Fix UCCL's consistency checks - #1151
Conversation
Signed-off-by: Pravein Govindan Kannan <pravein.govindan.kannan@ibm.com>
Signed-off-by: Pravein Govindan Kannan <pravein.govindan.kannan@ibm.com>
Signed-off-by: Pravein Govindan Kannan <pravein.govindan.kannan@ibm.com>
|
👋 Hi praveingk! Thank you for contributing to ai-dynamo/nixl. Your PR reviewers will review your contribution then trigger the CI to test your changes. 🚀 |
|
/ok to test 9f0ab11 |
|
/build |
|
i see this failure, which may be related. Checking if the CI node is healthy |
By any chance is it an AWS EFA node? Wondering why |
Signed-off-by: Pravein Govindan Kannan <pravein.govindan.kannan@ibm.com>
Signed-off-by: Pravein Govindan Kannan <pravein.govindan.kannan@ibm.com>
|
/ok to test 353baf0 |
|
/build |
|
@brminich Recently experimental EFA feature was pushed in UCCL. Probably, here it's hitting that on the AWS nodes. Can you please confirm if it's AWS nodes, and it was used previously? If so, I will need to turn off the EFA feature. |
|
@praveingk, no it is not AWS node. It is x86 with 1 h100 gpu and 1 BF3 NIC |
|
can you open this link https://nbuprod.blsm.nvidia.com/nbu-swx-nixl-main/blue/rest/organizations/jenkins/pipelines/nixl-ci-gpu/runs/108/nodes/199/log/?start=0? it contains some debug info including |
What?
Fix consistency check of UCCL
How?
UCCL backend was using the base address instead of the specific local/remote address during batches. This PR fixes that address and keeps multiple FIFO items for the batches instead of the single one used per handle.
A fix was also added to UCCL to enable memory check based on its range.