chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Fix nightly VLM accuracy: gemma3n TP fixes + removal, latency thresholds (#19401)

Browse Source 
Co-authored-by: Alison Shao <alisonshao@MacBook-Pro-D2W773R9CD.local>
This commit is contained in:
Alison Shao
2026-02-27 14:24:02 -08:00
committed by GitHub
co-authored by Alison Shao
parent e6da514c2c
commit 6ca7da3e7c
3 changed files with 15 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/eval/test_vlms_mmmu_eval.py
+4 -2
View File
@@ -34,7 +34,9 @@ MODEL_THRESHOLDS = {
0.270, 23.8
),
ModelLaunchSettings("google/gemma-3-4b-it"): ModelEvalMetrics(0.360, 10.9),
ModelLaunchSettings("google/gemma-3n-E4B-it"): ModelEvalMetrics(0.270, 17.7),
ModelLaunchSettings(
"google/gemma-3n-E4B-it", extra_args=["--tp=2"]
): ModelEvalMetrics(0.270, 17.7),
ModelLaunchSettings("mistral-community/pixtral-12b"): ModelEvalMetrics(0.360, 16.6),
ModelLaunchSettings("moonshotai/Kimi-VL-A3B-Instruct"): ModelEvalMetrics(
0.330, 23.5
@@ -50,7 +52,7 @@ MODEL_THRESHOLDS = {
ModelLaunchSettings(
"unsloth/Mistral-Small-3.1-24B-Instruct-2503"
): ModelEvalMetrics(0.30, 16.7),
ModelLaunchSettings("XiaomiMiMo/MiMo-VL-7B-RL"): ModelEvalMetrics(0.28, 32.0),
ModelLaunchSettings("XiaomiMiMo/MiMo-VL-7B-RL"): ModelEvalMetrics(0.28, 40.0),
ModelLaunchSettings("zai-org/GLM-4.1V-9B-Thinking"): ModelEvalMetrics(0.280, 30.4),
ModelLaunchSettings(
"zai-org/GLM-4.5V-FP8", extra_args=["--tp=2"]