[CI] Fit small KV-offload evals within shared memory (#52496)
Signed-off-by: Taneem Ibrahim <taneem.ibrahim@gmail.com>
Taneem Ibrahim committed
Aug 16, 2026 at 15:16 UTC
6b0b850a8b1764a66d7ffbb023c0b0e0bbdb900b
1 file changed
+2
tests/evals/gsm8k/test_gsm8k_offloading.py
+2
@@ -133,6 +133,7 @@ MODELS = [
133
connector="OffloadingConnector",
134
# Baseline ~0.49 on 200 questions (measured on GB200).
135
accuracy_threshold=0.39,
136
+ cpu_offload_gib=1,
137
),
138
OffloadingModelConfig(
139
id="offloading-gemma-4-e4b-it",
@@ -140,6 +141,7 @@ MODELS = [
141
connector="OffloadingConnector",
142
# Baseline ~0.64 on 200 questions (measured on GB200).
143
accuracy_threshold=0.55,
144
+ cpu_offload_gib=1,
145
),
146
OffloadingModelConfig(
147
id="offloading-qwen3.5-35b",