[Core] Deprecate "all" mamba cache mode (#55041)
Signed-off-by: Nick Hill <nickhill123@gmail.com>
Nick Hill committed
Sep 3, 2026 at 08:56 UTC
e55b93f2969d76b854eae158b6d17baea757d3e8
3 files changed
+18
-1
tests/model_executor/test_routed_experts_capture.py
+4
-1
@@ -459,7 +459,10 @@ def test_v2_model_runner_accepts_routed_experts(monkeypatch):
459
mode=CompilationMode.NONE,
460
pass_config=SimpleNamespace(enable_sp=False),
461
),
462
- cache_config=SimpleNamespace(kv_sharing_fast_prefill=False),
462
+ cache_config=SimpleNamespace(
463
+ kv_sharing_fast_prefill=False,
464
+ mamba_cache_mode="none",
465
+ ),
466
ec_transfer_config=None,
467
)
468
vllm/config/cache.py
+11
@@ -323,6 +323,17 @@ class CacheConfig:
323
self.user_specified_mamba_block_size = True
324
return self
325
326
+ @field_validator("mamba_cache_mode", mode="after")
327
+ @classmethod
328
+ def _validate_mamba_cache_mode(cls, mode: MambaCacheMode) -> MambaCacheMode:
329
+ if mode == "all":
330
+ logger.warning_once(
331
+ "Mamba cache mode 'all' is deprecated and will be removed in an "
332
+ "upcoming release. If this is a problem, please open an issue "
333
+ "at https://github.com/vllm-project/vllm/issues."
334
+ )
335
+ return mode
336
+
337
@field_validator("cache_dtype", mode="after")
338
@classmethod
339
def _validate_cache_dtype(cls, cache_dtype: CacheDType) -> CacheDType:
vllm/config/vllm.py
+3
@@ -2614,6 +2614,9 @@ class VllmConfig:
2614
# Will be added by https://github.com/vllm-project/vllm/pull/35045
2615
unsupported.append("KV sharing fast prefill")
2616
2617
+ if self.cache_config.mamba_cache_mode == "all":
2618
+ unsupported.append("mamba cache mode 'all'")
2619
+
2620
return unsupported
2621
2622
def _get_v1_model_runner_unsupported_features(self) -> list[str]: