[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>
This commit is contained in:
Itay Etelis
2026-06-18 20:43:35 +03:00
committed by GitHub
co-authored by Itay Etelis
parent a0df04e477
commit ea6078fe6a
3 changed files with 18 additions and 0 deletions
+14
View File
@@ -64,6 +64,7 @@ def make_mapper_from_offloading_spec(**kwargs) -> FileMapper:
"dcp_size", 1
)
mock_vllm_config.parallel_config.rank = kwargs.get("rank", 0)
mock_vllm_config.use_v2_model_runner = kwargs.get("use_v2_model_runner", False)
mock_kv_cache_config = MagicMock()
mock_kv_cache_config.kv_cache_groups = kwargs.get("kv_cache_groups", [])
@@ -210,3 +211,16 @@ def test_parallel_agnostic_excludes_mla():
)
assert fm.fields["tp_size"] == 2
assert fm.rank == 1
def test_parallel_agnostic_disabled_on_v2_model_runner():
# V2's KV layout is not known to be parallelism-invariant: don't collapse.
fm = make_mapper_from_offloading_spec(
tp_size=2,
rank=1,
kv_cache_groups=[_full_attention_group()],
use_v2_model_runner=True,
parallel_agnostic=True,
)
assert fm.fields["tp_size"] == 2
assert fm.rank == 1
@@ -37,6 +37,7 @@ def _make_vllm_config():
decode_context_parallel_size=1,
rank=0,
),
use_v2_model_runner=False,
)
+3
View File
@@ -84,10 +84,13 @@ class FileMapper:
]
# Only a single full-attention group is parallelism-invariant. MLA is
# excluded: its latent KV is replicated per rank, never head-sharded.
# The V2 model runner is excluded: its KV layout is not known to be
# parallelism-invariant.
groups = kv_cache_config.kv_cache_groups
spec = groups[0].kv_cache_spec if len(groups) == 1 else None
parallel_agnostic = (
parallel_agnostic
and not vllm_config.use_v2_model_runner
and isinstance(spec, FullAttentionSpec)
and not isinstance(spec, MLAAttentionSpec)
)