[CI] Increase ColQwen3 pooling test memory budget on H200 MIG (#55878)

Signed-off-by: khluu <khluu000@gmail.com>
2026-09-08 11:39:01 +00:00
parent 13cf9e05c1
commit 34b9899c8f
1 changed files with 3 additions and 1 deletions
@@ -45,7 +45,9 @@ TEXT_DOCUMENTS = [
 ]

 DTYPE = "half"
-GPU_MEMORY_UTILIZATION = 0.7
+# The Tomoro model needs room for its vision encoder and a 4096-token KV cache
+# on the 16 GiB devices used by the H200 MIG test lane.
+GPU_MEMORY_UTILIZATION = 0.8


 def _make_base64_image(