[kv_offload+HMA][0/N]: Support block-level preemption handling (#34805)

Signed-off-by: Or Ozeri <oro@il.ibm.com>
Co-authored-by: Nicolò Lucchesi <nlucches@redhat.com>
This commit is contained in:
Or Ozeri
2026-03-18 08:49:53 +02:00
committed by GitHub
co-authored by Nicolò Lucchesi
parent 86b7e3c95a
commit fcf0687b27
7 changed files with 27 additions and 25 deletions
@@ -25,8 +25,8 @@ The class provides the following primitives:
Worker-side: runs in each worker, loads/saves KV cache to/from
the Connector based on the metadata.
handle_preemptions() - called if there are preempted requests,
before their blocks are overwritten
handle_preemptions() - called for handling preempted requests
or request evicted blocks before they are overwritten
start_load_kv() - starts loading all KVs (maybe async)
wait_for_layer_load() - blocks until layer i load is done
@@ -288,9 +288,9 @@ class KVConnectorBase_V1(ABC):
"""
return
def handle_preemptions(self, preempted_req_ids: set[str]):
def handle_preemptions(self, kv_connector_metadata: KVConnectorMetadata):
"""
Handle preempted requests BEFORE their blocks are overwritten.
Handle preempted requests or evicted blocks BEFORE they are overwritten.
Needed for connectors which use async saves (e.g., OffloadingConnector)
"""
return
@@ -315,10 +315,11 @@ class MultiConnector(KVConnectorBase_V1):
for c in self._connectors:
c.set_host_xfer_buffer_ops(copy_operation)
def handle_preemptions(self, preempted_req_ids: set[str]):
def handle_preemptions(self, kv_connector_metadata: KVConnectorMetadata):
"""Handle preempted requests for all sub-connectors."""
for c in self._connectors:
c.handle_preemptions(preempted_req_ids)
assert isinstance(kv_connector_metadata, MultiKVConnectorMetadata)
for c, cm in zip(self._connectors, kv_connector_metadata.metadata):
c.handle_preemptions(cm)
def get_finished_count(self) -> int | None:
# TODO(https://github.com/vllm-project/vllm/issues/33400)
@@ -111,6 +111,7 @@ class OffloadingConnectorStats(KVConnectorStats):
class OffloadingConnectorMetadata(KVConnectorMetadata):
reqs_to_load: dict[ReqId, TransferSpec]
reqs_to_store: dict[ReqId, TransferSpec]
reqs_to_flush: set[str] | None = None
class OffloadingConnector(KVConnectorBase_V1):
@@ -146,9 +147,10 @@ class OffloadingConnector(KVConnectorBase_V1):
assert self.connector_worker is not None
self.connector_worker.register_cross_layers_kv_cache(kv_cache, attn_backend)
def handle_preemptions(self, preempted_req_ids: set[str]):
def handle_preemptions(self, kv_connector_metadata: KVConnectorMetadata):
assert self.connector_worker is not None
self.connector_worker.handle_preemptions(preempted_req_ids)
assert isinstance(kv_connector_metadata, OffloadingConnectorMetadata)
self.connector_worker.handle_preemptions(kv_connector_metadata)
def start_load_kv(self, forward_context: "ForwardContext", **kwargs) -> None:
assert self.connector_worker is not None
@@ -482,6 +484,7 @@ class OffloadingConnectorScheduler:
meta = OffloadingConnectorMetadata(
reqs_to_load=self._reqs_to_load,
reqs_to_store=self._get_reqs_to_store(scheduler_output),
reqs_to_flush=scheduler_output.preempted_req_ids,
)
self._reqs_to_load = {}
@@ -619,13 +622,13 @@ class OffloadingConnectorWorker:
attn_backends = {cross_layer_name: attn_backend}
self._register_handlers(kv_caches, attn_backends)
def handle_preemptions(self, preempted_req_ids: set[str]):
def handle_preemptions(self, kv_connector_metadata: OffloadingConnectorMetadata):
for job_id, transfer_spec in self._unsubmitted_store_jobs:
success = self.worker.transfer_async(job_id, transfer_spec)
assert success
self._unsubmitted_store_jobs.clear()
for req_id in preempted_req_ids:
for req_id in kv_connector_metadata.reqs_to_flush or ():
job_ids = self._store_jobs.get(req_id)
if job_ids:
self.worker.wait(job_ids)