[KV Connector][Offloading] Disable parallel-agnostic fs-tier cache on V2 model runner (#46044)
Signed-off-by: Itay Etelis <etelis2019@gmail.com> Co-authored-by: Itay Etelis <etelis2019@gmail.com>
Itay Etelis committed
Jun 18, 2026 at 20:43 UTC
ea6078fe6a7242e7a5a89798e617b807d2540466
3 files changed
+18
tests/v1/kv_offload/test_file_mapper.py
+14
@@ -64,6 +64,7 @@ def make_mapper_from_offloading_spec(**kwargs) -> FileMapper:
64
"dcp_size", 1
65
)
66
mock_vllm_config.parallel_config.rank = kwargs.get("rank", 0)
67
+ mock_vllm_config.use_v2_model_runner = kwargs.get("use_v2_model_runner", False)
68
69
mock_kv_cache_config = MagicMock()
70
mock_kv_cache_config.kv_cache_groups = kwargs.get("kv_cache_groups", [])
@@ -210,3 +211,16 @@ def test_parallel_agnostic_excludes_mla():
211
)
212
assert fm.fields["tp_size"] == 2
213
assert fm.rank == 1
214
+
215
+
216
+def test_parallel_agnostic_disabled_on_v2_model_runner():
217
+ # V2's KV layout is not known to be parallelism-invariant: don't collapse.
218
+ fm = make_mapper_from_offloading_spec(
219
+ tp_size=2,
220
+ rank=1,
221
+ kv_cache_groups=[_full_attention_group()],
222
+ use_v2_model_runner=True,
223
+ parallel_agnostic=True,
224
+ )
225
+ assert fm.fields["tp_size"] == 2
226
+ assert fm.rank == 1
tests/v1/kv_offload/tiering/test_obj_tier.py
+1
@@ -37,6 +37,7 @@ def _make_vllm_config():
37
decode_context_parallel_size=1,
38
rank=0,
39
),
40
+ use_v2_model_runner=False,
41
)
42
43
vllm/v1/kv_offload/file_mapper.py
+3
@@ -84,10 +84,13 @@ class FileMapper:
84
]
85
# Only a single full-attention group is parallelism-invariant. MLA is
86
# excluded: its latent KV is replicated per rank, never head-sharded.
87
+ # The V2 model runner is excluded: its KV layout is not known to be
88
+ # parallelism-invariant.
89
groups = kv_cache_config.kv_cache_groups
90
spec = groups[0].kv_cache_spec if len(groups) == 1 else None
91
parallel_agnostic = (
92
parallel_agnostic
93
+ and not vllm_config.use_v2_model_runner
94
and isinstance(spec, FullAttentionSpec)
95
and not isinstance(spec, MLAAttentionSpec)
96
)