[RFC][TorchElastic] topology info in training apps/ranks in #57
Conversation
There was a problem hiding this comment.
Do we have any details on what will consume WORLD_TOPOLOGY_DETAILS? Also can you share any details on how we'll compute the graph? Integrating with things like IB/NVLink seem pretty cluster specific -- can we autodetect that topology in all cases?
Also wondering about things like AWS with spline topologies etc which impact distributed performance pretty significantly
Profiling bandwidth/latency also seems tricky when there are many nodes so would be nice to see some details on that
| "measurement": "GB/s" | ||
| }, | ||
| "channels": { | ||
| "value": "4" |
There was a problem hiding this comment.
is this number of connections or lanes? multi-NIC vs # of nvlink lanes?
| - `RANK` - unique rank of a worker (0…WORLD_SIZE-1) | ||
| - `LOCAL_RANK` - unique rank of a worker on a node, typically used to exclusively assign accelerators on the host. | ||
|
|
||
| New proposed `WORLD_TOPOLOGY_FILE` environment variable will reference a local filesystem file that will provide information about the underlying topology. |
There was a problem hiding this comment.
would this be a required? if not what would the default topology be? if so, what is the proposal for backwards-compatibility?
| - Most of the can be easily detected at runtime by the trainer code | ||
| - More fine-grained details based on communication pattern (p2p vs collectives) | ||
|
|
||
| ### **Format of the topology information file** |
There was a problem hiding this comment.
could we define a TopologyInfo schema (e.g. dataclass or protobuf) and have a Reader API that can be extended to read from various sources? The default implementation could read from a simple json/yaml file, but I can imagine folks running on the cloud wanting to read from a database or directly from a cloud storage like s3 or (to your point below) auto-discovered and dynamically generated.
There was a problem hiding this comment.
+on API, adding that to proposal.
On additional datasources: we need a single point of discovery of this data at application level therefore it has to be controlled via underlying infra setup. We can extend to other datasources but I believe factory mechanism should be encapsulated.
…s into rcf-torchelastic-toplogy-info
|
Hi @kurman! Thank you for your pull request. We require contributors to sign our Contributor License Agreement, and yours needs attention. You currently have a record in our system, but the CLA is no longer valid, and will need to be resubmitted. ProcessIn order for us to review and merge your suggested changes, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA. Once the CLA is signed, our tooling will perform checks and validations. Afterwards, the pull request will be tagged with If you have received this in error or have any questions, please contact us at [email protected]. Thanks! |
Proposal to provide topology information to training apps/ranks that can be implemented as part of TorchElastic.