feat(aorta): disaggregated multi-node torchrun launch (AIMVT-173) - #329
Open
speriaswamy-amd wants to merge 1 commit into
Open
feat(aorta): disaggregated multi-node torchrun launch (AIMVT-173)#329speriaswamy-amd wants to merge 1 commit into
speriaswamy-amd wants to merge 1 commit into
Conversation
Runs the Aorta benchmark across N nodes from a single cluster.json instead of requiring N single-node cluster files. run() now branches on the launch mode resolved from the multi_node block: 'script' is today's behavior verbatim (experiment_script on the head node), 'torchrun' launches a rank-group on every node in parallel, all rendezvous-ing on the head. 'auto' - the default - picks script for 1-node clusters and torchrun above that, so existing single-node configs are unaffected. New helpers: - _resolve_launch_mode() - auto/script/torchrun; 'script' with >1 node is a config error rather than a silent single-node run. - _resolve_master_addr() - explicit override, else the head node's VPC/RDMA address from node_vpc_ips, else the plain node id. The mgmt/SSH address is frequently orchestrator-only and unreachable from peer nodes. - _pick_free_port_on()/_pick_master_port() - binds and releases a port over SSH on the head node, which is the host that actually holds the rendezvous socket. The snippet is passed as one shlex-quoted argument; as separate argv elements the remote shell re-splits it on its own semicolons. - _build_torchrun_command(), _run_single_node(). TraceLens/GEMM analysis now explicitly target the head node's container, and training-log discovery iterates the node list, rather than relying on locals left over from the single-node code path. Co-Authored-By: Claude <noreply@anthropic.com>
This was referenced Aug 14, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stack 4/6 — splits #171. Base: #328. Implements AIMVT-173.
Why
Run the Aorta benchmark across N nodes from a single
cluster.jsoninstead of requiring N single-node cluster files, mirroring the disaggregated pattern the PyTorch xDiT and SGLang multi-node suites already use.What changed
run()branches on the resolved launch mode:script— today's behavior verbatim (experiment_scripton the head node).torchrun— a rank-group on every node in parallel, rendezvous-ing on the head.auto(default) —scriptfor 1-node clusters,torchrunabove that. Existing single-node configs are unaffected.New helpers in
cvs/runners/aorta.py:_resolve_launch_mode()—scriptwith >1 node is a config error rather than a silent single-node run._resolve_master_addr()— explicit override → head node's VPC/RDMA address → plain node id. The mgmt/SSH address is frequently orchestrator-only and unreachable from peer nodes._pick_free_port_on()/_pick_master_port()— binds and releases a port over SSH on the head node, which is the host that actually holds the rendezvous socket. The snippet is passed as oneshlex-quoted argument; as separate argv elements the remote shell re-splits it on its own semicolons._build_torchrun_command(),_run_single_node().TraceLens/GEMM analysis now explicitly target the head node's container, and training-log discovery iterates the node list, rather than relying on locals left over from the single-node path.
Test
ruffclean. Unit tests 603 → 623 (20 new: mode resolution, port/addr selection, command construction,validate_config).