3b6c0bdae9 · tduncan/Nvidia-Optimization-Stack-VLLM

@@ -106,7 +106,7 @@ steps:

106 key: kernels-attention-test

107 device: l4

108 num_devices: 1

109 - timeout_in_minutes: 65

109 + timeout_in_minutes: 30

110 source_file_dependencies:

111 - csrc/attention/

112 - vllm/v1/attention

@@ -116,7 +116,7 @@ steps:

116 - tests/kernels/attention

117 commands:

118 - pytest -v -s kernels/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT

119 - parallelism: 2

119 + parallelism: 7

120 mirror:

121 amd:

122 label: ":amd: (MI300) Attention Kernels Shard %N"

@@ -169,19 +169,22 @@ steps:

169 key: kernels-quantization-test

170 device: l4

171 num_devices: 1

172 - timeout_in_minutes: 60

172 + timeout_in_minutes: 30

173 source_file_dependencies:

174 - csrc/quantization/

175 - vllm/model_executor/layers/quantization

176 - tests/kernels/quantization

177 commands:

178 - pytest -v -s kernels/quantization --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT

179 - parallelism: 2

179 + parallelism: 6

180 mirror:

181 amd:

182 label: ":amd: (MI300) Quantization Kernels Shard %N"

183 dind: false

184 device: mi300_1

185 + # Pin the AMD mirror at its current 2 shards so raising NVIDIA

186 + # parallelism does not multiply AMD copies (needs ci-infra #473).

187 + parallelism: 2

188 timeout_in_minutes: 120

189 source_file_dependencies:

190 - csrc/quantization/

@@ -277,9 +280,10 @@ steps:

280 - pytest -v -s kernels/attention/test_deepgemm_attention.py

281 - pytest -v -s quantization/test_cutlass_w4a16.py

282

280 -- label: ":nvidia: (B200) Kernels"

283 +- label: ":nvidia: (B200) Kernels Shard %N"

284 key: kernels-b200

282 - timeout_in_minutes: 80

285 + timeout_in_minutes: 30

286 + parallelism: 3

287 working_dir: "/vllm-workspace/"

288 device: b200-k8s

289 # optional: true

@@ -323,39 +327,49 @@ steps:

327 - tests/kernels/test_top_k_per_row.py

328 commands:

329 - nvidia-smi

326 - - python3 examples/basic/offline_inference/chat.py

327 - # Attention

328 - # num_heads2 broken by https://github.com/flashinfer-ai/flashinfer/issues/1353

329 - - pytest -v -s tests/kernels/attention/test_attention_selector.py

330 - - pytest -v -s tests/kernels/attention/test_flashinfer.py -k 'not num_heads2'

331 - - pytest -v -s tests/kernels/attention/test_flashinfer_trtllm_attention.py

332 - - pytest -v -s tests/kernels/attention/test_cutlass_mla_decode.py

333 - - pytest -v -s tests/kernels/attention/test_flashinfer_mla_decode.py

334 - - pytest -v -s tests/kernels/attention/test_minimax_m3_msa_cutlass_sparse_decode.py

335 - - pytest -v -s tests/kernels/test_fused_minimax_m3_qknorm_rope_kv_insert.py

336 - - pytest -v -s tests/kernels/test_top_k_per_row.py

337 - # Quantization

338 - - pytest -v -s tests/kernels/quantization/test_cutlass_scaled_mm.py -k 'fp8'

339 - - pytest -v -s tests/kernels/quantization/test_nvfp4_quant.py

340 - - pytest -v -s tests/kernels/quantization/test_silu_mul_nvfp4_quant.py

341 - - pytest -v -s tests/kernels/quantization/test_nvfp4_scaled_mm.py

342 - - pytest -v -s tests/kernels/quantization/test_flashinfer_scaled_mm.py

343 - - pytest -v -s tests/kernels/quantization/test_flashinfer_nvfp4_scaled_mm.py

