[kv_offload+HMA][1/N]: Support multiple KV groups in OffloadingSpec (#36610)

Signed-off-by: Or Ozeri <oro@il.ibm.com>
This commit is contained in:
Or Ozeri
2026-03-13 08:04:21 +00:00
committed by GitHub
parent 99a57bdf74
commit cfaf4668f7
5 changed files with 80 additions and 22 deletions
@@ -126,6 +126,7 @@ class OffloadingConnector(KVConnectorBase_V1):
):
super().__init__(vllm_config, role, kv_cache_config)
assert kv_cache_config is not None
spec = OffloadingSpecFactory.create_spec(vllm_config, kv_cache_config)
self.connector_scheduler: OffloadingConnectorScheduler | None = None
@@ -245,9 +246,10 @@ class OffloadingConnectorScheduler:
"""Implementation of Scheduler side methods"""
def __init__(self, spec: OffloadingSpec):
self.gpu_block_size = spec.gpu_block_size
self.offloaded_block_size = spec.offloaded_block_size
self.block_size_factor = self.offloaded_block_size // self.gpu_block_size
assert len(spec.gpu_block_size) == 1
self.gpu_block_size = spec.gpu_block_size[0]
self.offloaded_block_size = self.gpu_block_size * spec.block_size_factor
self.block_size_factor = spec.block_size_factor
self.manager: OffloadingManager = spec.get_manager()
self._requests: dict[ReqId, Request] = {}