A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 12, 06:00 UTC
Updated CI mirror 2025-11-25 (#29434)
Signed-off-by: Alexei V. Ivanov <alexei.ivanov@amd.com> Signed-off-by: Alexei-V-Ivanov-AMD <156011006+Alexei-V-Ivanov-AMD@users.noreply.github.com> Co-authored-by: Kevin H. Luu <khluu000@gmail.com>
Alexei-V-Ivanov-AMD committed
Dec 1, 2025 at 17:44 UTC
342c4f147266e7d01a8a6a64754df7e0f7f8991a
1 file changed
+80
-26
.buildkite/test-amd.yaml
+80
-26
@@ -39,9 +39,9 @@ steps:
39
# if this test fails, it means the nightly torch version is not compatible with some
40
# of the dependencies. Please check the error message and add the package to whitelist
41
# in /vllm/tools/pre_commit/generate_nightly_torch_test.py
42
- mirror_hardwares: [amdexperimental]
42
+ mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
43
agent_pool: mi325_1
44
- # grade: Blocking
44
+ grade: Blocking
45
soft_fail: true
46
source_file_dependencies:
47
- requirements/nightly_torch_test.txt
@@ -50,9 +50,9 @@ steps:
50
51
- label: Async Engine, Inputs, Utils, Worker Test # 10min
52
timeout_in_minutes: 15
53
- mirror_hardwares: [amdexperimental, amdproduction]
53
+ mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
54
agent_pool: mi325_1
55
- # grade: Blocking
55
+ grade: Blocking
56
source_file_dependencies:
57
- vllm/
58
- tests/multimodal
@@ -63,9 +63,9 @@ steps:
63
64
- label: Async Engine, Inputs, Utils, Worker, Config Test (CPU) # 15min
65
timeout_in_minutes: 20
66
- mirror_hardwares: [amdexperimental, amdproduction]
66
+ mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
67
agent_pool: mi325_1
68
- # grade: Blocking
68
+ grade: Blocking
69
source_file_dependencies:
70
- vllm/
71
- tests/test_inputs.py
@@ -115,9 +115,9 @@ steps:
115
- pytest -v -s basic_correctness/test_cpu_offload.py
116
117
- label: Entrypoints Unit Tests # 5min
118
- mirror_hardwares: [amdexperimental, amdproduction]
118
+ mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
119
agent_pool: mi325_1
120
- # grade: Blocking
120
+ grade: Blocking
121
timeout_in_minutes: 10
122
working_dir: "/vllm-workspace/tests"
123
fast_check: true
@@ -214,6 +214,7 @@ steps:
214
# test with internal dp
215
- python3 ../examples/offline_inference/data_parallel.py --enforce-eager
216
- TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/distributed/test_async_llm_dp.py
217
+ - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/distributed/test_eagle_dp.py
218
- TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/distributed/test_external_lb_dp.py
219
- TP_SIZE=1 DP_SIZE=4 pytest -v -s v1/distributed/test_internal_lb_dp.py
220
- TP_SIZE=1 DP_SIZE=4 pytest -v -s v1/distributed/test_hybrid_lb_dp.py
@@ -252,9 +253,9 @@ steps:
253
- torchrun --nproc-per-node=8 ../examples/offline_inference/torchrun_dp_example.py --tp-size=2 --pp-size=1 --dp-size=4 --enable-ep
254
255
- label: EPLB Algorithm Test # 5min
255
- mirror_hardwares: [amdexperimental, amdproduction]
256
+ mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
257
agent_pool: mi325_1
257
- # grade: Blocking
258
+ grade: Blocking
259
timeout_in_minutes: 15
260
working_dir: "/vllm-workspace/tests"
261
source_file_dependencies:
@@ -341,9 +342,9 @@ steps:
342
343
- label: V1 Test entrypoints # 35min
344
timeout_in_minutes: 50
344
- mirror_hardwares: [amdexperimental, amdproduction]
345
+ mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
346
agent_pool: mi325_1
346
- # grade: Blocking
347
+ grade: Blocking
348
source_file_dependencies:
349
- vllm/
350
- tests/v1
@@ -391,6 +392,20 @@ steps:
392
commands:
393
- pytest -v -s v1/attention
394
395
+- label: Batch Invariance Tests (H100) # 10min
396
+ mirror_hardwares: [amdexperimental]
397
+ agent_pool: mi325_1
398
+ timeout_in_minutes: 25
399
+ gpu: h100
400
+ source_file_dependencies:
401
+ - vllm/
402
+ - tests/v1/determinism/
403
+ commands:
404
+ - export VLLM_WORKER_MULTIPROC_METHOD=spawn
405
+ - pip install pytest-timeout pytest-forked
