[BUGFIX] Fix log capture in KV test (#49655)
Signed-off-by: zhenwei-intel <zhenwei.liu@intel.com> Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
liuzhenwei committed
Jul 25, 2026 at 08:36 UTC
6a1acac3fe2924ca221fdf048b74f9a020dabddd
1 file changed
+2
-2
tests/v1/core/test_kv_cache_utils.py
+2
-2
@@ -2183,7 +2183,7 @@ def test_mla_draft_prefers_standard_layout_when_pages_can_be_unified():
2183
)
2184
2185
2186
-def test_mla_with_incompatible_swa_uses_one_full_allocation_group(caplog):
2186
+def test_mla_with_incompatible_swa_uses_one_full_allocation_group(caplog_vllm):
2187
# Sparse MLA pages cannot be padded safely. Keeping the draft's attention
2188
# compute sliding-window while promoting only its allocation semantics lets
2189
# every layer share the target's block table and remain contiguous.
@@ -2206,7 +2206,7 @@ def test_mla_with_incompatible_swa_uses_one_full_allocation_group(caplog):
2206
assert promoted_draft.block_size == 64
2207
assert promoted_draft.sliding_window == draft.sliding_window
2208
assert specs["draft.0"] is draft
2209
- assert "attention compute is unchanged" in caplog.text
2209
+ assert "attention compute is unchanged" in caplog_vllm.text
2210
2211
2212
def test_get_kv_cache_spec_kind_prefers_specific_attention_subclasses():