jundot--omlx
wehub-resource-sync
129 行
3.8 KiB
Python
129 行
3.8 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
"""
|
|
Cache Recovery Manager for oMLX.
|
|
|
|
This module handles error recovery for cache corruption and other cache-related
|
|
failures, enabling the scheduler to continue processing after encountering errors.
|
|
"""
|
|
|
|
import gc
|
|
import logging
|
|
from typing import TYPE_CHECKING, Any, Dict, List, Optional
|
|
|
|
from ..exceptions import is_cache_corruption_error
|
|
|
|
if TYPE_CHECKING:
|
|
from ..request import Request, RequestStatus
|
|
from ..prefix_cache import BlockAwarePrefixCache
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
|
|
class CacheRecoveryManager:
|
|
"""
|
|
Manages cache error recovery for the scheduler.
|
|
|
|
This class provides methods to detect and recover from cache corruption,
|
|
including clearing caches and rescheduling affected requests.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
block_aware_cache: Optional["BlockAwarePrefixCache"] = None,
|
|
):
|
|
"""
|
|
Initialize the cache recovery manager.
|
|
|
|
Args:
|
|
block_aware_cache: Block-aware prefix cache (paged SSD-based).
|
|
"""
|
|
self.block_aware_cache = block_aware_cache
|
|
|
|
def is_cache_corruption(self, error: Exception) -> bool:
|
|
"""
|
|
Check if an error indicates cache corruption.
|
|
|
|
Args:
|
|
error: The exception to check.
|
|
|
|
Returns:
|
|
True if the error appears to be cache corruption.
|
|
"""
|
|
return is_cache_corruption_error(error)
|
|
|
|
def recover(
|
|
self,
|
|
batch_generator_holder: Any,
|
|
request_id_to_uid: Dict[str, int],
|
|
uid_to_request_id: Dict[int, str],
|
|
request_detokenizers: Dict[str, Any],
|
|
) -> None:
|
|
"""
|
|
Recover from cache corruption error.
|
|
|
|
This method clears the batch generator and all caches, resetting
|
|
the system to a clean state.
|
|
|
|
Args:
|
|
batch_generator_holder: Object holding the batch_generator reference.
|
|
request_id_to_uid: Mapping of request IDs to UIDs.
|
|
uid_to_request_id: Mapping of UIDs to request IDs.
|
|
request_detokenizers: Dict of request detokenizers.
|
|
"""
|
|
# Clear batch generator (this is the source of the corruption)
|
|
batch_generator_holder.batch_generator = None
|
|
if hasattr(batch_generator_holder, '_current_sampler_params'):
|
|
batch_generator_holder._current_sampler_params = None
|
|
|
|
# Clear cache
|
|
if self.block_aware_cache is not None:
|
|
self.block_aware_cache.clear()
|
|
|
|
# Clear UID mappings
|
|
request_id_to_uid.clear()
|
|
uid_to_request_id.clear()
|
|
|
|
# Clear detokenizer state to prevent contamination after recovery
|
|
request_detokenizers.clear()
|
|
|
|
# Force garbage collection
|
|
gc.collect()
|
|
|
|
logger.info("Cache recovery completed")
|
|
|
|
def reschedule_running_requests(
|
|
self,
|
|
running: Dict[str, "Request"],
|
|
waiting: Any, # deque[Request]
|
|
request_status_waiting: "RequestStatus",
|
|
) -> int:
|
|
"""
|
|
Move running requests back to waiting queue for retry.
|
|
|
|
Args:
|
|
running: Dictionary of running requests by ID.
|
|
waiting: Deque of waiting requests.
|
|
request_status_waiting: The WAITING status enum value.
|
|
|
|
Returns:
|
|
Number of requests rescheduled.
|
|
"""
|
|
count = len(running)
|
|
|
|
for request_id, request in list(running.items()):
|
|
# Reset request state
|
|
request.status = request_status_waiting
|
|
request.batch_uid = None
|
|
request.prompt_cache = None
|
|
request.cached_tokens = 0
|
|
request.remaining_tokens = request.prompt_token_ids
|
|
|
|
# Move to waiting queue (at front for priority)
|
|
waiting.appendleft(request)
|
|
del running[request_id]
|
|
|
|
if count > 0:
|
|
logger.info(f"Rescheduled {count} requests for retry")
|
|
|
|
return count
|