[KV Connector][NIXL][Bugfix] Fix NIXL handshake failures not honoring kv_load_failure_policy (#40364)
When NIXL handshake fails (e.g., due to compatibility hash mismatch between prefill and decode instances), requests fail with "engine dead" error instead of gracefully falling back to local recomputation as configured by kv_load_failure_policy='recompute'. Signed-off-by: NickLucche <nlucches@redhat.com>
This commit is contained in:
@@ -388,10 +388,12 @@ class NixlConnectorWorker:
|
||||
# Set of requests that have been part of a batch, regardless of status.
|
||||
self._reqs_to_process: set[ReqId] = set()
|
||||
|
||||
# invalid blocks from failed NIXL operations
|
||||
self._invalid_block_ids: set[int] = set()
|
||||
# Invalid blocks from failed NIXL operations (thread-safe queue of block ids)
|
||||
self._invalid_block_ids: queue.Queue[set[int]] = queue.Queue()
|
||||
# requests that skipped transfer (handshake or transfer failures)
|
||||
self._failed_recv_reqs: set[ReqId] = set()
|
||||
# Uses Queue for thread-safe cross-thread coordination with the
|
||||
# background handshake thread, matching the _ready_requests pattern.
|
||||
self._failed_recv_reqs: queue.Queue[ReqId] = queue.Queue()
|
||||
|
||||
# Handshake metadata of this worker for NIXL transfers.
|
||||
self.xfer_handshake_metadata: NixlHandshakePayload | None = None
|
||||
@@ -768,11 +770,7 @@ class NixlConnectorWorker:
|
||||
error=e,
|
||||
meta=meta,
|
||||
)
|
||||
if (
|
||||
req_meta := self._recving_metadata.get(req_id)
|
||||
) and not self._is_hma_required:
|
||||
self._invalid_block_ids.update(req_meta.local_block_ids[0])
|
||||
self._failed_recv_reqs.add(req_id)
|
||||
self._handle_failed_transfer(req_id, None)
|
||||
|
||||
fut.add_done_callback(request_ready)
|
||||
|
||||
@@ -1696,17 +1694,26 @@ class NixlConnectorWorker:
|
||||
done_sending = self._get_new_notifs()
|
||||
done_recving = self._pop_done_transfers(self._recving_transfers)
|
||||
|
||||
# add requests that skipped transfer to done_recving
|
||||
done_recving.update(self._failed_recv_reqs)
|
||||
self._failed_recv_reqs.clear()
|
||||
# Drain queue of requests where handshake or transfer setup failed.
|
||||
failed_recv_reqs = set[ReqId]()
|
||||
while not self._failed_recv_reqs.empty():
|
||||
try:
|
||||
failed_recv_reqs.add(self._failed_recv_reqs.get_nowait())
|
||||
except queue.Empty:
|
||||
break
|
||||
|
||||
# Add failed requests to done_recving for scheduler tracking
|
||||
# (blocks are already marked invalid, scheduler will handle recompute)
|
||||
done_recving.update(failed_recv_reqs)
|
||||
|
||||
if len(done_sending) > 0 or len(done_recving) > 0:
|
||||
logger.debug(
|
||||
"Rank %s, get_finished: %s requests done sending "
|
||||
"and %s requests done recving",
|
||||
"and %s requests done recving (%s failed)",
|
||||
self.tp_rank,
|
||||
len(done_sending),
|
||||
len(done_recving),
|
||||
len(failed_recv_reqs),
|
||||
)
|
||||
|
||||
block_ids_for_blocksize_post_process = defaultdict(list)
|
||||
@@ -1715,6 +1722,15 @@ class NixlConnectorWorker:
|
||||
# clean up metadata for completed requests
|
||||
meta = self._recving_metadata.pop(req_id, None)
|
||||
assert meta is not None, f"{req_id} not found in recving_metadata list"
|
||||
|
||||
# Skip KV sync and post-processing for failed requests
|
||||
if req_id in failed_recv_reqs:
|
||||
logger.warning(
|
||||
"Skipping KV post-processing for failed request %s",
|
||||
req_id,
|
||||
)
|
||||
continue
|
||||
|
||||
assert meta.remote is not None
|
||||
if self.use_host_buffer:
|
||||
self.sync_recved_kv_to_device(req_id, meta)
|
||||
@@ -1889,7 +1905,7 @@ class NixlConnectorWorker:
|
||||
transfers[req_id] = in_progress
|
||||
return done_req_ids
|
||||
|
||||
def _handle_failed_transfer(self, req_id: str, handle: int):
|
||||
def _handle_failed_transfer(self, req_id: str, handle: int | None):
|
||||
"""
|
||||
Handle a failed transfer by marking all (logical) blocks as invalid and
|
||||
recording the failure.
|
||||
@@ -1901,8 +1917,10 @@ class NixlConnectorWorker:
|
||||
# Use .get() here as the metadata cleanup is handled by get_finished()
|
||||
# TODO (NickLucche) handle failed transfer for HMA.
|
||||
if (meta := self._recving_metadata.get(req_id)) and not self._is_hma_required:
|
||||
self._invalid_block_ids.update(meta.local_block_ids[0])
|
||||
self.nixl_wrapper.release_xfer_handle(handle)
|
||||
self._invalid_block_ids.put(set(meta.local_block_ids[0]))
|
||||
self._failed_recv_reqs.put(req_id)
|
||||
if handle is not None:
|
||||
self.nixl_wrapper.release_xfer_handle(handle)
|
||||
self.xfer_stats.record_failed_transfer()
|
||||
|
||||
def start_load_kv(self, metadata: NixlConnectorMetadata):
|
||||
@@ -2221,14 +2239,7 @@ class NixlConnectorWorker:
|
||||
dst_engine_id=dst_engine_id,
|
||||
remote_rank=remote_rank,
|
||||
)
|
||||
if (
|
||||
meta := self._recving_metadata.get(request_id)
|
||||
) and not self._is_hma_required:
|
||||
self._invalid_block_ids.update(meta.local_block_ids[0])
|
||||
self.xfer_stats.record_failed_transfer()
|
||||
if handle is not None:
|
||||
self.nixl_wrapper.release_xfer_handle(handle)
|
||||
self._failed_recv_reqs.add(request_id)
|
||||
self._handle_failed_transfer(request_id, handle)
|
||||
|
||||
def get_mapped_blocks(
|
||||
self, block_ids: np.ndarray, block_size_ratio: int
|
||||
@@ -2366,8 +2377,13 @@ class NixlConnectorWorker:
|
||||
This is called by the scheduler to identify blocks that need
|
||||
to be retried after a NIXL transfer failure.
|
||||
"""
|
||||
result = self._invalid_block_ids
|
||||
self._invalid_block_ids = set()
|
||||
# Drain the queue (thread-safe, no lock needed).
|
||||
result: set[int] = set()
|
||||
while not self._invalid_block_ids.empty():
|
||||
try:
|
||||
result.update(self._invalid_block_ids.get_nowait())
|
||||
except queue.Empty:
|
||||
break
|
||||
return result
|
||||
|
||||
def __del__(self):
|
||||
|
||||
Reference in New Issue
Block a user