[ziqifan617]

Make the existing KVCR startup timeout configurable per instance via KVCRConfig.startup_timeout_seconds, allowing longer waits for large memory registrations.

None preserves the 30-second default; explicit values must be positive integers. The option applies to the local prepare() and start() waits only, without changing transfer deadlines or requiring matching values across nodes.

[karya0]

@ziqifan617 : Have you tried: UCX_REG_NONBLOCK_MEM_TYPES=host and UCX_IB_ODP_PREFETCH=n ? These solve the startup problem by using lazy initialization.

[ziqifan617]

@ziqifan617 : Have you tried: UCX_REG_NONBLOCK_MEM_TYPES=host and UCX_IB_ODP_PREFETCH=n ? These solve the startup problem by using lazy initialization.

Thanks @karya0 , I haven't. Will give it a shot when I have the hw to try it again.

for your suggestion, codex gives below concers:

1. ODP isn’t universally available. UCX checks device capabilities and can fall back to blocking registration. These host settings also don’t address CUDA-memory registration. UCX implementation
2. The timeout controls the whole initialization wait, including agent/backend initialization and metadata capture—not only host registration.

But overall, I had met a few times that the registration is slow somehow so exposing it out for future sake as well.

WDYT?