406
+ - pytest -v -s v1/determinism/test_batch_invariance.py
407
+ - pytest -v -s v1/determinism/test_rms_norm_batch_invariant.py
408
+
409
- label: V1 Test attention (B200) # 10min
410
timeout_in_minutes: 30
411
gpu: b200
@@ -401,9 +416,9 @@ steps:
416
- VLLM_DISABLE_FLASHINFER_PREFILL=1 pytest -v -s v1/attention # TODO: FI prefill is bugged and causes incorrectness, fix this
417
418
- label: V1 Test others (CPU) # 5 mins
404
- mirror_hardwares: [amdexperimental, amdproduction]
419
+ mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
420
agent_pool: mi325_1
406
- # grade: Blocking
421
+ grade: Blocking
422
source_file_dependencies:
423
- vllm/
424
- tests/v1
@@ -495,7 +510,7 @@ steps:
510
511
- label: PyTorch Compilation Unit Tests # 15min
512
timeout_in_minutes: 30
498
- mirror_hardwares: [amdexperimental]
513
+ mirror_hardwares: [amdexperimental, amdproduction]
514
agent_pool: mi325_1
515
# grade: Blocking
516
torch_nightly: true
@@ -512,7 +527,7 @@ steps:
527
528
- label: PyTorch Fullgraph Smoke Test # 15min
529
timeout_in_minutes: 30
515
- mirror_hardwares: [amdexperimental]
530
+ mirror_hardwares: [amdexperimental, amdproduction]
531
agent_pool: mi325_1
532
# grade: Blocking
533
torch_nightly: true
@@ -568,7 +583,7 @@ steps:
583
584
- label: Kernels Attention Test %N # 23min
585
timeout_in_minutes: 35
571
- mirror_hardwares: [amdexperimental]
586
+ mirror_hardwares: [amdexperimental, amdproduction]
587
agent_pool: mi325_8
588
# grade: Blocking
589
source_file_dependencies:
@@ -595,7 +610,7 @@ steps:
610
611
- label: Kernels MoE Test %N # 40min
612
timeout_in_minutes: 60
598
- mirror_hardwares: [amdexperimental]
613
+ mirror_hardwares: [amdexperimental, amdproduction]
614
agent_pool: mi325_8
615
# grade: Blocking
616
source_file_dependencies:
@@ -622,6 +637,26 @@ steps:
637
commands:
638
- pytest -v -s kernels/mamba
639
640
+- label: Kernels DeepGEMM Test (H100) # Nvidia-centric
641
+# Not replicating for CUTLAS & CuTe
642
+ timeout_in_minutes: 45
643
+ gpu: h100
644
+ num_gpus: 1
645
+ source_file_dependencies:
646
+ - tools/install_deepgemm.sh
647
+ - vllm/utils/deep_gemm.py
648
+ - vllm/model_executor/layers/fused_moe
649
+ - vllm/model_executor/layers/quantization
650
+ - tests/kernels/quantization/test_block_fp8.py
651
+ - tests/kernels/moe/test_deepgemm.py
652
+ - tests/kernels/moe/test_batched_deepgemm.py
653
+ - tests/kernels/attention/test_deepgemm_attention.py
654
+ commands:
655
+ - pytest -v -s kernels/quantization/test_block_fp8.py -k deep_gemm
656
+ - pytest -v -s kernels/moe/test_deepgemm.py
657
+ - pytest -v -s kernels/moe/test_batched_deepgemm.py
658
+ - pytest -v -s kernels/attention/test_deepgemm_attention.py
659
+
660
- label: Model Executor Test # 23min
661
timeout_in_minutes: 35
662
torch_nightly: true
@@ -1055,6 +1090,7 @@ steps:
1090
- pytest -v -s tests/kernels/moe/test_nvfp4_moe.py
1091
- pytest -v -s tests/kernels/moe/test_ocp_mx_moe.py
1092
- pytest -v -s tests/kernels/moe/test_flashinfer.py
1093
+ - pytest -v -s tests/kernels/moe/test_cutedsl_moe.py
1094
1095
- label: Blackwell Fusion and Compile Tests # 30 min
1096
timeout_in_minutes: 40
@@ -1064,11 +1100,19 @@ steps:
1100
- csrc/quantization/fp4/
1101
- vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
1102
- vllm/v1/attention/backends/flashinfer.py
1103
+ - vllm/v1/worker/
1104
+ - vllm/v1/cudagraph_dispatcher.py
1105
- vllm/compilation/
1106
# can affect pattern matching
1107
- vllm/model_executor/layers/layernorm.py
1108
- vllm/model_executor/layers/activation.py
1109
- vllm/model_executor/layers/quantization/input_quant_fp8.py
1110
+ - vllm/model_executor/layers/fused_moe/layer.py
1111
+ - tests/compile/test_fusion_attn.py
1112
+ - tests/compile/test_silu_mul_quant_fusion.py
1113
+ - tests/compile/distributed/test_fusion_all_reduce.py
1114
+ - tests/compile/distributed/test_fusions_e2e.py
1115
+ - tests/compile/fullgraph/test_full_graph.py
1116
commands:
