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 14, 05:30 UTC
[CI] Add Qwen3 MoE NVFP4 to Blackwell lm-eval (#26316)
Signed-off-by: mgoin <mgoin64@gmail.com>
Michael Goin committed
Oct 7, 2025 at 13:36 UTC
30a3e5af69b8d27af1d2ff350a6f69a61b605591
3 files changed
+8
-1
.buildkite/test-pipeline.yaml
+1
-1
@@ -868,7 +868,7 @@ steps:
868
- pytest -s -v tests/quantization/test_blackwell_moe.py
869
870
- label: Blackwell LM Eval Small Models
871
- timeout_in_minutes: 75
871
+ timeout_in_minutes: 120
872
gpu: b200
873
optional: true # run on nightlies
874
source_file_dependencies:
tests/evals/gsm8k/configs/Qwen3-30B-A3B-NVFP4.yaml
new
+6
@@ -0,0 +1,6 @@
1
+model_name: "nvidia/Qwen3-30B-A3B-FP4"
2
+accuracy_threshold: 0.89
3
+num_questions: 1319
4
+num_fewshot: 5
5
+max_model_len: 4096
6
+
tests/evals/gsm8k/configs/models-blackwell.txt
+1
@@ -2,3 +2,4 @@ Qwen3-0.6B-FP8.yaml
2
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