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 5, 09:45 UTC
Add nightly b200 test for spec decode eagle correctness (#38577)
Signed-off-by: Rishi Puri <riship@nvidia.com>
Rishi Puri committed
Apr 9, 2026 at 13:09 UTC
adaabb8a55b3109babaaee7e2821ceefe05d6db4
1 file changed
+34
.buildkite/test_areas/spec_decode.yaml
+34
@@ -12,6 +12,17 @@ steps:
12
commands:
13
- pytest -v -s v1/e2e/spec_decode -k "eagle_correctness"
14
15
+- label: Spec Decode Eagle Nightly B200
16
+ timeout_in_minutes: 30
17
+ device: b200
18
+ optional: true
19
+ source_file_dependencies:
20
+ - vllm/v1/spec_decode/
21
+ - vllm/v1/worker/gpu/spec_decode/
22
+ - tests/v1/e2e/spec_decode/
23
+ commands:
24
+ - pytest -v -s v1/e2e/spec_decode -k "eagle_correctness"
25
+
26
- label: Spec Decode Speculators + MTP
27
timeout_in_minutes: 30
28
device: h200_18gb
@@ -23,6 +34,18 @@ steps:
34
commands:
35
- pytest -v -s v1/e2e/spec_decode -k "speculators or mtp_correctness"
36
37
+- label: Spec Decode Speculators + MTP Nightly B200
38
+ timeout_in_minutes: 30
39
+ device: b200
40
+ optional: true
41
+ source_file_dependencies:
42
+ - vllm/v1/spec_decode/
43
+ - vllm/v1/worker/gpu/spec_decode/
44
+ - vllm/transformers_utils/configs/speculators/
45
+ - tests/v1/e2e/spec_decode/
46
+ commands:
47
+ - pytest -v -s v1/e2e/spec_decode -k "speculators or mtp_correctness"
48
+
49
- label: Spec Decode Ngram + Suffix
50
timeout_in_minutes: 30
51
device: h200_18gb
@@ -42,3 +65,14 @@ steps:
65
- tests/v1/e2e/spec_decode/
66
commands:
67
- pytest -v -s v1/e2e/spec_decode -k "draft_model or no_sync or batch_inference"
68
+
69
+- label: Spec Decode Draft Model Nightly B200
70
+ timeout_in_minutes: 30
71
+ device: b200
72
+ optional: true
73
+ source_file_dependencies:
74
+ - vllm/v1/spec_decode/
75
+ - vllm/v1/worker/gpu/spec_decode/
76
+ - tests/v1/e2e/spec_decode/
77
+ commands:
78
+ - pytest -v -s v1/e2e/spec_decode -k "draft_model or no_sync or batch_inference"