm/compilation/ |
| 431 | - tests/compile/fusions_e2e/ |
| 432 | commands: |
| 433 | - nvidia-smi |
| 434 | # Run all models but only FLASHINFER, Inductor partition and native custom ops |
| 435 | # include qwen/deepseek with +quant_fp8 as -quant_fp8 rms+quant fusion is not supported |
| 436 | # for ar-rms-quant-fp4, also sweep llama3 |
| 437 | - pytest -v -s tests/compile/fusions_e2e/test_tp2_ar_rms.py -k "(FLASHINFER and inductor_partition and not +rms_norm and (not +quant_fp8 or +quant_fp8 and (qwen3 or deepseek))) or Llama-3.1-8B-Instruct-FP4" |
| 438 | - pytest -v -s tests/compile/fusions_e2e/test_tp2_async_tp.py -k "FLASHINFER and inductor_partition and not +rms_norm and (not +quant_fp8 or +quant_fp8 and (qwen3 or deepseek))" |
| 439 | mirror: |
| 440 | amd: |
| 441 | label: ":amd: (MI355) Fusion E2E TP2 FP8 BF16 rewrites" |
| 442 | dind: false |
| 443 | device: mi355_2 |
| 444 | num_devices: 2 |
| 445 | timeout_in_minutes: 45 |
| 446 | working_dir: "/vllm-workspace/" |
| 447 | depends_on: |
| 448 | - image-build-amd |
| 449 | source_file_dependencies: |
| 450 | - vllm/compilation/ |
| 451 | - vllm/model_executor/ |
| 452 | - vllm/distributed/ |
| 453 | - vllm/v1/attention/ |
| 454 | - vllm/_aiter_ops.py |
| 455 | - vllm/platforms/rocm.py |
| 456 | - tests/compile/fusions_e2e/ |
| 457 | - tests/utils.py |
| 458 | - tests/conftest.py |
| 459 | - csrc/quantization/ |
| 460 | commands: |
| 461 | # Dummy-weight rewrite coverage; trained FP8 AR/RMS equivalence is unresolved. |
| 462 | # TODO: FP4 output-quant and trained MXFP4 AR/RMS fusion remain unsupported. |
| 463 | - amd-smi |
| 464 | - VLLM_ROCM_USE_AITER=1 VLLM_ROCM_USE_AITER_CUSTOM_AR=1 pytest -v -s tests/compile/fusions_e2e/test_tp2_ar_rms.py -k 'test_tp2_ar_rms_fp8_fusions and qwen3 and ROCM_AITER_UNIFIED_ATTN and +quant_fp8 and (dynamo_partition or not +rms_norm)' |
| 465 | - pytest -v -s tests/compile/fusions_e2e/test_tp2_rocm_moe.py |