[ROCm][Engine] Fix GPU memory leaks in engine shutdown and test workaround for async KV prefix cache reset (#38503)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
Andreas Karatzas
2026-04-25 05:25:20 +00:00
committed by GitHub
parent 07351e0883
commit 95995bbef8
8 changed files with 300 additions and 13 deletions
+25 -4
View File
@@ -45,16 +45,14 @@ def run_once(f: Callable[P, None]) -> Callable[P, None]:
@lru_cache
def supports_kw(
def _supports_kw(
callable: Callable[..., object],
kw_name: str,
*,
requires_kw_only: bool = False,
allow_var_kwargs: bool = True,
) -> bool:
"""Check if a keyword is a valid kwarg for a callable; if requires_kw_only
disallows kwargs names that can also be positional arguments.
"""
"""Internal cached implementation of supports_kw."""
params = inspect.signature(callable).parameters
if not params:
return False
@@ -99,6 +97,29 @@ def supports_kw(
return False
def supports_kw(
callable: Callable[..., object],
kw_name: str,
*,
requires_kw_only: bool = False,
allow_var_kwargs: bool = True,
) -> bool:
"""Check if a keyword is a valid kwarg for a callable; if requires_kw_only
disallows kwargs names that can also be positional arguments.
"""
# Unwrap bound methods so that the lru_cache key is the underlying
# function, not the instance. Caching bound methods pins the object
# (and all its GPU tensors) for the lifetime of the cache.
if hasattr(callable, "__func__"):
callable = callable.__func__
return _supports_kw(
callable,
kw_name,
requires_kw_only=requires_kw_only,
allow_var_kwargs=allow_var_kwargs,
)
def get_allowed_kwarg_only_overrides(
callable: Callable[..., object],
overrides: Mapping[str, object] | None,