[Perf][HiSparse] Reuse pinned host pool after graph profiling (#61668)

Signed-off-by: Itay Etelis <92247226+Etelis@users.noreply.github.com>
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Co-authored-by: Matthew Bonanni <mbonanni@redhat.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Itay Etelis
2026-10-15 12:33:03 -07:00
committed by GitHub
co-authored by Matthew Bonanni Claude Opus 5.5
parent 105605b8e9
commit 49f944be3e
6 changed files with 87 additions and 56 deletions
+20 -1
View File
@@ -223,7 +223,9 @@ class HiSparseHostPool:
self.shared_region: SharedOffloadRegion | None = None
def allocate(self, size: int) -> torch.Tensor:
assert self.backing is None, "HiSparse host pool is already allocated."
if self.backing is not None:
assert self.backing.numel() == size, "HiSparse host pool size changed."
return self.backing
config = self.kv_cache_config
if config.hisparse_shared_host_pool:
tensor_sizes = [
@@ -253,6 +255,23 @@ class HiSparseHostPool:
return self.backing
_host_pools: dict[str, HiSparseHostPool] = {}
def get_hisparse_host_pool(
vllm_config: VllmConfig, kv_cache_config: KVCacheConfig
) -> HiSparseHostPool:
"""Return the engine's host pool, which outlives the profiling KV cache."""
instance_id = vllm_config.instance_id
if instance_id not in _host_pools:
_host_pools[instance_id] = HiSparseHostPool(vllm_config, kv_cache_config)
return _host_pools[instance_id]
def drop_hisparse_host_pool(vllm_config: VllmConfig) -> None:
_host_pools.pop(vllm_config.instance_id, None)
def use_shared_hisparse_host_pool(vllm_config: VllmConfig) -> bool:
"""Whether replicated MLA host KV can share one local mmap."""
parallel = vllm_config.parallel_config