344 - - pytest -v -s tests/kernels/quantization/test_nvfp4_qutlass.py

345 - - pytest -v -s tests/kernels/quantization/test_mxfp4_qutlass.py

346 - - pytest -v -s tests/kernels/moe/test_nvfp4_moe.py

347 - - pytest -v -s tests/kernels/moe/test_mxfp4_moe.py

348 - - pytest -v -s tests/kernels/moe/test_ocp_mx_moe.py

349 - - pytest -v -s tests/kernels/moe/test_flashinfer.py

350 - - pytest -v -s tests/kernels/moe/test_flashinfer_moe.py

351 - - pytest -v -s tests/kernels/moe/test_trtllm_nvfp4_moe.py

352 - - pytest -v -s tests/kernels/moe/test_cutedsl_moe.py

353 - - pytest -v -s tests/kernels/mamba/test_gdn_prefill_cutedsl.py

354 - - pytest -v -s tests/kernels/test_bf16x3_router_gemm_cutedsl.py

355 - - pytest -v -s tests/kernels/attention/test_minimax_m3.py

356 - - pytest -v -s tests/kernels/test_ll_bf16_gemm.py

357 - # e2e

358 - - pytest -v -s tests/models/quantization/test_nvfp4.py

330 + - |

331 + set -eu

332 + case "$$BUILDKITE_PARALLEL_JOB" in

333 + 0)

334 + pytest -v -s tests/kernels/attention/test_attention_selector.py

335 + pytest -v -s tests/kernels/attention/test_flashinfer_trtllm_attention.py

336 + pytest -v -s tests/kernels/attention/test_flashinfer_mla_decode.py

337 + pytest -v -s tests/kernels/test_top_k_per_row.py

338 + pytest -v -s tests/kernels/quantization/test_silu_mul_nvfp4_quant.py

339 + pytest -v -s tests/kernels/quantization/test_flashinfer_scaled_mm.py

340 + pytest -v -s tests/kernels/moe/test_nvfp4_moe.py

341 + pytest -v -s tests/kernels/moe/test_ocp_mx_moe.py

342 + pytest -v -s tests/kernels/test_bf16x3_router_gemm_cutedsl.py

343 + pytest -v -s tests/kernels/test_ll_bf16_gemm.py

344 + ;;

345 + 1)

346 + python3 examples/basic/offline_inference/chat.py

347 + pytest -v -s tests/kernels/attention/test_minimax_m3_msa_cutlass_sparse_decode.py

348 + pytest -v -s tests/kernels/test_fused_minimax_m3_qknorm_rope_kv_insert.py

349 + pytest -v -s tests/kernels/quantization/test_flashinfer_nvfp4_scaled_mm.py

350 + pytest -v -s tests/kernels/quantization/test_nvfp4_qutlass.py

351 + pytest -v -s tests/kernels/moe/test_mxfp4_moe.py

352 + pytest -v -s tests/kernels/moe/test_flashinfer_moe.py

353 + pytest -v -s tests/kernels/moe/test_cutedsl_moe.py

354 + pytest -v -s tests/models/quantization/test_nvfp4.py

355 + ;;

356 + 2)

357 + pytest -v -s tests/kernels/attention/test_flashinfer.py -k 'not num_heads2'

358 + pytest -v -s tests/kernels/attention/test_cutlass_mla_decode.py

359 + pytest -v -s tests/kernels/quantization/test_cutlass_scaled_mm.py -k 'fp8'

360 + pytest -v -s tests/kernels/quantization/test_nvfp4_quant.py

361 + pytest -v -s tests/kernels/quantization/test_nvfp4_scaled_mm.py

362 + pytest -v -s tests/kernels/quantization/test_mxfp4_qutlass.py

363 + pytest -v -s tests/kernels/moe/test_flashinfer.py

364 + pytest -v -s tests/kernels/moe/test_trtllm_nvfp4_moe.py

365 + pytest -v -s tests/kernels/mamba/test_gdn_prefill_cutedsl.py

366 + pytest -v -s tests/kernels/attention/test_minimax_m3.py

367 + ;;

368 + *)

369 + echo "unexpected BUILDKITE_PARALLEL_JOB=$$BUILDKITE_PARALLEL_JOB" >&2

370 + exit 1

371 + ;;

372 + esac

373

374 - label: ":nvidia: (DGX) Spark b12x Linear Kernels Nightly"

375 key: b12x-linear-kernels-dgx-spark-nightly

[CI] Shard long kernel test groups (#54754)