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 7, 11:30 UTC
[CI] Defining extended V1 e2e + engine tests (#35580)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Andreas Karatzas committed
Mar 2, 2026 at 02:10 UTC
ec27b36b4b17ab51fe5a9fed4b0fbd4b39123058
3 files changed
+63
-5
.buildkite/test-amd.yaml
+29
-3
@@ -388,9 +388,7 @@ steps:
388
- label: V1 Test e2e + engine # 65min
389
timeout_in_minutes: 90
390
mirror_hardwares: [amdexperimental, amdproduction]
391
- # The test uses 4 GPUs, but we schedule it on 8-GPU machines for stability.
392
- # See discussion here: https://github.com/vllm-project/vllm/pull/31040
393
- agent_pool: mi325_8
391
+ agent_pool: mi325_1
392
optional: true
393
# grade: Blocking
394
source_file_dependencies:
@@ -402,6 +400,34 @@ steps:
400
- pytest -v -s v1/e2e
401
- pytest -v -s v1/engine
402
403
+- label: V1 Test e2e (2 GPUs) # 65min
404
+ timeout_in_minutes: 90
405
+ mirror_hardwares: [amdexperimental, amdproduction]
406
+ agent_pool: mi325_2
407
+ optional: true
408
+ # grade: Blocking
409
+ source_file_dependencies:
410
+ - vllm/
411
+ - tests/v1
412
+ commands:
413
+ # Only run tests that need exactly 2 GPUs
414
+ - pytest -v -s v1/e2e/test_spec_decode.py -k "tensor_parallelism"
415
+
416
+- label: V1 Test e2e (4 GPUs) # 65min
417
+ timeout_in_minutes: 90
418
+ mirror_hardwares: [amdexperimental, amdproduction]
419
+ # The test uses 4 GPUs, but we schedule it on 8-GPU machines for stability.
420
+ # See discussion here: https://github.com/vllm-project/vllm/pull/31040
421
+ agent_pool: mi325_4
422
+ optional: true
423
+ # grade: Blocking
424
+ source_file_dependencies:
425
+ - vllm/
426
+ - tests/v1
427
+ commands:
428
+ # Only run tests that need 4 GPUs
429
+ - pytest -v -s v1/e2e/test_spec_decode.py -k "eagle_correctness_heavy"
430
+
431
- label: V1 Test entrypoints # 35min
432
timeout_in_minutes: 50
433
mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
.buildkite/test_areas/engine.yaml
+33
-1
@@ -14,7 +14,7 @@ steps:
14
commands:
15
- pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py
16
17
-- label: V1 e2e + engine
17
+- label: V1 e2e + engine (1 GPU)
18
timeout_in_minutes: 45
19
source_file_dependencies:
20
- vllm/
@@ -36,3 +36,35 @@ steps:
36
commands:
37
- pytest -v -s v1/e2e
38
- pytest -v -s v1/engine
39
+
40
+- label: V1 e2e (2 GPUs)
41
+ timeout_in_minutes: 60 # TODO: Fix timeout after we have more confidence in the test stability
42
+ optional: true
43
+ num_devices: 2
44
+ source_file_dependencies:
45
+ - vllm/
46
+ - tests/v1/e2e
47
+ commands:
48
+ # Only run tests that need exactly 2 GPUs
49
+ - pytest -v -s v1/e2e/test_spec_decode.py -k "tensor_parallelism"
50
+ mirror:
51
+ amd:
52
+ device: mi325_2
53
+ depends_on:
54
+ - image-build-amd
55
+
56
+- label: V1 e2e (4 GPUs)
57
+ timeout_in_minutes: 60 # TODO: Fix timeout after we have more confidence in the test stability
58
+ optional: true
59
+ num_devices: 4
60
+ source_file_dependencies:
61
+ - vllm/
62
+ - tests/v1/e2e
63
+ commands:
64
+ # Only run tests that need 4 GPUs
65
+ - pytest -v -s v1/e2e/test_spec_decode.py -k "eagle_correctness_heavy"
66
+ mirror:
67
+ amd:
68
+ device: mi325_4
69
+ depends_on:
70
+ - image-build-amd
tests/v1/e2e/test_spec_decode.py
+1
-1
@@ -630,7 +630,7 @@ def test_eagle_correctness_medium(
630
False,
631
"auto",
632
0.8,
633
- marks=multi_gpu_marks(num_gpus=4),
633
+ marks=[*multi_gpu_marks(num_gpus=4), large_gpu_mark(min_gb=40)],
634
id="llama4_eagle",
635
),
636
pytest.param(