[Test] Pin block_size in auto-fit max_model_len test (#45914)
Signed-off-by: Ma, Liangliang <liangliang.ma@intel.com> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Liangliang Ma committed
Jun 24, 2026 at 22:56 UTC
2801b11156402ac4f3abb6c26ccbadf3cc502e58
1 file changed
+4
-2
tests/v1/e2e/general/test_context_length.py
+4
-2
@@ -75,8 +75,9 @@ def test_auto_fit_max_model_len_rejects_oversized_input(
75
must see this reduced value and reject prompts that exceed it,
76
rather than accepting them and hanging."""
77
78
- # Use a tiny KV cache budget to force auto-fit to a very small
79
- # max_model_len (e.g. ~16 tokens).
78
+ # Use a small KV cache budget to force auto-fit to a small
79
+ # max_model_len. Pin block_size=16 so the budget is independent
80
+ # of the platform's default block size.
81
kv_cache_bytes = 1_000_000 # 1 MB
82
83
with vllm_runner(
@@ -84,6 +85,7 @@ def test_auto_fit_max_model_len_rejects_oversized_input(
85
max_model_len=-1,
86
max_num_seqs=1,
87
enforce_eager=True,
88
+ block_size=16,
89
kv_cache_memory_bytes=kv_cache_bytes,
90
load_format="dummy",
91
) as vllm_model: