fix(pytorch_xdit): key benchmark output dirs on the cluster target, not hostname - #320
Open
speriaswamy-amd wants to merge 3 commits into
Conversation
…ot hostname
Both benchmarks named their per-node output directory from the remote
`hostname` command (wan_22_${hostname}_outputs, flux_${hostname}_outputs), and
the run step and the parse step each resolved it independently. When a node
reports a short name in one context and an FQDN in another, the two steps
disagree and parse fails with "no results found" even though the run wrote
them. It also scatters results across several directory trees for what is one
node.
The cluster node_dict key — the SSH target string the user already wrote in
cluster.json — is stable by construction and known without a remote call. New
utils_lib.cluster_target_output_label() normalizes it (strips whitespace, maps
"/" to "_", falls back to "unknown_node").
Both modules now derive output directories from that label in the run step and
in the parse step, including WAN's multi-node aggregation filter, which
previously built its expected-directory list from a second `hostname` probe.
Also pins _test_output_dir for single-node WAN runs, matching what FLUX already
did, so parse does not re-derive the path at all in the common case. `hostname`
is still collected, but only for the log line.
Adds a unit test for the new helper.
Co-Authored-By: Claude <noreply@anthropic.com>
This was referenced Aug 14, 2026
…t-07-output-dir-labels (bring branch up to date)
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Part 7 of 12 in a stack that replaces #184. Base: #319.
Why
Both benchmarks named their per-node output directory from the remote
hostnamecommand (wan_22_${hostname}_outputs,flux_${hostname}_outputs), and the run step and the parse step each resolved it independently. When a node reports a short name in one context and an FQDN in another, the two steps disagree and parse fails with "no results found" even though the run wrote them — and results scatter across several directory trees for what is one node.The cluster
node_dictkey — the SSH target string already incluster.json— is stable by construction and needs no remote call.What changed
utils_lib.cluster_target_output_label()— normalizes the key (strips whitespace,/→_, falls back tounknown_node). Unit tested.hostnameprobe to build its expected-directory list._test_output_dirfor single-node WAN runs, matching what FLUX already did, so parse does not re-derive the path at all in the common case.hostnameis still collected, but only for the log line.