1117
- nvidia-smi
1118
- pytest -v -s tests/compile/test_fusion_attn.py
@@ -1079,7 +1123,7 @@ steps:
1123
# Wrap with quotes to escape yaml
1124
- "pytest -v -s tests/compile/distributed/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm -k 'True and not +quant_fp8 and not +rms_norm'"
1125
# test_fp8_kv_scale_compile requires FlashAttention (not supported on default L4/L40)
1082
- - pytest -v -s tests/compile/distributed/test_full_graph.py::test_fp8_kv_scale_compile
1126
+ - pytest -v -s tests/compile/fullgraph/test_full_graph.py::test_fp8_kv_scale_compile
1127
1128
- label: Blackwell Fusion E2E Tests # 30 min
1129
timeout_in_minutes: 40
@@ -1101,7 +1145,7 @@ steps:
1145
commands:
1146
- nvidia-smi
1147
# Run all e2e fusion tests
1104
- - pytest -v -s tests/compile/test_fusions_e2e.py
1148
+ - pytest -v -s tests/compile/distributed/test_fusions_e2e.py
1149
1150
- label: ROCm GPT-OSS Eval
1151
timeout_in_minutes: 60
@@ -1216,6 +1260,7 @@ steps:
1260
- tests/v1/worker/test_worker_memory_snapshot.py
1261
commands:
1262
- TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_async_llm_dp.py
1263
+ - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_eagle_dp.py
1264
- TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_external_lb_dp.py
1265
- DP_SIZE=2 pytest -v -s v1/entrypoints/openai/test_multi_api_servers.py
1266
- pytest -v -s entrypoints/llm/test_collective_rpc.py
@@ -1251,7 +1296,7 @@ steps:
1296
1297
- label: Plugin Tests (2 GPUs) # 40min
1298
timeout_in_minutes: 60
1254
- mirror_hardwares: [amdexperimental]
1299
+ mirror_hardwares: [amdexperimental, amdproduction]
1300
agent_pool: mi325_2
1301
# grade: Blocking
1302
working_dir: "/vllm-workspace/tests"
@@ -1327,7 +1372,7 @@ steps:
1372
1373
- label: Weight Loading Multiple GPU Test # 33min
1374
timeout_in_minutes: 45
1330
- mirror_hardwares: [amdexperimental]
1375
+ mirror_hardwares: [amdexperimental, amdproduction]
1376
agent_pool: mi325_2
1377
# grade: Blocking
1378
working_dir: "/vllm-workspace/tests"
@@ -1432,7 +1477,7 @@ steps:
1477
- pytest -v -s tests/compile/distributed/test_fusion_all_reduce.py
1478
#- pytest -v -s tests/compile/distributed/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm
1479
- "pytest -v -s tests/compile/distributed/test_fusions_e2e.py -k 'not Llama-4'"
1435
- - pytest -v -s tests/compile/distributed/test_sequence_parallel.py
1480
+ - pytest -v -s tests/distributed/test_sequence_parallel.py
1481
- pytest -v -s tests/distributed/test_context_parallel.py
1482
- CUDA_VISIBLE_DEVICES=1,2 VLLM_ALL2ALL_BACKEND=deepep_high_throughput VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model Qwen/Qwen1.5-MoE-A2.7B --tp-size=1 --dp-size=2 --max-model-len 2048
1483
- pytest -v -s tests/v1/distributed/test_dbo.py
@@ -1464,7 +1509,7 @@ steps:
1509
- bash .buildkite/scripts/run-prime-rl-test.sh
1510
1511
- label: DeepSeek V2-Lite Accuracy
1467
- mirror_hardwares: [amdexperimental]
1512
+ mirror_hardwares: [amdexperimental, amdproduction]
1513
agent_pool: mi325_4
1514
# grade: Blocking
1515
timeout_in_minutes: 60
@@ -1475,8 +1520,8 @@ steps:
1520
commands:
1521
- bash .buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh 0.25 200 8010
1522
1478
-- label: Qwen3-30B-A3B-FP8-block Accuracy
1479
- mirror_hardwares: [amdexperimental]
1523
+- label: Qwen3-30B-A3B-FP8-block Accuracy (H100)
1524
+ mirror_hardwares: [amdexperimental, amdproduction]
1525
agent_pool: mi325_4
1526
# grade: Blocking
1527
timeout_in_minutes: 60
@@ -1486,3 +1531,12 @@ steps:
1531
working_dir: "/vllm-workspace"
1532
commands:
1533
- bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh 0.8 200 8020
1534
+
1535
+- label: Qwen3-30B-A3B-FP8-block Accuracy (B200)
1536
+ timeout_in_minutes: 60
1537
+ gpu: b200
1538
+ optional: true
1539
+ num_gpus: 2
1540
+ working_dir: "/vllm-workspace"
1541
+ commands:
1542
+ - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh 0.8 200 8020 2 1
\ No newline at end of file