Skip to content

Fix/materialize model deployment match#55

Open
Erotemic wants to merge 41 commits into
mainfrom
fix/materialize-model-deployment-match
Open

Fix/materialize model deployment match#55
Erotemic wants to merge 41 commits into
mainfrom
fix/materialize-model-deployment-match

Conversation

@Erotemic

@Erotemic Erotemic commented May 7, 2026

Copy link
Copy Markdown
Contributor

No description provided.

Erotemic and others added 11 commits March 28, 2026 16:58
HELM's public bucket reorganized so classic now lives at
gs://crfm-helm-public/classic/benchmark_output/runs/<ver>/<run_id> like
every other benchmark. The legacy gs://crfm-helm-public/benchmark_output/runs/
location is empty — every recent --list-versions for classic returned
nothing because of the special-case here. Remove it (and the now-redundant
names.add('classic') and 'benchmark_output' blocklist entry).

Verified via the GCS JSON API: classic/ is a normal top-level prefix
holding v0.1.0/v0.2.2/v0.2.3/v0.2.4/v0.3.0/v0.4.0; the legacy path
returns no objects and no prefixes.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
…-run.log

Replaces the ub.cmd(..., system=True) invocation with
system=False, capture=True so the wrapped helm-run process's stdout
and stderr are captured (in addition to streaming live to the terminal
via verbose=3). On every run we write the last 64 KB of each stream
to cmd_stdout.txt and cmd_stderr.txt next to the existing helm-run.log.

Motivation: when helm-run raises *before* its own logger initializes
— most commonly a TypeError from run_spec_function(**args) when a
run_entry contains a kwarg the function doesn't accept — the
traceback ends up only in the parent shell's stderr and is lost when
the run dir gets rsync'd. eval_audit's failure-classification
taxonomy can then only bucket these as truncated_or_incomplete_runtime
(the runtime log shows config-load + clean exit and nothing else).
With the captured streams persisted, eval_audit can recognize and
label them as malformed_run_entry.

Best-effort: a write failure on the cmd_*.txt files never masks the
underlying helm-run exit code.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant