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 3, 21:30 UTC
[CI] Migrate remaining B200 jobs to b200-k8s with test fixes (#42387)
Signed-off-by: khluu <khluu000@gmail.com>
Kevin H. Luu committed
May 12, 2026 at 02:00 UTC
1ff9d335355c6f60afc3b8a5b4c15f36ab1ccfa6
5 files changed
+31
-7
.buildkite/test_areas/lm_eval.yaml
+13
-1
@@ -40,7 +40,7 @@ steps:
40
- label: LM Eval Small Models (B200)
41
key: lm-eval-small-models-b200
42
timeout_in_minutes: 120
43
- device: b200
43
+ device: b200-k8s
44
optional: true
45
source_file_dependencies:
46
- csrc/
@@ -48,6 +48,18 @@ steps:
48
commands:
49
- pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-blackwell.txt
50
51
+- label: LM Eval Large Models (B200, EP)
52
+ key: lm-eval-large-models-b200-ep
53
+ timeout_in_minutes: 120
54
+ device: b200-k8s
55
+ optional: true
56
+ num_devices: 2
57
+ source_file_dependencies:
58
+ - csrc/
59
+ - vllm/model_executor/layers/quantization
60
+ commands:
61
+ - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-blackwell-ep.txt
62
+
63
- label: LM Eval Qwen3.5 Models (B200)
64
key: lm-eval-qwen3-5-models-b200
65
timeout_in_minutes: 120
.buildkite/test_areas/spec_decode.yaml
+2
-2
@@ -16,7 +16,7 @@ steps:
16
- label: Spec Decode Eagle Nightly B200
17
key: spec-decode-eagle-nightly-b200
18
timeout_in_minutes: 30
19
- device: b200
19
+ device: b200-k8s
20
optional: true
21
source_file_dependencies:
22
- vllm/v1/spec_decode/
@@ -40,7 +40,7 @@ steps:
40
- label: Spec Decode Speculators + MTP Nightly B200
41
key: spec-decode-speculators-mtp-nightly-b200
42
timeout_in_minutes: 30
43
- device: b200
43
+ device: b200-k8s
44
optional: true
45
source_file_dependencies:
46
- vllm/v1/spec_decode/
tests/evals/gsm8k/configs/models-blackwell-ep.txt
new
+3
@@ -0,0 +1,3 @@
1
+Qwen3-Next-80B-A3B-NVFP4-EP2.yaml
2
+Qwen3-Next-FP8-EP2.yaml
3
+Nemotron-3-Super-120B-A12B-NVFP4.yaml
tests/evals/gsm8k/configs/models-blackwell.txt
-3
@@ -3,6 +3,3 @@ Qwen2.5-VL-3B-Instruct-FP8-dynamic.yaml
3
Qwen1.5-MoE-W4A16-CT.yaml
4
DeepSeek-V2-Lite-Instruct-FP8.yaml
5
Qwen3-30B-A3B-NVFP4.yaml
6
-Qwen3-Next-80B-A3B-NVFP4-EP2.yaml
7
-Qwen3-Next-FP8-EP2.yaml
8
-Nemotron-3-Super-120B-A12B-NVFP4.yaml
tests/v1/e2e/spec_decode/test_spec_decode.py
+13
-1
@@ -488,6 +488,10 @@ def _run_eagle_correctness(
488
489
490
@single_gpu_only
491
+@pytest.mark.skipif(
492
+ current_platform.is_device_capability_family(100),
493
+ reason="DeepSeek head_dim=192 not supported on SM100/SM110 (Blackwell)",
494
+)
495
@pytest.mark.parametrize(
496
[
497
"model_setup",
@@ -718,7 +722,15 @@ def test_eagle_correctness_heavy(
722
["model_setup", "mm_enabled", "expected_accuracy_threshold"],
723
[
724
(("mtp", "XiaomiMiMo/MiMo-7B-Base", 1), False, 0.5), # ref: 65%-70%
721
- (("mtp", "ZixiQi/DeepSeek-V3-4layers-MTP-FP8", 1), False, 0.0), # dummy model
725
+ pytest.param(
726
+ ("mtp", "ZixiQi/DeepSeek-V3-4layers-MTP-FP8", 1),
727
+ False,
728
+ 0.0,
729
+ marks=pytest.mark.skipif(
730
+ current_platform.is_device_capability_family(100),
731
+ reason="DeepSeek MTP: TRTLLM MoE top_k check fails on Blackwell",
732
+ ),
733
+ ), # dummy model
734
(
735
("mtp", "Qwen/Qwen3.5-0.8B-Base", 1),
736
False,