[CI][Bugfix] Use a prompt that survives offload-resume rounding in mamba offload test (#53146)
Signed-off-by: Frederik Gossen <frgossen@meta.com> Co-authored-by: Claude <noreply@anthropic.com>
Frederik Gossen committed
Aug 21, 2026 at 10:14 UTC
5ee84d3c5207bd59faa53e19e60fa82e34961444
1 file changed
+9
-1
tests/v1/kv_connector/unit/test_offloading_connector.py
+9
-1
@@ -574,7 +574,15 @@ def test_mamba_cpu_offload_boundary(
574
)
575
576
_PROMPT_SIZE: int = block_size * 2
577
- _PROMPT_TEXT = "Hi. Give me a set of trivia questions and their answers "
577
+ # Cold prefill and offload-resume are not bit-exact (selective_scan_fn over
578
+ # the whole prompt vs selective_state_update for the recomputed token), so
579
+ # exact-text equality only holds while the rounding difference never flips a
580
+ # greedy argmax. Use a long, highly predictable prompt: it needs no "...."
581
+ # padding and keeps top-1 ahead of top-2 by >2.8 logprobs at every step.
582
+ _PROMPT_TEXT = (
583
+ "The cat sat on the mat. The cat sat on the mat. The cat sat on the mat. "
584
+ "The cat sat on the mat. The cat sat on the mat. The cat sat on the mat. "
585
+ )
586
587
# build prompt ids to match prompt_size
588
tokenizer = llm.get_tokenizer()