vllm.config.fault_tolerance
¶
Classes:
-
FaultToleranceConfig–Configuration for fault tolerance.
Functions:
-
ft_tp_barrier_required–Whether the per-step FT barrier over the TP cpu group is armed.
FaultToleranceConfig
¶
Configuration for fault tolerance.
Attributes:
-
engine_recovery_timeout_sec(int) –Timeout (in seconds) to wait for error handling instructions
-
request_failure_policy(Literal['recompute', 'fail']) –Policy for handling unfinished requests when the engine encounters a fault.
Source code in vllm/config/fault_tolerance.py
engine_recovery_timeout_sec = 120
class-attribute
instance-attribute
¶
Timeout (in seconds) to wait for error handling instructions before raising an exception. If the EngineCore encounters an error, it waits up to this many seconds for vLLM to receive instructions on how to handle the error and then recover from the fault. If vLLM does not recover during this time, the original error is raised.
request_failure_policy = 'fail'
class-attribute
instance-attribute
¶
Policy for handling unfinished requests when the engine encounters a fault.
'recompute': preempt requests back to the waiting queue to recompute their
KV cache and continue generation after recovery, without abort notifications.
'fail': abort unfinished requests immediately (default).
Passing params={"abort": true} to the retry instruction discards
queued requests even when the policy is 'recompute'.
ft_tp_barrier_required(parallel_config)
¶
Whether the per-step FT barrier over the TP cpu group is armed.
The barrier makes TP peers fail on the host within the gloo timeout instead of leaving an orphaned TP collective on the device stream.